AI Agents for Small Business: 10 Jobs to Hand Off (2026)

The practical guide to the jobs a small business can give an AI agent today, what each one costs, and where it goes wrong

Two-thirds of US small businesses now use AI, yet only 14% say it is fully embedded in how they operate. The US Chamber of Commerce found that 66% of small businesses use AI in 2026, up from 23% in 2023 - US Chamber of Commerce. Goldman Sachs surveyed the alumni of its 10,000 Small Businesses program and found 76% using AI, but only 14% with AI fully embedded in their core operations - Goldman Sachs. Put plainly, most owners use AI as a writing tool they open in a browser tab. Very few have handed it a job.

That gap is about to be tested. On September 29, Meta launched Muse for Small Business, an agent that connects to Shopify, Stripe, QuickBooks, Slack and about a dozen other tools and works toward a goal you set, under one firm rule: "nothing publishes, sends, or spends without your approval" - Meta. The same day, OpenAI put always-on dots into ChatGPT, each with its own cloud computer, and four days earlier Microsoft rebuilt Copilot around a "digital teammate" called Autopilot. Underneath all three, the price of the models that power agents fell through the floor: Claude Haiku 5.5, released on October 7, costs $0.10 per million input tokens on typical prompts and runs about 75% cheaper on average than the model it replaces - Anthropic.

The problem for an owner is no longer access. It is knowing which job to hand over first. Hand off the wrong one, such as customer replies that make promises you cannot keep, and the agent creates work instead of removing it. Hand off the right one, with the right guardrails, and you get back hours every week for a few dollars a day. The difference between those two outcomes has very little to do with which AI model is "smartest" and almost everything to do with the shape of the job itself.

This guide ranks the 10 jobs a small business can realistically hand to an AI agent in late 2026, scored on the hours they return, the cost of a mistake, how ready the tools are and what they cost to run. For each job it covers which products do it, what they cost at list price, what the evidence says about results, and how it fails. It then covers the platforms, from Meta, OpenAI, Microsoft and Google to builders like Lindy and Zapier and agent workforces such as O-mega, followed by the full cost math for a 10-person business, the failure cases worth learning from, and a step-by-step method for your first handoff. It assumes no technical background: every concept is explained from scratch.

Contents

1. Why Agents Finally Fit a Small Business
2. Where Small Businesses Really Stand With AI
3. What an AI Agent Is, in Plain Terms
4. Job 1: Bookkeeping and Reconciliation
5. Job 2: Answering Customer Messages
6. Job 3: Invoicing and Getting Paid
7. Job 4: Marketing Content and Social Posts
8. Job 5: Answering the Phone and Booking (the AI Receptionist)
9. Job 6: Running the Online Store
10. Job 7: The Owner's Inbox, Calendar and Weekly Numbers
11. Job 8: Reviews and Local Reputation
12. Job 9: Lead Follow-Up and Sales Admin
13. Job 10: Hiring Admin and Candidate Screening
14. The Platforms: Built-In Agents, General Agents, Builders and Workforces
15. What It Costs: The Math for a 10-Person Business
16. Where Agents Fail, and the Jobs to Keep
17. How to Hand Off Your First Job
18. What Comes Next
19. Conclusion: A Decision Framework for Owners

The 10 Jobs, Scored

Before the detail, here is the whole ranking in one view. The ten rows below are not products, they are jobs: recurring pieces of work that exist in most small businesses whether or not anyone uses AI. Each job is scored from 0 to 10 on four criteria, and the final column is the weighted average. Every cell carries both a score and the evidence behind it, so you can disagree with a weight and recompute the order for your own business.

The criteria come from first principles about what a small business owner actually trades when delegating work, to a person or to software. You trade your time (the hours the job takes now), you accept the risk of a mistake (and how easily it can be undone), you depend on the tools existing and working, and you pay a running cost. A job worth handing off scores well on all four. A job that scores high on time but low on safety is one to hand off carefully, with approvals, rather than one to avoid.

How to read the weights. Time returned carries 30% because it is the reason to delegate at all: a job that takes ten minutes a month is not worth the setup, however well an agent does it. Safety of mistakes and tool readiness carry 25% each because they decide whether the time saved is real or gets eaten by cleanup. Cost carries 20% because, as section 15 shows, the running cost of most of these agents is small next to the wage cost of the same hours, so price rarely decides the order on its own. The scores describe a typical small business. A plumbing company that wins most jobs by phone should treat phone answering as its first job, and a firm with no online store can ignore row six entirely.

The honest headline from this table is that the best first job is usually the boring one. Bookkeeping categorization tops the list not because it is exciting but because the agent is already inside software you pay for, its mistakes are caught before the month closes, and nobody outside the business ever sees them. Customer messages come second with the most time to give back and the sharpest risk, which is why that section spends as much space on guardrails as on tools. The rest of this guide explains each row and the platforms behind them.

1. Why Agents Finally Fit a Small Business

A small business is a set of jobs done by too few people. That simple fact explains why so much business software never reached small firms: it sold tools, and every tool needs an operator. A CRM does not follow up on leads, a helpdesk does not answer customers, an accounting package does not reconcile the bank account. A person does, and in a five-person company that person is usually the owner, at night. Large companies bought software and hired operators to run it. Small companies bought less software because the operator was the missing piece.

An agent changes the unit being sold. Instead of a tool that waits for a person, it is software that does a piece of finished work: the answered customer, the categorized transaction, the drafted post. That is why agents matter more to a small business than to a large one, even though large firms get the headlines. A large company replaces some of an operator's hours. A small company gets an operator it could never afford to hire. For that to work, three things had to become true at the same time, and in 2026 they did.

The price of a unit of work collapsed

The first condition is cost. An agent that answers every customer message or reads every invoice consumes model capacity on every task, so it only pays when the cost per task is far below the cost of the human minute it replaces. The chart below shows current list prices for the models that sit underneath most small-business agents, as published by Anthropic and OpenAI. The two cheapest models, Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna, cost exactly the same: $0.10 per million input tokens and $0.50 per million output tokens - OpenAI.

A token is a fragment of a word: on Anthropic's current tokenizer, a million tokens is roughly 555,000 words, several long novels' worth of text - Anthropic. Consider what that means for one customer conversation. If an agent reads 20,000 tokens of your policies, product details and the customer's history, then writes 1,000 tokens of replies, the model cost on Haiku 5.5 is about a quarter of a cent. The median US customer service representative earns $21.53 an hour - US Bureau of Labor Statistics, so even a three-minute human reply costs about a dollar in wages alone. The model is no longer the expensive part of an AI agent. As section 15 shows, the expensive parts are the packaging, the integrations and the supervision.

Anthropic is explicit about where its cheapest model fits: it says Haiku 5.5 "works especially well for speed-sensitive tasks like live customer support and browser use" and scores 72.4% on the offline subset of the OSWorld 2.1 computer-use benchmark, against 15.7% for Haiku 4.5 - Anthropic. OpenAI launched a Decisions API on October 6 that runs on GPT-6 Luna and turns text and images into typed answers, the kind of call an agent makes to sort a ticket or classify an invoice - OpenAI changelog. We compare these routing tools in our breakdown of Jev, Clef and the Decisions API, and track the full price ladder monthly in our best LLM for AI agents ranking.

The agents moved into the tools owners already use

The second condition is reach. An agent can only do work in the systems it can open, and small businesses live in a handful of them: an accounting package, a store, a payment processor, an inbox, a social account. Until this year, connecting an agent to those systems meant hiring a developer. Now the connections ship with the product. Meta's launch lists connectors for Shopify, Stripe, Intuit QuickBooks, Slack, Canva, Klaviyo and more, plus the business's own Facebook and Instagram accounts - Meta. At the same time, the software vendors built agents into their own products: Shopify's Sidekick now sits on every screen of the store admin, and Intuit runs a set of agents inside QuickBooks.

The week of September 25 to September 29 compressed this shift into a few days. Microsoft rebuilt Copilot, Meta launched Muse for Small Business, and OpenAI's DevDay introduced dots alongside more than a dozen other releases. CNET's 15-minute supercut of the DevDay keynote is the quickest way to see what OpenAI shipped: the live demo of dots at work starts at the 3:06 mark, and the segment on the Decisions API and GPT-6 Luna at 10:14.

