Project: an AI adoption plan for your role

Choose a task you do at least weekly that involves reading or writing text, avoiding anything where one unseen mistake causes serious harm. The best first tasks are frequent, low-stakes and easy to check. Write the plan, try it on ten real examples and record what happened.

In Daveth Pro (AI Engineering)

What this page answers

Which tasks should I use AI for first?

Frequent, low-stakes tasks that are easy to check, such as summarising meeting notes or drafting replies to a common kind of email. Read it in “The brief”

Which tasks should I not give to AI?

Tasks where one mistake causes serious harm and no human would see it first: sending money, medical or legal decisions, or anything about a person's job. Read it in “The brief”

The brief

Choose a task you do at least weekly that involves reading or writing text: summarising meeting notes, drafting replies to a common kind of email, extracting fields from documents, first drafts of reports, triaging requests. Avoid tasks where one mistake causes serious harm and no human would see it before it does (sending money, medical or legal decisions, anything about a person's job). The best first tasks are frequent, low-stakes and easy to check.

Then write the plan, try it on ten real examples, and record what happened. A plan that looks good on paper but hasn't been tried is a guess; lesson E3 showed how much the details of a prompt matter.

Your plan

The task

What is it, how often do you do it, and how long does it take by hand?

Why this task

Why is it a good first candidate: frequent, low-stakes, easy to check?

What the tool sees

Exactly what you'll paste in or connect, and what you'll remove first (names, personal or confidential data).

Which tool

Which AI tool, and is it approved for this kind of data at your organisation?

The prompt

Your instructions and two or three worked examples, in the format you want back.

What comes back

The exact format of the output and where it goes next.

How you'll check it

What you check, how often, and what counts as wrong.

What could go wrong

Invented details, leaked data, biased or harmful output, over-reliance. For each: how you'd notice and what you'd do.

Who's accountable

Who signs off the result and answers for mistakes.

Trial on ten real examples

How many were right, what failed, and what you changed.

Is it worth it

Time saved per week, minus time spent checking. Would you keep doing it?

Grade yourself

Count the statements that are true of your plan. Twelve or more of the fifteen is a plan ready to show someone; fewer means a section needs more work.

The task is one I really do at least weekly.

A mistake in the output would be caught before it caused serious harm.

I've listed exactly what data the tool sees.

Personal and confidential details are removed or the tool is approved for them.

The prompt includes clear instructions about the output format.

The prompt includes at least two worked examples.

The material the answer depends on is in the prompt, not left to the model's memory.