Your first AI agent: start with one process

Most AI projects that fail, fail on the choice of starting point. Not on the technology. Start with a platform instead of a process and you build something nobody was waiting for.

What an agent is, and what it is not

An AI agent is software that carries out a task on its own: it receives an instruction, consults systems, makes a decision within predefined rules and performs an action. Not a chat window someone types a question into, but a step in your process that happens by itself.

That difference decides the business case. A chat assistant makes individual employees faster. An agent takes work away.

Which processes are suitable

Four properties make a process suitable. If two are missing, choose a different process.

• Repetition. It happens dozens of times a week, in a similar way.
• Rules. Someone can explain when which choice is made, even if it is written down nowhere.
• Systems with an interface. The data sits in ERP, CRM, HR or Microsoft 365 and can be retrieved.
• An owner. Someone is responsible for the result and decides on exceptions.

Decision rules: the real work

Process analysis almost always shows that the procedure is not what actually happens. There are exceptions nobody wrote down, and those decide whether an agent becomes usable.

So we record three things: what the agent may do on its own, when it escalates to a human, and what happens when it does not know. Escalating in case of doubt is nearly always the right behaviour; an agent that guesses costs more than it returns.

Let it run alongside first

We never put a new agent straight into production. It first runs alongside the existing work: it makes its proposal, an employee does the work as usual, and we compare. After a few weeks you know exactly where the rules need adjusting, and the team has built confidence.

That phase is also the cheapest place to discover that an integration does not give what you hoped.

What we most often start with

At customers these were the first agents to go into production: classifying incoming post and mail with entry into the right system, a draft quote based on a request and the company price list, categorising service desk tickets, and on- and offboarding of staff. The last is popular because it also brings a compliance benefit: every step is logged, so you can demonstrate that permissions were revoked.