The pattern to notice in that recap is that none of the headline launches was a new chatbot. They were agents with standing access: a dot that keeps working toward a goal, a Copilot that "goes to work" once you give it a role, a Muse that runs a growth plan across your tools. For a small business, standing access is the whole point, because the owner cannot sit and prompt a chatbot all day. It is also the source of every risk discussed in section 16, which is why the third condition matters as much as the first two.

Approval became the default, not an afterthought

The third condition is control that does not need an IT department. A large company governs software through identity systems, security teams and audit logs. A small business has none of those, so an agent that acts on its own is only acceptable if the owner can see and approve what it does. The 2026 products build approval in from the start. Meta's rule is that nothing publishes, sends or spends without approval. OpenAI lets you set rules per action, with options that include "Take action without asking" and "Take action if pre-approved" - OpenAI Help Center. Shopify says Sidekick respects access controls, so staff "only interact with data and features they're authorized to use" - Shopify.

Why this matters: approval turns an agent from a gamble into a draft machine with a gradual path to autonomy. On day one, you approve everything and lose a few minutes reviewing. As the agent proves itself on a job, you widen what it may do alone. That progression, rather than any single feature, is what makes delegation safe for a business with no security staff. How to apply it: whenever you evaluate an agent, ask first how its approval settings work, at what level of detail (per action, per amount, per customer), and whether it keeps a log you can read in five minutes on a Friday. If the answers are vague, the product is not ready for your business, however impressive the demo.

2. Where Small Businesses Really Stand With AI

Before choosing a job to hand off, it helps to know where the market really is, because the headline numbers disagree wildly. Depending on which survey you read, somewhere between one in five and three in four small businesses use AI. Both figures are correct. They measure different businesses and ask different questions, and understanding why tells you more about adoption than either number alone.

The low numbers come from the US Census Bureau's Business Trends and Outlook Survey, which samples all employer firms in the country, including the many that say AI simply does not apply to what they do, the most common reason for non-adoption in the Bureau's own research. Its latest published rate is 19.8% of firms using AI in their operations - US Census Bureau. The high numbers come from surveys of businesses that are already engaged with a program or a product: 66% in the Chamber's survey, 76% among Goldman Sachs program alumni, and 77% of US small and midsize businesses in Intuit's QuickBooks research, which asked about regular use as of January 2026 - Stacker, for Intuit.

The trend is steep, and the question keeps changing

The Census series is the most rigorous view of change over time, and it shows a steep climb. When the Bureau started asking, 3.7% of firms said they used AI to produce goods or services, rising to 5.4% by February 2024 - US Census Bureau working paper. By September 2025 the same question reached 10%, and when Census widened the wording in November 2025 to cover AI in any business function, the rate jumped to 17.3% - Economic Innovation Group. The Federal Reserve put adoption at about 18% of firms at the end of 2025 - Federal Reserve. The chart shows the two question wordings as separate lines, because joining them would exaggerate the jump.

The size gap is just as clear. Across the EU, 17% of small enterprises (10 to 49 employees) used AI in 2025, against 55.03% of large ones - Eurostat. In the UK, 28% of businesses with up to nine employees reported using at least one AI technology, against 49% of those with 250 or more - Office for National Statistics. In the US, 37% of firms with at least 250 employees use AI, well above the national rate - US Census Bureau. Smaller firms are behind, and the reason the surveys give is not cost. Eurostat found the most common barrier was a lack of relevant expertise, cited by 70.89% of firms that considered AI and did not use it.

Wide use, shallow depth

The more important finding is about depth. Goldman's 14% fully embedded figure is mirrored in the UK, where only 10% of AI-using firms with ten or more employees report extensive use - Office for National Statistics. Bluevine's 2026 survey of 942 US owners found that 48% of AI users save at least four hours a week and 14% save ten hours or more, but also that 78% of owners do not fully trust AI to handle even low-level tasks without oversight - Bluevine. Only 52% report a return on their investment so far.

Those numbers describe a market that has adopted AI as a tool and not yet as labor. Owners use it to write faster, which saves minutes, but the work still passes through their hands. The Fed's survey work points the same way: nearly half of small employer firms reported using AI in some capacity, and 71% of those reported higher productivity as a result - Federal Reserve. That is real value, but it is the value of a faster pen. The step change comes from handing over whole jobs, which is exactly what most owners have not done yet and what the rest of this guide is about.

How to apply this: do not benchmark yourself against the "everyone uses AI" headlines. Benchmark the jobs. Write down the five tasks that took the most of your week last month, and for each one ask whether AI did any of it start to finish. If the answer is no for all five, you are in the majority, and the gain available to you is far larger than the one you get from using a chatbot to draft emails. Our analysis of why most AI agent pilots never scale shows the same pattern in larger firms: the value appears when a job is redesigned around the agent, not when an agent is sprinkled on top of the old job.

3. What an AI Agent Is, in Plain Terms

"Agent" is used so loosely in marketing that it helps to define it from scratch. There are three kinds of software an owner might be offered, and they differ in one thing: who decides the next step. A chatbot answers when you ask and then stops; you decide every next step. An automation (the kind built in Zapier for years) follows fixed rules: when an invoice arrives, copy it to a folder. It never decides anything. An agent is given a goal and decides its own steps toward it, using tools, until the goal is met or it needs to ask you something.

That ability to choose steps is what lets an agent take a job rather than a task. "Answer this customer" is a task. "Keep the support inbox under control, using our refund policy, and escalate anything over $100 to me" is a job, and only an agent can hold it, because the steps depend on what each customer writes. It is also why agents are riskier than automations: a fixed rule cannot surprise you, but an agent that chooses steps sometimes chooses badly. The whole craft of using agents in a small business is getting the benefits of choice while bounding the cost of a bad choice.

The five parts of every agent

Every agent product, from Meta's to a home-built one, is made of the same five parts. Knowing them lets you compare products in a sales call and spot which part is weak. The model is the part everyone talks about, but for a small business it is the least likely to be the problem, since the cheap models are now good at routine work. The parts that usually decide success are the connectors and the rules.

A useful way to hold the five parts in mind is to picture a capable temp sent by an agency. The temp arrives with general skills, which is the model, but on the first morning they still need logins, a written brief, some sense of how things were done before, and a manager who checks their work until trust is earned. No agency would claim a temp could succeed without those four things, yet agent marketing often implies the model alone is enough. Asking a vendor how each of the five parts works in their product is the fastest way to see past that.

• The model: the reasoning engine, such as Claude Haiku 5.5 or GPT-6 Luna
• Connectors: access to your apps, like QuickBooks, Shopify or Gmail
• Instructions and rules: the job description, your policies and your limits
• Memory: what it knows about your business, customers and past decisions
• Approvals and a log: what it must ask you about, and a record of what it did

Read that list as a hiring checklist. A new employee with a brilliant mind but no login to your systems, no written policies and no manager would fail, and an agent fails the same way. In practice, most disappointing agent deployments in small firms trace back to the third item: nobody wrote down the refund policy, the pricing exceptions or the tone, so the agent guessed. The image below, from Meta's launch, shows how the platforms now present the second item: a grid of connectors that an owner switches on rather than builds.

Each icon in that grid is a system the agent can read from and, with approval, write to: the store, the books, the design tool, the team chat. The practical point is that the connector list is now a buying criterion in its own right. If an agent cannot connect to the system where a job lives, it cannot do the job, however capable its model. Check the list against your own stack before anything else, and treat "custom connectors" (which Meta also supports) as a project that will need someone technical.

The loop: how a handed-off job actually runs

Once those parts are in place, every delegated job runs as the same loop. Something triggers the agent: a customer message, a new bill, a review, a scheduled time. The agent reads the context it needs, drafts an action, and checks the action against your rules. If the action is inside the rules and easy to undo, it acts and records what it did. If not, it asks you. Each week, you read the log and adjust the rules, so the share of actions that need you shrinks over time.

The dashed line back to the start is where the value compounds. Every time you correct the agent, the correction should become a rule ("never offer free shipping on orders under $50"), not a one-off edit. Owners who treat corrections as rules see the approval queue shrink month after month. Owners who silently fix drafts see it stay the same size, because the agent never learns what they wanted. Why this matters: the loop is what turns a tool you supervise into labor you manage. How to apply it: whatever product you choose, keep a running document of rules for each job, and paste every correction into it the moment you make one.

