Monday morning. Your inbox is already ahead of you. There is a request to sort, a document to find and the same information to enter into three different tools. That is often a sensible place to start with AI: a recurring task your team can describe without a slide deck.

Find the repetition that actually gets in the way

For a few days, record the steps people repeat: reading an enquiry, identifying its purpose, finding a procedure, drafting a reply and handing over a file. Identify the input, the expected output and the person who checks it. “Manage customers” is too broad. “Sort website enquiries and prepare a contact record” is a workable starting point.

Choose something frequent with an outcome people can inspect. Your team should be able to explain why a proposed result is right, incomplete or irrelevant. If the expected result is unclear, automation will carry that ambiguity forward.

Three practical starting points

Request triage can identify the service, language and missing information before suggesting who should handle it. Document assistance can find a current procedure and provide its source link. Follow-up preparation can summarize an exchange and suggest a next action.

These are illustrative scenarios, not reported client results. A business can begin with one of them while keeping message sending and record changes under human control.

Not every step needs AI

A fixed rule such as “create a record when this form arrives” may be handled by ordinary automation. AI becomes useful when it must interpret free-form messages, ambiguity or different wording. The two approaches can work together: interpret the request, then use defined steps to record and route the information.

Anthropic distinguishes predefined workflows from agents that choose their next actions. That distinction is useful when deciding what responsibility a system should have.

Give it a clear way to hand over

An unusual request, conflicting documents or an unreadable attachment needs an explicit path. The system can flag the problem and pass the gathered context to a colleague. It should not fill gaps with a confident guess.

For an initial trial, we recommend letting the tool prepare work while a person approves consequential actions. Its role can grow once the results justify it.

Look beyond the impressive demo

Test representative enquiries, including awkward cases. Check incorrect classifications, missing details, necessary edits and how easily colleagues can take over. An attractive interface is not evidence of reliable follow-up.

A worthwhile first project remains useful after the demonstration. At ARKHAME, defining an AI agent starts with the work people need to complete and the systems that must connect.

Further reading

Would you like to discuss a project for your business? Tell us what you need.