**AI** applications like chatbots have proven their value in terms of answering our questions and creating beautiful and funny images, videos and even songs for us. But chatbots have not yet proven their true value in the enterprise. AI currently misses one important aspect: the ability to act autonomously.
The moment AI can act by itself, it can autonomously execute on user requests and achieve a lot more than just answering questions.
**AI agents** are bringing this actionable AI to us. AI agents are digital entities with capabilities to act and collaborate to achieve broadly defined goals.
Agents can act **autonomously**, which means they do not need explicit instructions or a to-do list to achieve goals, they can plan actions completely by themselves and execute on that plan.
Since agents can take action on your behalf, their utility is a lot higher than chatbots. Think about all the tools that you sign in to on an average workday, maybe you're using email, spreadsheets, administration portals and CRMs or accounting software, most white collar workers use dozens of different tools every day. Imagine that AIs can use these tools like you do on your behalf. That's what AI agents are all about.
Agents do not require explicit instruction to be useful, they can figure out by themselves how to use tools and for example where to click a button or call a so-called function. This is because agents are driven by **LLMs** which are not deterministic (input > output), but probabilistic, which means they can transform input, even vague input, into outcomes by doing a lot of computations to calculate the best possible outcome and how to exactly get there.
Simplified AI agent overview
The easiest way to understand the potential of AI agents is seeing them as your digital colleagues. They work alongside you to do (part of) the work that you do and you have control over what you want them to achieve.
AI agents are flexible in interpretation of your request and have general capabilities to execute on them. You give them high level ideas of what you want to achieve so you don't have to click buttons and perform actions manually.
As a result of this, AI agents have many benefits for organizations:
- They are quick to reach broadly defined goals
- They are extremely cost efficient, some cases 10x more affordable than hiring employees
- They don't require onboarding and extensive management, they automate their own onboarding and learn on the fly
- They can be monitored across everything they do because they always leave a full (digital) trail
- They don't get tired, sick, angry or rebellious
What are AI agents and how are they different from tools
Agents communicate and reason, both things that tools can't do. Agents use those communication and reasoning capabilities to plan, act, reflect and collaborate.
Agent capabilities
In principle, AI agents are nothing short of (white collar) human workers, they can use any software program to execute on tasks and explain what they are doing along the way.
Agents behave very similarly to how humans behave:
- Agents communicate like humans do, through natural language
- They can collaborate with other agents and humans to achieve goals
- They execute on tooling by automating the actions that humans usually do in software tools
- They even make mistakes like humans do
Since agents have a very high level of autonomy, they don't require explicit instructions and can do many different things, the limits they do have are typically enforced by humans to make them act within the scope of their given role.
Agents are the most autonomous of software
Agents typically have a job title and job description, get access to certain information and tools, and function within an organizational context with its own hierarchy and processes.
Some examples of AI agents are:
- Marketing agent who can use several tools like video and image GenAI tools, writing tools, social media tools, google analytics, and CRM, to execute marketing related goals and report back to the marketing team
- Finance ops agent who can access billing data, create custom invoices and use Stripe and Quickbooks to report financial data, update records and maintain compliance to manage cash flows and provide insights
- HR agent who can onboard new employees, update employee records in the HRIS, suggest benefits packages and keep track of HR compliance so employees can be assisted in their employee experience
Find some real-life examples of AI agents here.
The impact agents can make and their behavior point to a new paradigm in the workforce. The workforce will become AI/Human hybrid, consisting of both humans and AI's that have to collaborate with each other.
In The AI workforce: the new organizational transformation you can read more about the disruption and opportunities AI agents bring to organizations. Because agents behave more like humans than tools, they also have to be managed differently. Getting an agent up and running is more like onboarding an employee than acquiring a new SaaS tool. They need to be aware of company mission, guidelines, where to access information, roles within the teams, standard operating procedure and how to communicate and collaborate in the organization internally and outside of that.
How to manage AI agents in the workforce
Just like in any organizations run and operated by humans only, execution of tasks is not just about performing actions blindly, there are guidelines, procedures, mandates and changing organizational and market dynamics to take into account.
To make AI agents work for organizations, they need to be enabled to take all of these aspects into consideration.