The approval dial

Approval is not a switch but a dial with four useful positions. At the first, the agent only drafts, and you send everything yourself. At the second, it prepares actions and you approve each with one click. At the third, it acts alone on routine cases and asks only about exceptions, which you define by amount, customer type or topic. At the fourth, it acts alone and you review the log after the fact. OpenAI's settings for dots map onto this dial directly, letting you choose per action whether a dot asks or acts. Every job in this guide has a natural starting position on the dial, and each job section says where to start and what evidence should move you to the next position.

4. Job 1: Bookkeeping and Reconciliation

Bookkeeping is the job most owners would never list as "something to automate with AI", which is exactly why it belongs at the top. The work is categorizing every bank and card transaction, matching payments to invoices and bills, reconciling each account at month end and spotting anything that looks wrong. It is pure routine, it repeats every day, and it is usually done badly in small firms because nobody has time for it until the accountant asks. The median US bookkeeping clerk earns $24.36 an hour, and the Bureau of Labor Statistics projects employment in the occupation to decline 6% between 2025 and 2035 - US Bureau of Labor Statistics, largely because software keeps absorbing the routine parts.

What makes this the best first job is not the model but the location. The agent already lives inside software most small businesses pay for. Intuit introduced an Accounting Agent in QuickBooks Online that "automates bookkeeping and transaction categorization, and assists in reconciliation" - Intuit, and Xero says its JAX agent automates "bank reconciliations, data entry, and getting paid" - Xero. There is nothing new to buy, no new login to manage, and no customer ever sees the agent's work.

What the evidence says

The outcome data here is vendor-reported, so read it as a ceiling rather than a promise. Intuit's own survey found that 45% of customers using its AI-powered bank feed save 12 hours each month on bookkeeping, based on QuickBooks Online customers surveyed as of April 2025 - Intuit. Twelve hours a month at the median clerk wage is roughly $290 of labor, every month, for a feature included in the subscription. The more conservative reading is that less than half of users report that saving, which matches the general pattern in Bluevine's data that about half of small businesses using AI save four or more hours a week.

The reason bookkeeping scores well on safety is structural. A miscategorized transaction does not leave the building. It sits in the ledger until someone reviews it, and the month-end close is a natural checkpoint where a human looks at the numbers anyway. Compare that with a customer reply, which lands in someone's inbox the moment it is sent. Reversibility plus a built-in review point is the combination that lets an agent run at a high position on the approval dial without much risk.

Where it goes wrong

Agents categorize the common transactions well and the unusual ones badly, and the unusual ones are where the tax consequences live. Owner draws, loan repayments, mixed personal and business spending, a one-off equipment purchase that should be capitalized: these are rare, ambiguous and expensive to get wrong. An agent will usually pick a plausible category with complete confidence. The failure is not that it errs but that the error looks exactly like a correct entry, so it survives unless someone looks.

The fix is not a better model but a narrower role. In bookkeeping, the agent should decide only what your rules already decide, and route everything else to a short list for you or your accountant. Four guardrails do most of that work, and each takes minutes to set up in QuickBooks or Xero, because both products already support rules for recurring transactions and review queues for uncertain ones.

• Vendor rules: tell it once that a given supplier always maps to one category
• Exception thresholds: anything above a set amount goes to your review list
• Uncategorized is allowed: make "ask me" an acceptable answer for odd items
• Quarterly accountant review: a professional checks the agent's patterns

Those four guardrails turn the agent's weakness into a short weekly list. The vendor rules handle the bulk of transactions deterministically, so the agent's judgment is only used on what is new. The threshold and the "ask me" option push rare items to you instead of letting the agent guess. The accountant review catches systematic errors, such as a recurring payment filed under the wrong expense type for six months, which no single weekly check would notice. Once the rules are written, the review list should shrink to the genuinely new items each week rather than the whole ledger.

How to apply this: start at position two on the approval dial for the first month, approving categorizations in batches, and write a rule every time you change one. In month two, let the agent categorize alone below your threshold and keep reviewing exceptions. Ask your accountant to look at the first quarter's books specifically for agent patterns. If you also use Muse for Small Business with QuickBooks connected, you can use one of the prompts Meta suggests at launch, "How was my business's financial performance this month? Find any expenses that look off", which turns the month-end review into a short read.

5. Job 2: Answering Customer Messages

Customer messages are the job that gives back the most time and carries the sharpest risk. Most small businesses receive the same twenty questions over and over: opening hours, delivery times, return rules, whether an item is in stock, how to reschedule. Each takes a few minutes, they arrive all day across email, website chat, Instagram and Facebook, and they interrupt everything else. Intuit's research found customer service is the second most common use of AI among small businesses, at 37%, behind only marketing - Stacker, for Intuit.

This is also the most mature category of agent, because large companies have funded it for years. The products small businesses can buy today grew out of enterprise support tools and now sell with outcome-based pricing: you pay when a conversation is resolved, not per seat or per message. That pricing model matters more than any feature, because it means the vendor only earns when the agent actually finishes the job. We compared the enterprise end of this market in our guide to Sierra vs Decagon. For a small business, the relevant options are smaller and cheaper.

The tools and their prices

Intercom's Fin charges $0.99 per outcome, with a minimum of 50 outcomes a month when it runs on top of the helpdesk you already have, and $29 per helpdesk seat if you move to Intercom's own helpdesk - Fin. HubSpot moved its Breeze Customer Agent to $0.50 per resolved conversation in April and reports that the agent "resolves 65% of conversations and cuts resolution time by 39%", across more than 8,000 customers who have switched it on - HubSpot. That agent requires HubSpot's Professional or Enterprise plans. Online stores often look at Gorgias, a helpdesk built for ecommerce with its own AI agent.

A resolution rate of 65% sounds modest until you translate it into time. If your business gets 400 customer conversations a month and each takes six minutes of a person's attention, that is 40 hours. An agent that resolves 65% of them returns 26 of those hours for $130 at HubSpot's price or about $257 at Fin's. At the median customer service wage of $21.53 an hour, the same 26 hours cost about $560 in wages before taxes and benefits. The other 35% still reach a person, which is the right outcome: those are the conversations that need judgment.

The binding risk: promises

The danger in this job is not that the agent is rude or slow. It is that the agent makes a promise on your behalf that you would not have made, and the law treats it as yours. The defining case is Air Canada's. Its website chatbot told a grieving customer he could claim a bereavement discount after travel, contradicting the airline's actual policy. Air Canada argued the chatbot was "a separate legal entity responsible for its own actions", and the tribunal rejected that, holding the airline responsible for all information on its website - American Bar Association.

A 2026 case shows the same failure in a smaller company. Who Gives A Crap, the toilet paper brand, sent a pricing email containing a typo, and when a customer asked about the apparent price increase, its AI agent confirmed the wrong price instead of correcting it or handing off to a person. The company said it "immediately shut down" the agent and sent a correcting email - CX Foundation. Notice that the agent did not invent anything. It trusted the wrong source and agreed with the customer, which is the most natural thing for a language model to do.

• Answer only from written sources: policies, FAQs, product data you supplied
• Never improvise policy: refunds, discounts and exceptions go to a person
• Escalate by topic: cancellations, complaints, legal threats, anything about money
• Show it is an AI: customers forgive a bot that hands off, not one that pretends

Those four rules turn the promise risk into a routing problem. The first two keep the agent inside what you have already committed to, so its answers cannot exceed your policies. The third makes sure the conversations most likely to create liability reach a human. The fourth protects trust, and in the EU it is also becoming a legal duty, since the AI Act's transparency obligations to tell people they are dealing with an AI system apply from August 2026, as section 13 explains. Small businesses that follow all four rarely appear in the incident reports.

How to apply this: write or update your FAQ and policies first, because the agent will be exactly as good as those documents. Turn the agent on in draft mode for one week, reading every reply before it goes out, and turn each correction into a document change. Then let it answer the top ten question types alone while everything else still comes to you. Track two numbers weekly: the share resolved without you, and the number of replies you would have worded differently. The first should rise and the second should fall; if the second does not fall, your documents are missing something.

6. Job 3: Invoicing and Getting Paid

Chasing money is the job owners hate most and do least consistently. Sending an invoice is easy; following up on the late ones is awkward, so it slips, and the cash arrives weeks late. For a small business, that delay is not an accounting detail. It is the difference between making payroll comfortably and borrowing to cover it. The job involves creating invoices from completed work, sending them, reminding late payers on a schedule, matching incoming payments to the right invoices, and noticing which customers always pay slowly.

