AI agents in business: where to start

An AI agent is a program that relies on a language model to understand a request, consult authorised sources and carry out a sequence of actions: looking up information, drafting a reply, filling in a form, preparing a decision. The real question is not “do we need agents?” but “on which task does an agent deliver measurable value at a controlled level of risk?”. This guide sets out a four-step approach to answer it.

Choose a business task

A good first use case is a task that is frequent, well described and verifiable. Projects often fail because they start from an ambition that is too broad (“an assistant for the whole company”) rather than from a precise problem.

For each candidate task, ask three questions:

• Who does it today, and how much time does it take? A task without a clear owner has no one to judge the quality of the result.
• How do we know it has been done well? If there is an expected answer, a procedure or a document template, evaluation is possible.
• What happens if it goes wrong? An error in a draft reviewed by a person does not carry the same weight as an error in a payment.

Document search, triage of incoming requests, preparing replies and summarising case files are often good starting points: they are repetitive, their output is easy to review and they free up time for teams.

Review the data

An agent is only as good as the information it can reach. Before choosing a solution, list the sources the task depends on:

Question	Why it matters
Where are the useful documents?	Scattered or outdated sources lead to incomplete answers.
Who is allowed to read them?	The agent must respect the same access rights as people do.
Do they contain personal data?	The applicable data protection rules depend on it.
Are they up to date?	An agent quoting an obsolete procedure misleads your teams.

This step often reveals some groundwork on sorting or updating sources. That work pays off on its own: teams benefit from it even before the agent goes live.

Set boundaries

A well-designed agent knows what it can do on its own, what it should propose and what it must decline. These boundaries are decided with business owners, not by the technical team alone.

• Permitted actions: for example searching, summarising, drafting.
• Actions requiring approval: sending a message to a customer, changing a record, committing spending.
• Escalation situations: an ambiguous request, missing information, a sensitive topic. The agent then hands over to a person, with the context.

Plan for traceability too: which sources the agent consulted, what it produced and who approved it. These records help correct errors and account for the decisions made with the agent.

Evaluate the pilot

A pilot turns an intuition into a decision. It is prepared before development starts:

1. Build a set of representative examples: real cases, including difficult ones, with the expected answer validated by business experts.
2. Set success criteria: accuracy of answers, share of cases handled without intervention, time saved, user satisfaction.
3. Test with a small group for a few weeks, collecting feedback and errors.
4. Decide: extend, adjust or stop. A pilot that concludes “stop” has also done its job: it has prevented a misdirected investment.

Only after a conclusive pilot do you industrialise: continuous monitoring, access management, team training and source maintenance.

Getting started in practice

Bring together a business owner, someone who performs the task day to day and a technical counterpart. In a single session you can list candidate tasks, select one or two and identify the data they need. That is the starting point for an assessment: it weighs expected value, risks and effort, and leads to a pilot plan measured against your own criteria.