Guides · AI adoption
What an agent touches, and where it stops.
Documents
Searches internal documentation and answers citing its source.
Email
Drafts and sends, within the limits you set.
Data
Reads from the CRM, a spreadsheet or a business system.
An approval step
A person reviews before the action happens.
AI agents are already in production at large companies: among organisations with over $1bn in annual revenue, the share scaling agents in one or more functions went from 27% to 40% (McKinsey, The state of AI 2026). At smaller organisations the figure holds at 22%, so the useful question is no longer whether to use AI agents but on which task, and with what control.
What are AI agents?
AI agents are «software systems that use AI to pursue goals and complete tasks on behalf of users», with reasoning, planning, memory and a degree of autonomy to decide, learn and adapt.
In practice an agent takes a goal, breaks it into steps and uses tools to complete them: it looks up data, drafts a document, sends an email or updates a system. The difference from using Claude, ChatGPT or Microsoft Copilot in a conversation is that the agent does not wait for you to name each step.
For a director of Operations, HR or L&D, the most useful definition is an operational one: an agent is a delegated task with a goal, a set of permissions and someone responsible for reviewing what it does.
How is an AI agent different from a chatbot or an assistant?
An AI agent acts proactively to meet a goal, an assistant responds to what you ask, and a bot follows predefined rules (Google Cloud, What are AI agents? 2026). The difference is the degree of autonomy, and who makes the final call.
| Criterion | AI agent | AI assistant | Bot |
|---|---|---|---|
| How it acts | Proactive, goal-driven | Reactive, responds to requests | Reactive, responds to triggers or commands |
| What it can do | Complex multi-step actions; learns and adapts | Answers, informs and completes simple tasks | Follows predefined rules, basic interactions |
| Who decides | The agent, within the limits you set | The person: the assistant can recommend, you decide | The rule, written in advance |
| Fits | Whole workflows | Help with one-off tasks | Simple, repetitive tasks or conversations |
The practical consequence: a good share of what is presented today as an "agent" is solved by an assistant used well, or by rule-based automation. Picking the right level makes the project cheaper and the risk smaller.
What do companies use AI agents for?
Companies use AI agents mostly in IT, knowledge management and software engineering — the functions where respondents most often report scaling them (McKinsey, The state of AI 2026). Around 2 in 10 respondents say their organisation has reached the scaling stage with AI agents.
The pattern shifts by sector, per the same report: technology companies lead in software engineering, consumer and retail in marketing and sales, and advanced manufacturing in supply chain and production.
These are tasks that fit an agent, by area:
- Knowledge management: search internal documentation and answer citing the source document.
- IT: triage incoming tickets and propose a fix to the technician handling them.
- Sales and marketing: prepare an account summary before a commercial meeting.
- Finance and admin: match invoices against orders and flag discrepancies for review.
Non-technical teams are growing fastest
Agent use is no longer confined to technical teams. Since February 2026, weekly active users of Codex, OpenAI's agent, grew 108x in legal, 41x in sales and recruiting and 26x in marketing among its enterprise customers, against 5x in engineering (OpenAI, From assistance to execution 2026). These are growth rates on a small base, measured only among OpenAI customers, but they show where the use is moving.
For those teams the challenge is not technical: it is deciding which task to delegate and at what level of autonomy, which is the criterion the next sections set out.
When an AI agent makes sense (and when something simpler will do)
An AI agent pays off when a task meets 4 conditions at once: it has several steps a person chains by hand today, it needs to look up data or use tools such as email, the CRM or a spreadsheet, it repeats often enough for the time saved to show, and it tolerates a margin of error a person can review.
If the task is one-off, an assistant is enough. If it always follows the same path, rule-based automation is cheaper and easier to maintain.
In our AI consulting we grade every task on a three-level scale: AI executes and a person reviews, AI accelerates and a person validates, or a person leads and AI assists. That scale is a quick way to decide how much room to give an agent before building it.
How do you govern an AI agent?
An AI agent is governed with controls proportional to its level of autonomy, not with one policy for every agent. Gartner expects that by 2027, 40% of companies will downgrade or retire autonomous AI agents over governance gaps that only surface after incidents in production (Gartner, press release, May 2026).
Gartner proposes four levels of autonomy, each with its own controls:
| Level | What the agent does | What the person does |
|---|---|---|
| 1. Observe | Reads defined data sources without changing them — summarises documents, retrieves information. | Uses the result, which only the requester sees |
| 2. Recommend | Produces recommendations, drafts or proposed actions — writes emails or reports. | Reviews and carries out the action by hand |
| 3. Act with approval | Writes data, sends communications or changes settings | Approves each action before it happens |
| 4. Act autonomously | Acts on its own, within defined limits | Reviews exceptions, logs and aggregate results |
Source: Gartner, press release, May 2026.
Gartner warns that applying the same controls to every agent over-restricts the simple ones and under-controls the autonomous ones. The practical rule is to start at level 1 or 2 and move up only once you have data on how the agent behaves.
Governance also means the basics: a usage policy and a list of approved tools. You can check what evidence you are missing with a quick test on Article 4 of the AI Act.
Where to start with AI agents in your company: 4 steps
- Measure where the time goes. Before choosing a tool, identify the tasks that eat the most hours and how often they repeat. The savings calculator gives an indicative estimate of the hours your team would recover and what they are worth.
- Pick one task and its autonomy level. Apply the 4 criteria above and give the task a level from 1 to 4. A first level-2 agent, which proposes and leaves the decision to a person, gives you real data without exposing critical processes.
- Train the team that will work with the agent. The EU AI Act asks companies deploying AI systems to take AI literacy measures for their staff, an obligation in force since 2 February 2025 (European Commission, AI Literacy: Questions & Answers, Jul 2026). It does not require a certificate: internal training records are enough. In our practical workshops, each team takes away a replicable demo agent built on its own use case. It is a demo for learning how to build one, not a deployment into your systems.
- Measure before you scale. Compare the hours the task took before the agent against the hours 60 to 90 days later. If the result improves and errors are under control, move up a level of autonomy or take the agent to another team.
Next step
If you want to know which task in your company is the best candidate for a first agent, we will come back with a starting point in under 1 working day.