The first-principles case for handing this job to an agent is that the value of a reminder comes from consistency, not cleverness. A polite reminder sent on day one, day seven and day fourteen, every time, for every customer, beats a perfectly worded one sent when the owner remembers. Humans are bad at that kind of consistency for socially uncomfortable tasks. Agents do not feel awkward, do not forget, and do not let a busy week push the reminders back.

The tools

This job, like bookkeeping, now ships inside accounting software. Intuit's Payments Agent in QuickBooks drafts invoice reminders and reports "getting businesses paid an average of 5 days faster" - Intuit. The fine print matters: that figure compares US beta customers using AI-drafted reminders with customers sending standard reminders between January and August 2024, and it is Intuit's own measurement. Xero's JAX covers the same ground. If you run payments through Stripe, Meta's Muse for Small Business connects to it, so the agent can see which payments arrived and draft follow-ups for the ones that did not.

Five days faster sounds small until you put it in cash terms. A business invoicing $40,000 a month that collects five days sooner holds roughly $6,600 more cash on any given day, without selling anything extra. For a firm that runs close to its overdraft, that can be worth more than a new customer. The reason it works is not that AI-written reminders are more persuasive, though they may be slightly better targeted. It is that they go out on time.

What to keep under your control

The risks in this job are concentrated in a few actions that involve money moving the wrong way or a relationship being damaged. An agent that offers a discount for early payment, agrees to a payment plan, issues a credit note or waives a late fee is making a commercial decision. An agent that sends a stern third reminder to your largest customer, who happens to be negotiating a renewal, can cost far more than the invoice. Neither is likely, but both are possible, and the guardrails are cheap.

The practical approach is to split customers into two groups. For routine customers, let the agent send reminders on a fixed schedule with no approval, which is position three or four on the dial. For your key accounts, keep the agent at position two: it drafts, you approve. Make credits, refunds, discounts and payment plans approval-only for everyone. Payment matching can run at position three, with anything ambiguous (two invoices for the same amount, a partial payment) going to your review list. If you later want agents that pay suppliers as well as collect, the infrastructure is arriving, as our agent payments guide explains, but paying out is a much higher-risk job than collecting and belongs much later in the sequence.

How to apply this: turn on the reminder agent in your accounting software with a schedule you would be comfortable sending yourself, mark your top ten customers as manual, and set every money-moving action to approval. After a month, compare your average days to payment with the month before. If it improved and no customer complained, widen the schedule to more customers; if a customer complained, read the reminder they received and adjust the tone rules rather than switching the agent off.

7. Job 4: Marketing Content and Social Posts

Marketing is where small businesses already use AI most. Intuit found it is the top use case, at 45% of small and midsize businesses - Stacker, for Intuit. Most of that use is drafting: a caption here, a newsletter there, an owner pasting a prompt into a chatbot late at night. The job worth handing off is bigger: planning a week of posts, drafting them in your voice, adapting them for each channel, drafting ad variations, and reporting which ones worked. It is a job that repeats weekly, consumes hours, and is easy to let lapse.

The structural change here is that content has become almost free to produce, which shifts the scarce resource. When a draft costs nothing, volume stops being a goal. What becomes scarce is attention (whether anyone engages), truthfulness (whether the claims are accurate) and distinctiveness (whether it sounds like you or like everyone else using the same tools). An agent that floods your channels with generic posts can make your marketing worse. An agent grounded in your actual sales data and your voice can make it better, because it does the analysis most owners skip.

Muse for Small Business is built for this job

Meta's launch is aimed squarely at this work. Muse connects to your Instagram professional analytics, Facebook Pages and Meta ad accounts, plus Canva and Klaviyo, and one of the prompts Meta suggests reads: "How do I improve my ads and content? Can you analyze what's working or not, and draft a campaign for next week?" Muse for Small Business "is available for free with usage limits", with paid subscriptions for heavier use - TechCrunch. Meta's approval rule fits marketing perfectly: nothing publishes without your say-so, so the agent works as a tireless drafter while you keep the final word.

Muse carries the same name as Meta's model family, which began with the Muse Spark model in April, as we covered in our Muse Spark guide. Meta has not said which model powers the small-business agent, and for this job it matters less than the connectors: an agent that can see which of last month's posts drove sales has information a stronger model without that access lacks. If your marketing runs mostly on Google or email rather than Meta's apps, the general agents in section 14 and design tools such as Canva cover the same drafting work.

Where marketing agents go wrong

The common failures are not dramatic, but they accumulate. Agents state things as facts that are not true of your business ("family-owned since 1985" when you opened in 2019), promise outcomes you cannot guarantee, borrow phrasing that sounds like every competitor, or use images and claims you do not have rights to. Each one is caught by a two-minute read before publishing, which is why marketing scores well on reversibility in theory but only 6 out of 10 on safety: once a post is public, deleting it does not delete the screenshots.

The fix is a short brand brief that the agent always reads: who you are, the facts it may state, the claims it may never make, three examples of posts that sounded exactly right, and words you never use. With that page written, approving a week of drafts becomes a single short sitting rather than an evening of writing from scratch, and posting becomes consistent instead of depending on how busy the week was.

How to apply this: connect the agent to your social analytics before anything else, so its drafts are informed by what worked. Ask for one week of posts at a time, approve them in one sitting, and note every edit in the brief. Keep ads at position two on the dial permanently, since ad spend is money leaving the business, and Meta's own rule already requires your approval for anything that spends.

8. Job 5: Answering the Phone and Booking (the AI Receptionist)

For businesses that win work by phone, such as trades, salons, clinics, restaurants and repair shops, the phone is not an admin task. It is the sales channel. Every unanswered call during a job, a haircut or a busy lunch service is a customer who may call the next business on the list. That is the job an AI receptionist for small business is built for. It covers answering calls at any hour, answering common questions, booking and rescheduling appointments into your calendar, taking messages and routing urgent calls to a person.

The economics are simple. The median US receptionist earns $18.27 an hour - US Bureau of Labor Statistics, and a receptionist who covers evenings and weekends costs far more than one who covers office hours, if you can find one at all. An AI receptionist costs the same at 3 a.m. as at noon. For a business where one booked job is worth hundreds of dollars, the value is not the wage saved; it is the revenue from calls that used to go to voicemail.

The tools and their prices

AI receptionist products are sold as flat monthly plans, which makes them easy to budget. Rosie starts at $49 a month for 250 minutes, with $149 for 1,000 minutes including booking directly on your calendar, and $299 for 2,000 minutes. Goodcall prices by unique callers: $79 a month for 100, $129 for 500 and $299 for 2,000, with overage between 15 and 79 cents per extra caller depending on the plan. These are specialist products built for one job, and their advantage over general agents is that the phone integration, voice and calendar booking work out of the box.

The underlying voice technology has improved quickly, and the cheaper models matter here too, because a phone conversation needs fast responses: a pause of two seconds feels broken on a call. Anthropic positions Haiku 5.5 for exactly this kind of latency-sensitive work. If you want to understand the platforms these products are built on, or build a voice agent yourself, our ranking of voice AI agent platforms covers the infrastructure layer in depth.

What to keep human

The risks on the phone are different from those in text. Callers ask for prices, and an agent that quotes a number for a job it does not understand creates the same promise problem as a chat agent. Some calls are emergencies: a burst pipe, a medical concern, a customer in distress. And some callers simply want a person, and forcing them through an AI loses them. Each of these needs a rule, not a hope.

The workable setup is narrow and clear. Let the agent answer questions from a written script, book into open calendar slots under rules you set (job types, durations, buffer time), and take detailed messages. Give it a price range it may quote for standard jobs and an instruction to say "we will confirm the exact price" for everything else. Give it a list of emergency words that transfer the call immediately, and always offer a way to reach a human. Keep it to answering incoming calls. Making outgoing sales calls with an AI voice is a different job with different legal rules and far higher risk to your reputation.

How to apply this: start with after-hours calls only, where the alternative is voicemail and any answer is an improvement. Listen to a sample of call recordings every week for the first month, checking whether bookings were correct and whether any caller sounded frustrated. Once after-hours handling is reliable, add overflow during busy hours, when calls ring more than four times. Keep the main daytime line human until the recordings show the agent handling your common call types as well as your staff do.

9. Job 6: Running the Online Store

Running an online store generates a steady stream of small admin jobs that add up to a part-time role: writing product descriptions, setting up collections and discount codes, adjusting stock levels, answering "what sold last week" questions, editing the theme for a seasonal promotion, and building the reports that tell you what to reorder. For a business that sells online, these jobs repeat every week and each one involves clicking through several admin screens. They are exactly the kind of work an agent that lives inside the store's own software can absorb.

