The task audit

Start with a list, not a licence

Most workplace AI failures start the same way. Somebody buys a subscription, sends a launch email, and then everyone goes looking for something to do with it. That is backwards. The unit of adoption is not a tool. It is a task.

So the first real exercise in this course is dull, and it is the one that decides whether any of the rest helps you. Take last week. Write down every task you did that took more than twenty minutes. Not projects — tasks. "Wrote the parents' newsletter." "Chased six unpaid invoices." "Read the new procurement circular and worked out what changed." "Wrote up eleven patient handover notes." Most people land somewhere between twelve and twenty-five items and are mildly surprised by the list, because a week does not feel like that from inside it.

Now three questions per task. They take about ten seconds each.

The three questions

1. Where does the truth live? This is the sorting rule from the previous lesson, applied item by item. If everything needed to do the task is in a document you can hand over, the truth is in the text. If the task depends on your state's current fee schedule, what your supplier actually delivered, or what your manager said in the corridor, the truth is out in the world and the model is guessing.

2. What does a wrong answer cost? Be specific and be honest. A clumsy sentence in an internal email costs nothing. A wrong appeal deadline in a refusal letter costs somebody their home and you your job. Most tasks are far cheaper than people assume, and two or three are far more expensive.

3. What does checking cost? This is the question nobody asks and it is the one that decides. Not "can I check it" — how long does checking take, in minutes, compared with doing the work yourself?

The trap in the third question

Here is the pattern that makes people quietly abandon AI after a month, and it has nothing to do with quality.

A solicitor asks for a summary of a forty-page expert report. It arrives in ninety seconds and it reads well. But she is going to rely on it in a hearing, so she reads the forty pages anyway to confirm nothing was dropped. Total time: the original reading, plus ninety seconds, plus the summary. She has spent more time than before and produced the same document.

That is not a bad summary. That is a task where verification cost exceeds production cost, and no improvement in the model fixes it. If you must check every line against the source to be safe, you have to read the source, and reading the source was the work.

The fix is almost never "use it anyway". It is to split the task. In that example: let the model produce a navigation aid rather than a summary — a list of which paragraphs discuss causation, which discuss quantum, which discuss the claimant's history. Wrong entries are visible in seconds because you open the paragraph. The reading stays hers. The value is real and the checking is cheap.

Ask this of every task you are tempted by: what is the cheapest thing to check here? Usually it is structure, location and format. Usually it is not facts.

The sort

With three answers per task, sort into three piles.

• Hand over. Truth is in the text, a wrong answer is cheap, checking is a skim. First drafts of routine correspondence. Turning notes into prose. Reformatting. Sorting eighty survey comments into themes. Start here, today.
• Assist. You do the task; the model does a defined part of it. The formula but not the answer. The counter-argument but not the position. The structure but not the citations. Most professional work lives here and this is where the course spends most of its time.
• Never. A wrong answer is unrecoverable, or checking costs more than doing, or the judgment is the job. Dosages. Filed accounts. Whether to dismiss someone. A named patient's history going into an unapproved account.

Write the pile next to each item. That page is your curriculum, and it is more useful than any list of prompts on the internet, because the tasks are yours.

What people find

Two things happen almost every time, so you may as well expect them.

The first: the tasks that move are not the technical ones. People assume AI will help with the hardest thing they do. It usually helps most with the most repetitive writing they do — the third-reminder email, the standard risk paragraph, the monthly report nobody reads but somebody must write. Unglamorous, and it is where the hours are.

The second: two or three items on the list turn out not to need AI at all. They need a template, a saved reply, or a decision that the task should stop. An audit finds those too, and they save more time than any model will.

Do this today

Do the list. Twelve to twenty items, three answers each, three piles. Twenty minutes.

Then pick one item from the hand-over pile — the most boring one — and only that one, for a week. Note how long it used to take and how long it takes now, including checking. You now have a number, which is more than most organisations have after a year of enthusiasm.

Set a reminder to redo the audit in a month. Tasks move between piles as the tools change and as your checking gets faster, and a classification nobody revisits stops being true.