CoreD Journal

What should your first AI agent do?

Pick one repeatable task, give the agent an approved source, and decide what a person must still approve.

AI agents4 min read

Amber routing card and a document tray on a charcoal desk

Short answer

Start with one reversible task, such as classifying enquiries or drafting answers from approved material. Keep staff in charge of sending, define when the agent should stop, and compare the pilot with the way the team works today.

Pick a task you can check

A useful first agent does one job with a clear finish. Take a support inbox: you might ask the agent to classify incoming questions and draft answers from your approved help pages. A staff member reviews the draft and sends it. This is easier to test than asking an agent to ‘handle support’.

Before you build, collect a small set of routine examples and a few awkward ones. Include a question with no answer in your documents, an old policy, a duplicate request and a customer asking to speak to a person. Write down the response you would expect in each case.

Give the agent a source and a limit

Choose one owner for the knowledge source. Agree which documents the agent may read, what it should do when two documents disagree and where it should send a question it cannot answer. Give it read access before giving it permission to make changes.

For an agent connected to a CRM, write a short list of allowed actions: read a lead record, prepare a note, or suggest an owner. Changing a deal stage, sending a message or deleting a record should require separate approval until your team agrees otherwise.

Measure the review, not the demo

A demo shows that one example worked. In a pilot, track how often staff accept a draft, how often they have to correct it and why they send it back. Compare the review time with how long the same task takes today.

Start with a queue someone owns. Review the first batch together, add the missed cases to your test set and decide whether the next step should be better coverage, a single integration or no expansion at all.

Where an agent fits better than a rule

Use a rule when the condition and the action are known. A form tagged “billing” can go directly to a billing queue. An AI agent can help when the input varies, such as a customer describing an invoice question without using the word “billing”. Both can run in the same workflow. Let the rule enforce permissions and let the agent interpret the wording.

A first pilot can work in draft mode. Classify a support email, show which policy the draft relied on and leave the Send button with staff. Give the agent an “unclear” outcome so an unusual message still reaches the right person. The reviewer needs the original question and the reason the draft was routed there.

Make a small acceptance set

Collect routine cases from the team, remove unnecessary personal information and write the expected outcome for each. Include two lookalike requests that should take different paths. A billing question and a request to change banking details may sound similar, but only the first belongs in an FAQ workflow.

Keep a separate set for checking changes to source material, instructions or the model. Look for serious errors staff might miss. A draft that confidently cites an outdated policy may cost more than one that simply asks for human help.

Decide what happens when the pilot ends

Review time spent per case, how many drafts staff actually use and why they reject the rest. If people spend more time checking evidence than they save writing, improve the sources and format before adding more actions. Name a person who can pause the workflow and agree how the team will handle incoming requests during a pause.

Does a first agent need access to the whole CRM?

Usually one read-only view or a small set of approved fields is enough for a pilot. Add permission to write or send after the team has agreed on an approval step and knows how to revoke access.

References

Explore AI agent plans

Keep reading