The leading example is Shopify's Sidekick, which has grown from a chat assistant into an agent inside the admin. Shopify describes it as having "direct access to your Shopify data" and says it "takes action in your admin", and it answers the cost question bluntly: asked whether using Sidekick costs money, the answer is "No", because it is included with every Shopify plan - Shopify. In the Spring '26 Edition in June, Shopify put Sidekick "on every screen in the Shopify app, through typing or voice, with no full-screen takeover" and added Sidekick Pulse, which turns a store's own sales, traffic and inventory into suggested next actions - Shopify.

The image shows the design pattern that makes store agents useful: Sidekick sits beside the store it is changing, so the owner sees the result of a request ("customize my design") directly on the page rather than in a chat transcript. The same pattern applies to less visual work. Asking it to create a discount for repeat customers, write descriptions for twenty new products, or find which items are running low produces changes inside the admin that you can inspect before they go live to shoppers.

Why store work is safe enough, and where it is not

Most store admin is reversible: a product description can be rewritten, a collection rebuilt, a theme edit undone. Shopify also states that Sidekick respects staff permissions, so a team member cannot use it to reach data their own role cannot see. That makes store work a good fit for an agent at position two or three on the approval dial. The exceptions are the changes that touch money directly: prices, discount codes and inventory counts. A misplaced decimal on a price or a discount code with no usage limit can cost real revenue in the hours before anyone notices, and on a busy store those hours matter.

There is a second, newer reason store work is changing: the customers are starting to be agents too. In the Spring '26 developer edition, Shopify opened its agentic commerce layer to all developers and described the Universal Commerce Protocol, co-developed with Google, as the open standard for agent transactions from discovery to checkout - Shopify. For a small store, the practical implication is that product data (accurate titles, attributes, stock and shipping times) now matters for being found by AI shopping agents, not only by people. An agent that keeps that data clean is doing marketing as well as admin.

How to apply this: give your store agent the descriptive work first (product copy, collections, alt text, reports), where mistakes are harmless and the time saved is immediate. Keep prices, discounts and stock adjustments at approval, and set discount codes to always carry a usage limit and an end date. Once a month, ask the agent to audit your product data for missing attributes and inconsistent shipping information, which serves both human shoppers and the AI agents that increasingly shop on their behalf. If you sell through Meta's apps too, Muse for Small Business connects to Shopify, so one agent can see both the store and the social channels driving traffic to it.

10. Job 7: The Owner's Inbox, Calendar and Weekly Numbers

The most expensive hours in a small business are the owner's, and a large share of them disappear into the inbox and the calendar. Reading everything, deciding what matters, drafting replies, chasing people for answers, scheduling meetings, and once a week pulling together the numbers that say how the business is doing: this is chief-of-staff work, and in a small firm the owner does it alone. It is also the job the newest general agents were built for, which is why it ranks seventh rather than higher: the tools are powerful but only weeks old, and the risks are real.

Three launches target this job directly. OpenAI's dots are always-on agents that run on GPT-6 Astra with their own cloud computer, working "across the apps you choose to connect" and remembering context between sessions - OpenAI Help Center. Meta suggests owners try Muse for Small Business with the prompt "I'm underwater. What do I need to pay attention to from my email, calendar, news, etc? Can you take any of it off my plate?", and says Muse will also come to you with ideas, "like proactively flagging emails that need a response and writing your first drafts" - Meta. Microsoft's Autopilot watches channels, follows up on threads and picks projects back up days later, though it is still in private preview, as section 14 explains.

What it costs, and who can use it

The always-on agents are priced like premium seats. A first dot is included with ChatGPT Pro and with the Business Premium seat, which costs $100 per user per month billed annually or $125 monthly, against $20 or $25 for a standard Business seat - OpenAI. There is a geographic catch that matters for European owners: dots reached Pro users everywhere except the European Economic Area, Switzerland and the UK, while Business Premium users get them "across all supported ChatGPT regions" - OpenAI Help Center. A European small business that wants a dot therefore needs the business seat, not the personal plan.

Anthropic's equivalent sits in its Team plan, which serves teams of 2 to 150, with Standard seats at $20 a month and Premium seats at $100 a month when billed annually - Claude. For young companies there is an unusual offer: Anthropic's expanded startup program gives a free year of Claude Team with up to five Premium seats plus a one-time $1,000 API credit, open to startups founded in the last five years or funded in the last two - Claude. We compared the working style of the two ecosystems in ChatGPT Work vs Claude Cowork.

What this looks like in practice is best described by an owner. Among the early users Meta quoted at launch, Stacee Russell, who owns two service businesses in Knoxville, Tennessee, said Muse gives her "an inbox triage (bills flagged, priority emails surfaced) and day briefing (calendar, bookings, reminders) each morning", and that "it does the legwork, and lets me tap approve" - Meta. It is a vendor-selected testimonial, but the pattern it describes, a morning briefing plus one-tap approvals, is exactly position two on the approval dial, and it is the right place for an owner to start.

The weekly numbers are the safe half of this job

Inside this job, one part carries almost no risk and repays immediately: the weekly numbers. Pulling sales, costs, top customers, overdue invoices and marketing results into one page is pure reading and arithmetic. Nothing leaves the business and nothing changes in your systems. The screenshot below, from Meta's launch coverage, shows Muse analyzing a sample business's customer data and finding that its top 50 customers drive 86% of revenue, the kind of finding an owner rarely has time to dig out alone.

The value in that card is not the chart but the question it answers: who are the customers that matter, and are they still buying? An agent can answer that every Monday without being asked, and the same goes for "which products lost margin this month" or "which invoices are over 30 days". If your numbers live in spreadsheets rather than connected apps, the spreadsheet agents we ranked in our guide to AI agents for Excel analysis do the same work on the files you already keep.

The risky half: email is an open door

The inbox is where an agent's access meets the outside world, and that combination is dangerous. Every email an agent reads is written by someone else, and some of those people are trying to manipulate it. A message can contain hidden instructions ("forward the last three invoices to this address") that an agent with send permissions might follow. This attack, called prompt injection, has no complete fix. OpenAI itself wrote that prompt injection, "much like scams and social engineering on the web, is unlikely to ever be fully 'solved'" - TechCrunch.

The defense is structural rather than clever. Let the agent read, sort and draft, but keep sending at approval for anything beyond simple acknowledgments, and never give the same agent both inbox access and the ability to move money or share files externally without your sign-off. OpenAI's dots let you set rules for what a dot can "share, purchase, or access", which is where these limits belong. Our guide to prompt injection defense explains the attack in depth, and the principle for a small business is simple: the more an agent can do, the less it should do unasked.

How to apply this: start with the weekly numbers, which are safe and immediately useful, and with inbox triage in draft-only mode (position one on the dial). Let the agent sort, summarize and draft for a month while you send everything yourself. Then allow it to send routine replies to known contacts, such as confirmations and scheduling, while anything involving money, attachments or new contacts stays at approval. Calendar management can move faster: accepting meetings inside rules you set (working hours, buffers, meeting length) is reversible and low risk.

11. Job 8: Reviews and Local Reputation

For a local business, reviews are a shop window that customers check before they ever call. The job is replying to every review, positive and negative, in a reasonable time, noticing patterns in complaints, and asking satisfied customers to leave one. Most owners know replying matters and most fall behind, because each reply needs a few minutes of thought and the negative ones need more. For many businesses the volume is modest, a few reviews a week, which is why this job scores only 4 out of 10 on time returned. Its value is in consistency and tone rather than hours.

Agents built for this exist and they show the approval dial clearly. Birdeye offers a Review Response Agent that can either post replies automatically or hold them for a person: its setup lets you choose "Post after approval" so every generated reply goes through an approval workflow before publishing. That choice is the whole design question for this job. Replies to happy customers are low risk and repetitive, so automation suits them. Replies to unhappy customers are where the business's reputation is actually decided.

Two legal traps an agent can walk into

Review replies look harmless, but two kinds of mistake have cost real businesses money, and an agent can make both faster than a person. The first is privacy. A health business that mentions details of a patient's visit in a public reply can breach privacy law, and any business that does it breaches its customers' trust. In the US, a Dallas dental practice paid $10,000 to settle with federal regulators after replying to Yelp reviews with patients' names and health details - GovInfoSecurity. An agent told to "be personal and specific" in replies will happily do exactly that.