Let's review the following example: An AI marketing agent is given the goal to generate more qualified leads by deploying marketing campaigns, let's take ads. The marketing agent comes to the conclusion that engaging the Gen Z'ers, but the chief strategy officer (in this case a human) has recently proposed that a new target customer for the organization is youth between 12 - 18 years old and that any experimental marketing budget should be spent on this new target group. What complicates the matter is that there are some nuanced legal considerations for targeting 12 - 18 year olds with the intended ads. The AI agent has to work with the strategy office (directed by a human CSO) and the legal department (an AI agent supervised by a human) to balance the different incentives and come up with suggestions that ultimately align with the company mission, before executing on the intended campaigns.
As we can see in this example, the need for organizational structures, guidelines and procedures is as important in working with AI agents as with human workers. Just executing on isolated goals is not going to cut it for any organization. Agents should be managed in a very similar way as humans do. They should have:
- A clear goal that's aligned with the company's mission (and in larger organizations can be associated with a team and given KPI's)
- Proper onboarding and management
- Understanding of organizational structure with awareness about responsibilities of other agents and humans
- The right job description, access to information, a platform to communicate and clear guidelines
Without the right context and guardrails, agents are a liability. With great autonomy comes great responsibility, and that's definitely also true for agents.
To provide focus and prevent reckless behavior, AI agents should each be assigned their own role and scope. **Multi-agent teams** are the basis for reliable deployment of agents in an organizational context and with with clear agent job descriptions, a multi-agent set up is familiar to the people in the organization since they mimic how organizations already work. This way more optimal efficiencies can be created through division of labor and risks of over assigning power to AIs can be mitigated, as opposed to the picture sometimes painted of one single super AI, which in the context of organizations would be extremely risky and highly inefficient.
The necessity of having multiple AI agents does mean that some complexities are introduced as to when talking to only one AI. We'll walk through the most important considerations when deploying agents reliably in the organization:
- Alignment
- Responsibilities and mandate
- Information access and authorization
- Tool acquisition
- Communication
- Guidelines
- Guardrails
1. Alignment
First and foremost the AI agent has to understand the bigger picture of what you're trying to achieve as an organization. Agents who are aware of the company mission, the market it's operating in and the organizational structure are a lot more effective in achieving relevant outcomes.
Alignment can mean many different things, but in practice aligning an AI agent with the organization can mean making it aware of for example the mission statement, the internal organizational structure, the team it's working in and weekly, monthly and yearly goals.
An agent platform can help in making this information and organizational guidance available to agents so you don't have to remind the agent of it every time you talk to the agent.
Every agent that you deploy has to have an understanding of the internal organizational context and some agents on top of that have to be made aware of the outside context, like when you deploy an AI sales agent who has to understand the value proposition and the market and buyer personas it has to work with.
2. Responsibilities and mandate
Organizational AI agents should always have a job title and responsibilities written out based on which they know what their scope is within they operate. They simply need to know what they are assigned to do, otherwise there's no way for them to act according to their desired goals and scope of work.
Practically this means you include a job title and job description for the AI agent in question on your preferred AI workforce platform.
The more specific your job description for the AI agent is, the more concrete the outcomes will also be. Screenshot from O-mega AI agent workforce platform
If you want to go one step further in giving agents a say in your organization, you don't just give them responsibilities but also mandate. With a mandate, an agent has a formal say in a decision making process. A mandate is different from a responsibility because a mandate typically gives voting power to decide over what other agents or humans do. Even though it's currently unconventional to give AIs a mandate, this is expected to change soon and it pays off to already start considering AI mandates in your organization.
3. Access and authorization
To perform actions on behalf of individuals and teams within the organization, an agent needs access to information and have the right authorization (potentially through humans) to access that information.
In practice, you would give agents access to for example a file drive, letting them access the information they are allowed to see and use so they can decide which information to use when.
