Becoming the AI person on your team: a 90-day plan

What the role actually is

Nobody will give you this title. It forms when colleagues start asking you, "Can we use AI for this?" and your answers turn out to be right. A good answer is usually one of three:

• "Yes, and here is the prompt and the check I use."
• "Partly — it drafts the first version, but you still need to verify X."
• "No, because the data can't go there / the error would be invisible."

The plan below builds the evidence and habits behind those answers.

Days 1–30: find the boundaries

The aim of the first month is to know the rules better than anyone else on the team.

If your log shows that verifying took longer than doing the work, record that too. It is a finding, and it will save a colleague from wasting the same week.

Days 31–60: make one workflow better and measure it

Take the chosen task and turn it into a repeatable procedure, not a clever one-off chat.

1. 1Write the procedure down: inputs, the exact prompt or template, what the tool produces, and the check a human performs.
2. 2Measure before and after on the same kind of work. A simple approach: time yourself doing five instances the old way and five the new way, including verification time. Write down any errors caught during checking.
3. 3Keep a failure list. Every time the output was wrong, note how you noticed. This becomes the "watch out for" section when you teach it.
4. 4Show your manager a one-page summary: what changed, the measured time including checks, the error types, and what data the workflow does and doesn't touch.

Resist the urge to automate three workflows. One well-documented, honestly measured example is more persuasive than three demos, and it gives you something to point to at appraisal time.

Days 61–90: teach it and set up the question queue

Now make yourself useful to others without becoming the bottleneck.

• Run a 30-minute walkthrough for the team using your real workflow, including one failure you caught. Showing a failure builds more trust than a flawless demo.
• Publish a short internal page: the approved tools, your summary of the policy, the workflow, and a "don't do this" list (no customer data in unapproved tools, no final figures from chat output).
• Create a simple intake: a shared sheet where colleagues can suggest tasks for AI help. Triage each one with the three-answer framework above.
• Pick the next candidate from the intake, and tell your manager which one and why.

Traps that end the reputation early

Overclaiming. If you say a tool "saves hours" and a colleague finds it saves ten minutes after checking, your credibility drops for every later suggestion. Quote only what you measured.

Tool-hopping. Demonstrating a new app every week signals enthusiasm, not judgement. Depth in the approved tool matters more.

Doing everyone's work. If you personally run every AI task, you have become a service desk. Teach the procedure and hand it over.

Ignoring security. One incident of confidential data in the wrong tool will outweigh every good result. When unsure, ask before trying.

A checklist for day 90

• I can state our company's AI policy in five sentences and link to the source.
• I have one documented workflow with measured time, including verification.
• I have a written list of failures and how I caught them.
• At least two colleagues have used the workflow without my help.
• My manager has seen a one-page summary and knows what I plan next.

Once the checklist is done, the natural next step is proposing something bigger; see how to propose your first GenAI pilot on one page. If you want structured practice building evaluated assistants and agents alongside your job, the Generative & Agentic AI add-on is designed as a three-month skills track taught live online.