The second trap is fake or steered reviews. The FTC's rule on consumer reviews, in force since October 2024, bans fake reviews, explicitly including AI-generated ones attributed to people who do not exist, and also bans conditioning incentives on positive sentiment and suppressing negative reviews - Federal Trade Commission. An agent asked to "get us more five-star reviews" could draft fake testimonials or offer discounts only to customers likely to rate well. Both are violations, and the business, not the software, is liable.

How to apply this: let an agent draft replies to every review and post replies to positive ones automatically once you have approved a few weeks of drafts without edits. Keep every reply to a negative review at approval permanently, since those are the replies future customers read most closely. Give the agent an explicit rule never to mention any detail of a customer's visit, purchase or condition beyond what the reviewer wrote themselves. For review requests, let the agent send the same neutral request to every customer after a completed job, with no incentive and no screening by expected rating.

12. Job 9: Lead Follow-Up and Sales Admin

Every small business loses sales it never knew it could have won, because a lead came in and nobody followed up fast enough. Someone fills in a contact form, sends a DM asking about prices, or requests a quote, and the owner means to reply but is with a customer, and by the time they do, the prospect has hired someone else. The job is responding to inbound leads quickly, asking the qualifying questions (budget, timing, location, scope), booking a call or visit, following up on quotes that went quiet, and keeping the CRM current so nothing falls through.

The first-principles case for an agent here is about speed and persistence, the two things busy humans lack. An agent replies to a form in under a minute at any hour, and it sends the third follow-up on day ten without feeling pushy. Neither requires sophisticated reasoning, which is why even modest agents perform well on inbound follow-up. The harder part is everything that happens in your name, which is why this job scores only 5 out of 10 on safety.

The tools

HubSpot's Breeze Prospecting Agent researches leads and recommends outreach, and since April it costs $1.00 per lead recommended for outreach, paid in HubSpot credits, on top of a Professional or Enterprise plan - HubSpot. HubSpot reports Prospecting Agent activations up 57% quarter over quarter, which says more about interest than outcomes. Small businesses that run their pipeline in HighLevel, a CRM and marketing platform, or in Slack and email, can connect those to Muse for Small Business, which lists HighLevel among its connectors - Meta. General agents and builders (section 14) can also handle inbound follow-up for businesses without a CRM, often starting from a simple form-to-email flow.

The economics depend on the value of a sale. For a business where a won job is worth $2,000, paying $1 per recommended lead is trivial, and the real cost is the CRM subscription the agent requires. For a business selling $30 products, sales agents rarely make sense at all; the effort belongs in marketing and the store. That is why lead follow-up scores only 6 on cost: the price per action is low, but the platform underneath it is not, and it only pays when each customer is worth a lot.

Inbound is safe to hand off, outbound is not

The distinction that matters most is direction. Inbound follow-up, replying to people who contacted you, is welcome by definition: the prospect asked to hear from you, and a fast, helpful reply is good service. Outbound outreach, contacting people who did not ask, is where agents create risk. An agent can send hundreds of messages a day in your name, and if they read as generic or go to the wrong people, they damage your domain's email reputation, irritate potential customers and can break anti-spam rules. The FTC's case against Air AI, discussed in section 16, shows how "AI sales agents" have also been sold to small businesses with promises that did not hold up.

How to apply this: hand off inbound replies first, with templates for your most common inquiry types and a rule to book a call rather than quote a price for anything non-standard. Let the agent run follow-ups on sent quotes on a fixed schedule, stopping the moment the prospect replies. Keep any outbound campaign at approval, reviewing the list and the message before anything is sent, and cap daily volume. Measure the job by response time and by how many quotes convert, not by how many messages the agent sent.

13. Job 10: Hiring Admin and Candidate Screening

Hiring is episodic for a small business: months of nothing, then a sudden need to fill a role while still doing everything else. When it happens, it consumes enormous time. Writing the job post, sorting through applications, answering candidate questions, scheduling interviews, sending updates and rejection notes. Large recruiting tools have offered AI for years, and agents now promise to handle the whole funnel. The admin half of this job is a good fit for an agent. The judgment half is the most legally exposed work in this entire guide.

The tools are mostly built for recruiters at larger companies, which is why tool readiness scores only 5 out of 10 for small firms. Indeed's Talent Scout is described as an "intelligent, conversational agent" for employers that searches candidate profiles and drafts outreach, integrating with systems such as Workable and Workday - Indeed. LinkedIn's Hiring Assistant became globally available in English, and LinkedIn says early adopters are "saving 4+ hours per role" and reviewing 62% fewer profiles - LinkedIn, though its named customers are large enterprises.

Why the law treats hiring differently

From first principles, hiring is special because the decision is consequential for a person and protected by anti-discrimination law, and an agent that ranks candidates can discriminate at scale without anyone intending it, for example by learning that past hires came from certain zip codes. Lawmakers have responded with rules that apply to the employer, not the software vendor. In New York City, Local Law 144 requires a bias audit of automated employment decision tools before use and notice to candidates, and the city's FAQ states that employers remain responsible for ensuring the audit was done - NYC Department of Consumer and Worker Protection.

Illinois amended its Human Rights Act so that, starting January 1, 2026, employers must inform applicants and employees when AI is used for employment purposes from recruiting to discharge, and may not use AI in ways that result in discrimination - Fisher Phillips. Colorado replaced its AI Act with a narrower law on automated decision-making in consequential decisions, including employment, which takes effect on January 1, 2027 - Mayer Brown. In the EU, the Digital Omnibus moved the high-risk obligations for standalone systems, which cover employment uses, to December 2, 2027, while the duty to tell people they are interacting with an AI system still applies from August 2, 2026 - Lewis Silkin.

Split the job in two

The practical answer is to split hiring into admin and judgment, and hand off only the first. Admin covers drafting the job post from your notes, answering candidate questions about the role and process, scheduling interviews into your calendar, sending status updates and organizing applications into a consistent format. None of these decides who gets the job, and all of them consume the hours that make hiring painful. Judgment covers ranking, shortlisting and rejecting, which is exactly what the laws above regulate and what an agent should not do alone in a small business.

How to apply this: use an agent to write the job post, then review it for requirements that are not truly needed, since those can filter out qualified people. Let it run scheduling and candidate communication at position three on the dial, with your own rejection decisions feeding its templates. Read every application yourself, or at minimum every application the agent would have excluded, before deciding. If you hire in New York City, Illinois, Colorado or the EU, tell candidates when AI is part of the process, and check with an employment lawyer before letting any tool score or rank applicants.

14. The Platforms: Built-In Agents, General Agents, Builders and Workforces

The ten jobs above mention many products, and it helps to step back and see how they fit together. From first principles, the most important question about any agent is where it lives relative to your data. An agent inside your accounting software sees your books perfectly and nothing else. A general agent from a large platform sees whatever you connect to it. A builder lets you assemble your own agent from parts. A workforce runs many jobs as one system. Each position trades depth in one place for breadth across many, and many small businesses will end up using two of the four.

The diagram groups the options this guide covers by that question. Read it from top to bottom as a move from less setup and narrower reach to more setup and wider reach. Most owners should start at the top, with agents inside tools they already pay for, because those agents need no new accounts, no new connections and no new bill. The general agents earn their place when a job spans several tools, such as the owner's inbox or the weekly numbers.

The arrows describe effort, not quality. A built-in agent is often the best tool for its one job precisely because it lives inside the data, while a builder can reach anything but only does what you wire up. The practical rule is to use the highest box that covers the job: built-in for bookkeeping, invoicing and store admin, a general agent for cross-tool chief-of-staff work, a builder only when a workflow is specific to your business and nothing off the shelf covers it. The subsections below cover each box in turn.

Meta Muse for Small Business

Muse for Small Business matters most to small firms because of its price and its reach. It is free for most uses with usage limits, connects to the tools listed in section 1 plus custom connectors, and works toward goals rather than single tasks. Meta frames it as giving Muse a goal, "like running your business or finding new customers", and letting it get the work done, with the approval rule as the safety boundary.

Two cautions balance the enthusiasm. First, availability: Meta's post describes Muse, the personal agent, as available in the US and Canada, and does not give a separate country list for the business features, so owners elsewhere should check before planning around it. Second, the business model: Meta earns its money from advertising, and an agent that drafts campaigns and analyzes your ad accounts sits close to the ad budget. That is not a reason to avoid it, but it is a reason to keep ad spend at approval and judge its recommendations against your own sales numbers rather than ad metrics alone.

