The Careful Person's First Month With AI Agents
A week-by-week plan for careful experts adopting AI agents: which tools to try first, what to measure, and how fast adopters can build the judgment AI lacks.
Every team has two people who worry about AI for opposite reasons. One has used it every day since 2023 and ships a lot. The other hasn’t touched it, and is the person everyone asks when the numbers look wrong.
Which of them is better placed for the next five years? We don’t know yet, and the research is thinner than the confidence on either side. What we can say is that the two of them have different next moves, and that one of those moves is much easier than the other.
This post argues that the careful expert who avoids AI has the cheapest, highest-return move available to anyone right now: adopt, and bring the checks along. The fast adopter’s move is harder, because no model hands out judgment. Below is a four-week plan for the first person, and a practice routine for the second.
One controller, before her first month
Ruth is a financial controller at a mid-sized distributor. She is a composite, built to show the shape of the situation, not a real person. She has run month-end close for years and has never used an AI tool for work. Her team trusts her because she catches things.
Look at what she already has. Every task ends in a check, and every check is something a person does to output they didn’t fully trust. That is exactly the work AI output demands. Ruth has spent her career building the half of the system that most fast adopters are missing.
What she lacks is the cheap first draft. Her variance commentary takes most of a day because she writes all forty lines herself.
The model: judgment and AI use are separate axes
Here is the model in one sentence: whether you use AI and whether you have good judgment are independent, and AI multiplies whatever judgment is already there rather than supplying any.
The careful person’s suspicion is the common position, not an eccentric one. The 2025 Stack Overflow Developer Survey found 84 percent of developers using or planning to use AI tools, while 46 percent said they distrust the accuracy of the output and 3 percent said they highly trust it. Almost everyone is using tools almost nobody trusts. The person who arrives with working checks is the one equipped for that situation.
Move 1 is a tooling problem. Move 2 is a skill problem. Tooling problems are cheaper.
Week 1: a chat assistant on drafts nobody audits
Pick one chat assistant and stay with it all week. Switching between three is a way to feel busy without learning any of them. The tasks should be ones where a mistake costs you a minute of embarrassment at most, and where you would reread the whole thing regardless.
The log matters more than the drafts. Write down how long the task usually takes, how long it took with the assistant including your edit, and every error you caught: a wrong date, a softened commitment, a summary that dropped the one point that mattered. Ruth’s first recap invented an action owner who wasn’t in the meeting. She would have caught that anyway. Now she has it on paper, and that is how you learn what this particular tool gets wrong.
Week 2: a task you could do yourself, done both ways
In 2023, Fabrizio Dell’Acqua and colleagues ran a randomized experiment with 758 consultants at Boston Consulting Group. On creative product tasks that GPT-4 handled well, consultants with AI finished 12.2 percent more tasks, faster and at higher rated quality. On a business problem built to sit just outside what the model could do, the AI group was 19 percentage points less likely to get the right answer. The authors called the boundary a jagged frontier, because tasks of similar difficulty landed on different sides of it.
You can’t see the frontier from the outside. You find it by doing a task you know cold, once by hand and once with the tool, and diffing the results. A developer might hand Claude Code or Cursor a small refactor they would normally do in an afternoon. Ruth asks the spreadsheet assistant to build the accrual roll-forward she has built for years, then ties it to her own version.
The week’s measure is a second column of errors: the ones in the AI version that your own version didn’t have. That column is your personal map of the frontier, for your work, with this tool.
Week 3: deep research, with every citation opened
Deep research tools, such as ChatGPT’s deep research mode and Gemini Deep Research, search the web for a while and return a long report with citations. The citations are what make them useful to a careful person, because they turn the report into a checklist.
Pick a question you already know something about. Ruth asked for the current treatment of a lease accounting edge case she had dealt with the year before. Then open every citation and mark it. The middle outcome is the one to watch for: the source exists and is relevant, but it says something weaker or narrower than the report claims. A fast reader skims past those. A careful one has been catching that exact move in other people’s memos for years.
Week 4: delegate a multi-step task, with gates
By week four you have a log that says where your tool is reliable and where it isn’t. Now hand it a task with several steps and put yourself between them. Agentic tools already support this shape. Claude Code, by default, asks for confirmation before it edits a file or runs a shell command. Coding agents in Cursor and GitHub Copilot propose changes you review before accepting. Outside code, the same pattern works with a document or spreadsheet agent if you split the job and refuse to let it run end to end.
The measure this week is where you stopped it. Count the gates where you rejected or rewrote the agent’s work, and note how much rework that saved. Ruth stopped hers at the variance gate: it had flagged a timing difference as a real variance. The two she redid by hand are what told her to look.
The harder move: building judgment after adopting fast
The fast adopter’s problem shows up in the best-known field study of AI at work. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied a chat assistant rolled out to more than five thousand customer support agents (NBER working paper in 2023, published in the Quarterly Journal of Economics in 2025). The newest and least skilled agents gained the most, while the most experienced saw small speed gains and slight quality declines. The authors’ reading is that the tool spread the practices of the best agents to everyone else. On scripted work, a novice can borrow judgment from the tool. That is real, and it is also borrowed. It covers the cases the top agents already solved and nothing past them.
So the fix can’t be more tool. In October, Berkeley News reported on randomized trials by Grace Liu, Brian Christian and colleagues with 1,222 participants solving math and reading problems. When the AI was removed, people who had mostly asked it for direct answers did worse and gave up sooner than people who never had it. People who asked for hints stayed close to the control group.
The practical version is the review rep. Before you prompt, write your own answer, even a rough one. Generate. Diff the two. Log who was right and why. Ask for hints rather than answers on anything you want to keep being able to do. It is slower, deliberately, for a few months, and it gives a fast adopter what a careful expert built over years: a record of where they are wrong. For the engineering version of this, where review becomes most of the job, see the engineer after the editor.
What this gets wrong
Judgment isn’t one number. Ruth’s checks are excellent inside accounting and worth little if she asks an agent to write a script against the general ledger’s database. A careful expert outside their domain is a fast adopter with good manners. The four weeks should stay inside the work you already know how to check, and the plan says nothing about the rest.
The evidence is narrower than the model. The frontier study used one firm and a 2023 model; the frontier has moved since, and probably not uniformly. The persistence trials measured people right after a short session on math and reading, which says little about a year of daily use. The support study covered scripted work, where the best answers already existed to be copied.
And the plan relies on a log you keep yourself. A careful person can still go easy on a tool they have started to like. If you can, have a colleague spot-check two entries a week. Then there is the chance that the axes stop being independent: tools may get good enough at checking their own work that the careful person’s edge shrinks. If that happens, the annual review in the annual review for an AI year is where you would see it first.
Your next move
If you are the careful one, block thirty minutes on Monday, open one chat assistant, and hand it the most boring draft on your list. Start the log before you start the draft. Four weeks from now you will have something the fast adopters around you mostly don’t: a written record of where the tool is right, where it is wrong, and how you know.
If you are the fast one, write your own answer before your next prompt. Do it once a day for a month, and keep the log. For why the shape of the help matters, ten minutes of answers goes deeper on the persistence trials.