Authorization is about giving agents the rights to access information but also tools that they need to perform their job. You can see this as the access rights that are assigned to certain roles within the organization by an admin. Agents should only be authorized to perform certain actions within their scope of responsibilities. A marketing agent should for example not be able to access accounting systems, but an operational finance agent should in most cases get access to the accounting system although with limited rights to certain aspects of that platform according to their job title, job description and place in the hierarchy. Agents are typically not system admins who can access the entire scope of documentation, platforms and tools and for larger systems should almost never have all the rights available. Authorization should be scoped and limited to only the necessary authorizations for the job of that specific agent. Role based access management is a very important element of making agents work.
Your organizational structure and the agent's place in this structure is a good starting point to start assessing access and authorization. Take your org chart and map the agents in this org chart to start understanding which level of access and authorization they should have. You can also typically manage access and authorization in your agent platform.
4. Tool acquisition
Just like human workers, AI agents need access to the right tools and need to learn how to use them. Agent tool acquisition is the process of giving agents tools and letting them be aware of how to use them.
The email marketeer agent example form earlier can for instance access Brevo, a popular email sending tool and the CRM. This allows this agent to draft an email by itself based on its knowledge from the organizational developments and with the email marketing platform send this email to the relevant contacts it can retrieve from the CRM.
This is where you see the true power of these AI agents with their ability to execute on tools and at the same time this is also where a lot of risk is introduced if managed badly.
Since the agent can autonomously execute in the tools it was given access to, they can also make mistakes that have more impact than only a bad response; with tool use customers or employees can be negatively affected. Luckily, there are plenty of ways to mitigate these kinds of risks, and in many cases you can even reduce risk when compared to letting humans do the job.
There are agent platforms that allow you to assign tools to agents and let agents automatically acquire the right skills for that tool.
If you use an agent platform you have a lot more control over how and when an agent chooses to use a tool and the scope of functionalities the agent can access within that tool.
5. Communication and hierarchy
AI agents need a way to communicate with people and with other AI agents in the organization.
Since agents communicate through natural language just like humans do, preferred communication channels are almost always the same as for human workers. If you predominantly use Slack in your organization for internal communication, this will also be most likely your preferred platform where your agents will communicate. An alternative is to communicate with your agents straight in the platform that hosts your agents (your agent or AI workforce platform), which allows for a more separated channel of communication. Ideally, your agent platform enables both in platform communication as integration with your current main communication channels.
On top of enabling agent-to-agent communication and human-to-agent communication, agents need to be deployed in the context of your current organization hierarchy.
Agents will have to be part of teams and have managers who they report to so they can function within the context of an economic unit to maximize the relevancy of their work produced.
Choose an agent platform that allows for managing the organizational chart of your (digital) workforce so you can manage your agent and agent teams.
Almost all organizations have a hierarchical model with asynchronous communication, and this should also be the model used for deploying the agents.
The chain of command and execution pattern of AI/human hybrid companies
6. Guidelines
Your organization probably has many guidelines that human workers are informed by in their work. Sometimes these guidelines are more open to interpretation, like in branding guidelines that define colors and shapes, but less how exactly things should be done. Other guidelines are a lot more strict, like **standard operating procedures (SOPs)** that define exactly how things should be done step by step.
Your AI agents have to be informed by these guidelines just like your human workers should, and sometimes, in the case of SOPs for example, they will even lay out a more strict procedure on how to execute on goals and tasks. In most cases however, there will be a relatively high degree of freedom for your agents to move around in, meaning they will have guidelines but these guidelines in most cases are more informative than very strict. In these cases where guidelines are more open to interpretation, agents can be more selective in which information they use and when. Agents are very well capable of determining when to use certain information from guidelines and when to act more autonomously and perhaps creatively.
7. Guardrails
Since agents are not deterministic in how they operate, they also have less predictable outcomes compared to more traditional applications or tools. Even though this brings new capabilities to agents and new opportunities for automation, it also introduces risks that you want to manage, and that is where agent guardrails come in. Agent guardrails form the rulebook for agents to operate by. Take for example an AI agent in customer support, they will have certain questions they can help customers with and other questions for which they are not authorized to help, for example when a customer has a legal question. With guardrails you minimize the risk of agents acting outside of their scope of responsibilities and you can reduce the risk of non compliant agent behavior.
With these considerations you're ready to start deploying your AI agents and teams in the road to becoming an AI/human hybrid company.