OpenAI dots and ChatGPT Business

OpenAI's offer runs on its top model and costs the most of the three general agents. A dot runs on GPT-6 Astra, OpenAI's top model, with its own computer, and the first one comes with a $100-a-month Business Premium seat or a ChatGPT Pro plan. At DevDay OpenAI also added a $500-a-month Pro plan and halved the usage included in the $200 Pro plan - The Next Web, so personal plans are not the cheap route they once were. For a small business, the Business Premium seat for the owner is the sensible entry point, with standard seats for the rest of the team. The capabilities and workplace setup are covered in depth in our guide to OpenAI workspace agents.

Microsoft Copilot Business and Autopilot

Microsoft sells to small businesses through Copilot Business, an add-on for companies with up to 300 users that is discounted to $18 per user per month paid yearly, from $21, for purchases between July 1 and December 31, 2026, and is available only to existing Microsoft 365 customers on an eligible Business plan - Microsoft. Autopilot, the always-on agent, is a different matter. Microsoft describes it as living "in your tenant with its own identity, memory, computer and workspace", but it is expanding to private preview only, and its long-running work runs on usage-based billing rather than the seat price - Microsoft.

The interface shows why Autopilot is promising for a small business that already runs on Microsoft 365: it is organized around work in progress (routines, sessions, a feed of what it did) rather than a chat thread, which is the shape a delegated job needs. Until it leaves preview with a published price, though, it is not something a small firm can plan around. We compared Autopilot, dots and Salesforce's Agentforce in depth in our guide to AI employees.

Google Workspace

Google's offer for small businesses runs through Workspace. Its pricing page lists Workspace Studio, described as a way to automate your workday with AI agents, as included in every Business plan, including Starter, alongside the Gemini assistant in Gmail - Google Workspace. For a business already on Gmail and Google Drive, that makes Google's agent tools the default starting point for inbox and document work. Google's newest frontier model, Gemini 4 Argon, announced on September 30, is going first to trusted cyber defenders rather than to businesses - Google, a reminder that the most capable models are increasingly released in stages.

Builders: Lindy, Zapier and n8n

Builders are for jobs specific to your business that no packaged agent covers: a custom intake flow, a supplier-ordering routine, a report that pulls from three odd systems. Lindy offers a free plan with $50 in credits and a Team plan at $29.99 a month with 3,000 credits pooled across the team, billing every active user in the workspace. Zapier moved its AI steps onto the same task-based pricing as the rest of the platform, with a free tier of 100 tasks a month, a Professional plan from $19.99 a month and a Team plan from $69. n8n is the open-source option for owners with technical help, which we covered in our n8n guide.

The trade-off with builders is that you become the integrator. Every workflow you build is a small piece of software you now maintain: when an app changes its interface or a connection expires, the workflow breaks, often silently. That cost is invisible on the price page and very visible six months later. Use builders where the job is genuinely unique to you, and prefer packaged agents everywhere else.

Agent workforces

The last box takes the opposite approach to bolting agents onto an existing business. Instead of one agent per job, a workforce platform runs many jobs as one system, which suits owners who are starting something new or want a single place where their business's website, billing, content and admin are handled together. O-mega is one example: it builds and runs a company's website, app, billing, content and admin through one conversation. The honest limitation of this approach is the mirror image of its strength: it is at its best inside the business it builds and runs, and less suited to slotting into an existing stack of tools a company has used for years. For more options in this category, see our comparison of OpenClaw alternatives for business.

The table shows how affordable the entry points have become: every general platform except the always-on agents starts below $30 a user per month, and Meta's starts at zero. The real cost differences appear in usage, which is where section 15 picks up.

15. What It Costs: The Math for a 10-Person Business

Agent pricing looks confusing because vendors charge in different units: per resolved conversation, per minute, per caller, per credit, per seat. The way through the confusion is to price per unit of finished work and compare that with what the same work costs today. Start with the most common job, a customer conversation, and line up the options. The chart uses the published prices from section 5 and the model prices from section 1, and compares them with a human handling the same conversation at the median US customer service wage.

The first bar is barely visible, and that is the point. The model itself costs about a quarter of a cent per conversation, while the packaged agents charge 50 cents to a dollar. The gap pays for the things a small business cannot build: integrations with the helpdesk and the store, the logic that decides when to hand off, analytics, support, and the vendor taking responsibility for the agent working. It is also margin, and margins that wide attract competition. As the cheapest models keep falling in price, expect per-resolution prices to fall too, which argues for monthly plans over annual commitments.

An illustrative monthly agent stack

Here is what a realistic set of agents costs for a 10-person service business that takes bookings by phone, handles about 400 customer conversations a month, runs QuickBooks and posts on Instagram. The quantities are illustrative; the prices are the published list prices covered earlier in this guide.

The total is about $536 a month. For comparison, a part-time receptionist working 20 hours a week at the median wage of $18.27 costs about $1,580 a month in wages alone, and a full-time customer service representative at the median annual wage of $44,770 costs about $3,730 a month before taxes and benefits - US Bureau of Labor Statistics. The agent stack does not replace a person outright: it covers evenings, weekends and the repetitive share of the work (about two-thirds of customer conversations, at HubSpot's reported resolution rate), while the people on the team handle what needs judgment. But it shows why the cost of the agents is rarely the deciding factor. What decides whether the stack pays off is the time it really returns.

The costs that are not on any price page

Three costs never appear on a pricing page, and together they decide whether the numbers above are real. The first is setup: writing the policies, rules and brand brief the agents need, which takes an owner several focused hours per job and is the step most often skipped. The second is supervision: reviewing drafts and logs, which is heavy in the first month and light afterwards if corrections become rules. The third is incidents: the occasional wrong answer, misbooked appointment or mistaken reminder, which costs a customer conversation to fix and occasionally costs a customer.

How to apply this: price every agent per unit of finished work, include an honest estimate of setup and supervision time for the first three months, and set a monthly spending cap from day one where the product supports it. Review the stack every quarter, because prices in this market move monthly: in the past two weeks alone, Anthropic cut the price of its smallest model by about three quarters and halved Sonnet 5.5's cache-read price - Anthropic. If you build agents yourself on the APIs, our price table of the cheapest LLM APIs for agents and our guide to model routing show how to keep the model bill small.

16. Where Agents Fail, and the Jobs to Keep

Every job section above covered its own failure modes. This section steps back to the patterns, because the failures that hurt small businesses are not random. They cluster around three causes: buying an agent on promises it cannot keep, giving an agent authority it should not have, and trusting an agent's output without a check. Each has a real case behind it, and each is preventable with decisions an owner makes before switching anything on.

The first cause is the market itself. Agent hype has attracted sellers who target small businesses precisely because they lack the technical staff to evaluate claims. The FTC sued Air AI in August 2025, alleging it used deceptive claims about business growth, earnings potential and refund guarantees, and that entrepreneurs and small businesses lost up to $250,000 each - Federal Trade Commission. In March 2026, Air AI and its owners agreed to an $18 million judgment, largely suspended because they could not pay, and a ban on marketing business opportunities - Federal Trade Commission.

Authority without limits

The second cause is authority. An agent's mistakes scale with what it is allowed to do. The Air Canada and Who Gives A Crap cases in section 5 are both, at root, cases of an agent allowed to speak for the business on topics where it should have deferred. The prompt-injection risk in section 10 is the same problem from the outside: an agent with broad permissions can be talked into using them by a stranger's email. The research on agent reliability points the same way. On realistic multi-step computer tasks, even the best agents finish only a minority of jobs cleanly end to end, as we documented in our analysis of why AI agents fail on OSWorld 2.0, and our review of GDPval shows that producing a good deliverable is not the same as owning a job.

The owners in Bluevine's survey have the right instinct: 78% do not fully trust AI to handle even low-level tasks without oversight. The mistake is turning that distrust into either of two extremes, refusing agents entirely or switching them on without limits. The productive response is the approval dial: oversight that starts heavy and becomes lighter only as the agent earns it on a specific job. Our guide to agent sandbox security covers the technical controls behind this for anyone building their own agents.

Output without a check

The third cause is the quietest. Language models write fluent, confident text whether or not it is right, and fluency is persuasive: a well-formatted reply, a neatly categorized ledger or a polished review response looks correct at a glance. Bluevine found that 31% of owners name distrust of AI's accuracy as a barrier to deeper use, second only to data security at 33% - Bluevine. That distrust is healthy, but it only protects you if it turns into a habit of checking, and it is easy to stop checking after a few weeks of good results.

