Your team is already using AI. Now what?

Find out how AI is already being used.

Ask people what they use AI for, what they put into it, what comes back, and how they check the answer. A short anonymous survey often gets more honest answers than a policy meeting.

Keep the questions about work rather than enthusiasm:

• What task are you trying to complete?
• Which tool do you use?
• What information do you give it?
• What do you do with the output?
• Who checks the result before it is used?
• Where does the tool save or retain the information?

Do not frame the inventory as surveillance. People are less likely to disclose risky habits if they think the survey is a trap. Explain that you are trying to make useful work safer and easier.

Choose work worth testing.

A good first pilot is narrow, repeatable, and easy for a knowledgeable person to review. It should solve a real nuisance without putting an important decision in the model's hands.

Good early candidates

• Turning approved notes into a first draft.
• Summarizing a long internal document for a reviewer.
• Classifying routine messages before a person responds.
• Comparing information across a known set of files.
• Preparing a checklist from an established procedure.

Poor early candidates

Do not start with hiring, discipline, legal advice, financial decisions, health or safety decisions, unattended customer commitments, or work where nobody can reliably judge the answer.

Decide what information stays out.

Before the pilot begins, decide which tools are approved and which information can go into them. The answer depends on your account type, vendor terms, settings, contracts, industry, and location. Do not assume every version of the same AI product handles data in the same way.

A sensible default for an early test is to use public, synthetic, or properly anonymized information. If real internal data is necessary, confirm the approved account, controls, retention, and vendor terms first.

Keep a person responsible.

AI can prepare, compare, summarize, classify, or suggest. A named person still owns the work. Human review is not a temporary inconvenience. It is part of the process.

Write down four things before the first live use:

1. Owner: the person accountable for the workflow.
2. Reviewer: the person qualified to check the output.
3. Checks: the facts, sources, tone, calculations, or requirements that must be verified.
4. Escalation: what happens when the answer is uncertain or the work falls outside the approved use.

Run one 30-day pilot.

Keep the pilot small enough to understand. One team, one workflow, one owner, and one way to judge the result is plenty.

1. Week 1
   Record the baseline.

   Collect ten recent examples. Note the time, common errors, rework, and current approval steps.

2. Week 2
   Build the test.

   Write the instructions, prepare safe examples, choose the tool, and agree on the review checklist.

3. Week 3
   Use it on real work.

   Log what the AI produced, what the reviewer changed, what went wrong, and how long the full task took.

4. Week 4
   Compare and decide.

   Review the evidence with the people doing the work. Stop, revise, or expand the test.

A pilot is allowed to fail. Finding out that a workflow is a poor fit is cheaper than forcing adoption because the organization has already announced a program.

Measure the whole task.

Do not count only the seconds it takes the model to produce an answer. Include preparation, review, corrections, handoffs, and cleanup.

Decide what happens next.

The pilot should end with a decision, not a vague plan to use more AI.

Most teams do not need a custom system for their first useful AI workflow. They need a clear problem, an approved tool, good examples, a responsible reviewer, and evidence that the process is better than what it replaced.