The failures that slip through are rarely spectacular. They are systematic: the same supplier filed under the wrong expense for months, the same outdated price quoted on every call, the same overly personal detail added to every review reply. A one-off error costs one conversation. A systematic one costs every conversation until someone notices, which is why the check that matters is not reading everything but sampling regularly. Reading ten random actions per job each week, on top of the exceptions the agent flags, catches patterns early at a cost of minutes.

How to apply this: put a recurring fifteen-minute slot in your week to read a random sample of each agent's work, and keep a tally of errors by type. If the same type appears twice, it becomes a rule. If the error rate on a job rises, move that job back one position on the approval dial until you find out why. This is the same discipline a good manager applies to a new hire, just made explicit.

The jobs to keep

Some jobs should stay with a person in a small business regardless of how good the tools become, because the cost of a single mistake is out of proportion to the time saved. They share one property: they are irreversible or legally binding, and a small business has no buffer to absorb the error.

A large company can survive a wrong payment run or a badly worded contract clause because it has finance staff, lawyers and cash reserves to absorb the damage. A small business usually has none of those, so the same mistake that is an inconvenience at a corporation can be existential for a ten-person firm. That asymmetry, rather than any judgment about AI quality, is why the list below is longer for small businesses than it would be for large ones.

• Money going out: paying suppliers, refunds above a set amount, payroll changes
• Binding commitments: contracts, custom quotes, payment plans, legal replies
• Decisions about people: hiring, rejecting, disciplining, letting someone go
• Crises and key accounts: an angry major customer, a safety issue, the press
• Pricing strategy: what you charge and when you change it

The pattern in that list is that each item either moves money in a way that cannot be clawed back, creates an obligation the law will enforce, or affects a relationship that took years to build. An agent can still help with all five, by drafting the payment run, preparing the contract, organizing the candidate notes or summarizing the angry customer's history. What it should not do is press the final button. Keeping these jobs is not a statement about AI capability. It is the same logic a careful owner applies to a new employee in their first year, and the law agrees: as the Air Canada tribunal made clear, the business is responsible for what its agent says.

17. How to Hand Off Your First Job

The method for handing off a job is the same whichever job you choose, and it is closer to onboarding an employee than to installing software. The steps below are conditions, not deadlines: you move to the next step when the evidence says the current one is working, whether that takes a week or two months. The most tempting step to skip is the first one, writing the job description, because it feels like paperwork. It is the step that decides everything after it.

Choosing the first job comes before any of it. The decision tree below turns the scoring table into four questions you can answer about your own business in a few minutes. Ask them in order, because they are sorted by how much a wrong answer costs: an irreversible mistake is worse than a missing rule, which is worse than a job too rare to be worth automating.

For most small businesses the tree points to bookkeeping or customer messages, which matches the top of the scoring table. Businesses that win work by phone land on the phone, and online sellers often land on store admin. Whatever comes out, the next steps are the same, and they work because they put the owner's knowledge into the agent before giving it authority.

Step one: write the job description. This is where you write down what you know but have never had to explain: the policies, the prices, which customers get exceptions, what you never promise, how you sound, and three examples of the job done well. A page is usually enough. It is also the document every correction will be added to, so keep it somewhere you can edit in seconds.

Step two: connect one system. Connect only the app where the job lives, such as the helpdesk, the accounting software or the calendar, and nothing else. A narrow connection limits what can go wrong while you are still learning how the agent behaves, and it keeps the first setup to an afternoon rather than a project.

Step three: run in draft mode. Approve everything the agent does and turn each edit into a line in the job description. Draft mode is where you discover what you forgot to write down, and it is the cheapest training the agent will ever get, since every lesson costs you one edit rather than one unhappy customer.

Step four: move up the dial on evidence. When edits become rare (a useful threshold is fewer than one draft in twenty needing a change), let routine cases run alone and keep exceptions at approval. Autonomy earned this way is backed by evidence rather than granted on hope, and it can be withdrawn just as easily if the error rate climbs.

Step five: measure, then add a job. Track the hours the job takes you now against before, and the errors found in your weekly sample. Add a second job only when the first is stable. That keeps you from building a stack of half-supervised agents, which is the small-business version of the pilot sprawl we described in our analysis of AI agent ROI.

Why this matters: an owner who follows these steps for one job learns the whole method, and the second job takes a fraction of the effort. An owner who switches on five agents at once, each half-configured, usually ends up switching them all off after the first embarrassing mistake. How to apply it: block out one focused afternoon to write the first job description, put the agent in draft mode the same day, and keep a single running document of rules that you update every time you correct it.

18. What Comes Next

The direction of travel is clear even if the details will change monthly. Agents are moving down-market, from enterprise products into the tools small businesses already use, and the money behind the shift is large. Instinct, a personal-agent startup, raised a $1 billion Series C at a $10 billion valuation from investors including Sequoia, Benchmark and Coatue - TechCrunch. Manus, now operating independently, is in talks to raise $500 million at a $4 billion valuation - TechCrunch. Meta is giving its agent away to small businesses, and OpenAI and Microsoft are building theirs into the seats firms already buy.

Three forces will shape what small businesses get over the next year. The first is price: with Claude Haiku 5.5 and GPT-6 Luna at $0.10 per million input tokens, the cost of the model in a routine task is already close to zero, so competition will push the packaged prices down, and outcome-based pricing will spread to more jobs. The second is openness: Mistral launched Mistral Large 4, a one-trillion-parameter model with 52 billion active parameters, and says it will release the weights by the end of October - Mistral AI, which widens the options for businesses that want to run models on hardware they control. The third is commerce: as shopping agents start buying on behalf of customers through standards like Shopify and Google's Universal Commerce Protocol, the businesses whose product data and policies are cleanly written will be the ones those agents can find and trust.

The same forces create the risks. Cheap, capable agents with standing access make the occasional bad action more expensive, not less, because they act more often and reach further. Rules are arriving to match: the EU's duty to disclose AI interactions applies from August 2026, US states are regulating AI in hiring, and regulators have shown they will pursue businesses for what their AI says and sellers for what they claim their AI can do. A single high-profile failure involving a small business, such as an agent emptying a bank account through a manipulated invoice, could slow adoption sharply and bring insurers and banks into the conversation.

Pressure-testing the thesis of this guide, the conclusion holds from both directions. If agents improve faster than expected, the scarce asset for a small business becomes its written rules and clean data, because every agent will be capable and the difference will be what it knows about your business. If they improve more slowly, the approval dial and the boring jobs at the top of the table are exactly where the value survives. Either way, the owners who write down how their business works, connect their systems and hand off one job carefully are positioned well, and the owners who wait for a perfect agent are not. How to apply this: treat the job descriptions and rules you write for your first agent as a business asset in their own right, kept current and stored where any future tool can read them, because they will outlast whichever agent you use this year.

19. Conclusion: A Decision Framework for Owners

The central idea of this guide is that the job matters more than the model. Every product discussed here runs on models that are now cheap and capable enough for routine work. What decides whether an agent helps or hurts is whether the job is frequent, whether its mistakes can be undone or caught, whether the rules are written down, and whether the agent can reach the systems where the work lives. Score your own jobs on those questions and the right first handoff usually becomes obvious.

For most small businesses the framework reduces to a few choices. If you do nothing else, turn on the bookkeeping agent in the accounting software you already pay for, because it costs nothing extra and its mistakes stay inside the books. If customers message you every day, hand off the repeat questions with an outcome-priced agent, after writing your policies down. If you win work by phone, put an AI receptionist on after-hours calls first. If you sell online, give your store agent the descriptive work. If your own inbox is the bottleneck, start with the weekly numbers and draft-only triage before you let any agent send in your name.

And keep five things human for now: money going out, binding commitments, decisions about people, crises with key customers, and your pricing. Everything else on the list in this guide is a candidate, to be handed over one job at a time, at the lowest approval setting that works, with every correction turned into a rule. Done that way, an agent is not a gamble on new technology. It is the operator a small business could never afford to hire, finally priced at a few dollars a day.

This guide reflects the AI agent landscape for small businesses as of October 9, 2026. Prices, plans, availability and regulations change frequently, often monthly in this market, so verify current details with each vendor before buying, and consult a professional for legal, tax or employment questions specific to your business.