Expertise gains value when you write down how you decide, let AI use it and keep the ability to check the result. Here is how, with an audit, a prompt and a 30-day plan.
Audit which parts of your work AI has made cheap, then encode the scarce part as a checklist or brief using the interview prompt. Agree what AI does, what you own and how it’s checked, share the standard with one colleague, and score it monthly for use, persistence and impact. No salary or job-security promise is made.
Recall, first drafts and routine analysis got cheap quickly, and many experienced professionals responded by defending the wrong thing. They guard the facts they know, when facts are the part a tool now produces in seconds. What stays scarce is knowing which facts matter in this case, which answer to distrust and what the client or the CFO will do with it. That is expertise, and it gains value only when other people, and the tools you work with, can get at it.
A finance analyst at month end asks an assistant to explain a swing in a cost line and gets a fluent, plausible paragraph back. Her seniority shows in the two sentences she doesn’t believe, because she knows which accrual is usually late and which vendor splits its invoices. She fixes the paragraph in a few minutes. She has never written down why she doubted those two sentences, so the next person to use the same prompt takes the paragraph at face value.
My view is that expertise gains value when you encode how you decide into something AI can use, and keep the ability to check what comes back. Encoding means writing down what you look at first, what makes you distrust an answer and what you change in nearly every draft. Checking means staying able to tell good output from fluent output. Skip the first and your judgment stays in your head while the tool works from generic habits. Skip the second and you’ve handed the judgment over without noticing. For the wider question of which human skills gain value, see our piece on human skills in the AI era. This guide is about the packaging.
Start with an audit of your own job, because the advice differs by role. A consultant’s scarce part is framing the client’s real question. A marketer’s is knowing which claim the brand can’t make. An HR partner’s is reading what a manager isn’t saying, and a senior individual contributor’s is knowing which shortcut will break in production. Here is a worksheet filled in for a finance analyst. Do yours the same way, with your own last month of work in front of you.
| Part of the work | Cheaper now? | Still scarce | What I’d do about it |
|---|---|---|---|
| Recalling policy and accounting treatment | Yes. Quick to look up and explain | Knowing which rule applies to this contract | Let AI draft the lookup, then check the clause myself |
| First-draft variance commentary | Yes | Knowing which variance is real and which is timing | Encode my doubts as a checklist the draft must pass |
| Routine reconciliations and formatting | Yes | Spotting the one line that shouldn’t reconcile | Automate the routine, keep the exception review |
| Deciding which number the CFO will question | No | Knowing the audience and their history with this line | Write the questions down so a junior can prepare them |
| Telling a business partner not to cut a budget | No | Trust, and accountability for being right | Prepare with AI, deliver myself |
A completed example. Replace each row with a task from your own last month. The cheaper-now column is a judgment about today’s tools, so revisit it each quarter.
Run it in four steps:
The cell people tend to leave blank is the scarce part. Without it the audit is only a list of tasks, and the whole point is to find the sentence you could hand to a junior colleague.
There is a reason to write your expertise down yourself instead of waiting for someone else to. The clearest evidence comes from a study of 5,179 customer support agents who were given a generative AI assistant. Productivity rose 14% on average and 34% for novice and low-skilled agents, with minimal impact for the most experienced. The authors also report suggestive evidence that the model spreads the working practices of the more able agents.[1]
That is one company and one kind of work, so read it as research on a particular setting, not as a law. What it suggests is useful, though. If your edge is a way of working that can be described, it will be copied, by a tool or a colleague, whether or not you ever packaged it. The part that stays yours is applying it to the case in front of you and answering for the result.
So the choice is not whether your know-how circulates. It’s whether it circulates in a form you wrote and can correct, or one that gets inferred. For what happens on a team when strong people pull ahead of the rest, see AI as a system that multiplies expertise.
Most experts can’t answer “how do you do that?” on the spot, which is why the useful trick is to let the tool ask the questions. Give it an example of your recent work and ask it to interview you about the decisions inside it. The answers become a checklist, a brief or a set of standing instructions you can reuse.
Write down how you decide before you ask AI to decide anything for you.
This is the prompt. It works as written in Claude, ChatGPT or Copilot, with the brackets filled in.
| I’m a [role, for example senior marketing manager] and I review [type of work, for example campaign briefs] for [audience or standard]. I’m going to paste one example of my work with all names and figures replaced by placeholders. Interview me, one question at a time, about the decisions inside it: what I look at first, what makes me distrust a section, what I change in nearly every draft, and what I would refuse to sign off. After ten questions, write a one-page checklist in my words with three parts: what good looks like, the five things I check, and the mistakes I see most often. Mark anything you inferred rather than heard from me with [INFERRED] so I can correct it. |
Copy, fill in the brackets and paste your anonymised example after it. The [INFERRED] marker is what keeps the checklist yours.
A consultant who runs this after a proposal comes out with a checklist like this one:
The first run usually produces something partly right and generic in places. Correct it, run it on the next piece of work and correct it again. After a few rounds it reads like you, which is the point.
One warning before you paste anything. Client and employer material doesn’t belong in a public AI tool, so replace names, figures and anything identifying with placeholders, and check your organisation’s policy on which tools are approved. Our guide to using AI without leaking company data covers the practical rules.
A good checklist creates a new risk, which is trusting the output because it was produced by your own rules. Fluent answers look the same whether they’re right or wrong. In a preregistered experiment with 758 consultants at Boston Consulting Group, people using AI on a task outside the tool’s capability were 19% less likely to produce correct solutions than people working without it.[2] Past success with the tool told them nothing about where it would fail.
The takeaway for you is practical. Your checklist should say where the tool is weakest on your work, and a person has to own each judgment explicitly. This is the split to agree before you delegate anything, filled in for six roles.
| Role | What AI does | What the human still owns | How it gets checked |
|---|---|---|---|
| Consultant | Drafts the first structure of a client memo from my notes | The framing of the client’s real question and the recommendation | I trace each claim to a source I’ve opened; a partner reviews the logic |
| Marketer | Produces ten headline variants against the brief | Which claims the brand can make, and the final pick | Claims checked against the approved-claims list; one review before anything goes live |
| Finance analyst | Drafts variance commentary from the ledger extract | Which variances are real, and what goes to the CFO | Headline figures tied back to the source system each month |
| HR partner | Groups exit-interview comments into themes | What the themes mean for this team, and what is said to the manager | Sample of original comments read by me; sensitive cases never go through the tool |
| Senior engineer or analyst | Writes a first pass of the routine code or query | The design decision and what ships | Tests and a peer review before merge |
| Manager | Turns my own notes into talking points | The assessment of the person and the conversation | Notes come from my own record; nothing AI-generated is treated as a judgment about a person |
A Future Factors working split. Adjust it to your organisation’s policy and to what the tool can reach.
Put a name next to every judgment the tool touches, and make it yours if you can’t name anyone else.
The rows that matter most are the ones where a person is affected by the result. Those are slow on purpose. For more on building the checking habit, see our guides on critical thinking for AI and why human in the loop matters.
A checklist in your own folder helps you. One that other people use makes you the person who set the standard, and that is where expertise starts to compound. How you package it depends on the role:
Here is one asset filled in, in the format worth reusing for any of them.
| Field | Entry |
|---|---|
| What it covers | Turning a quarter’s exit-interview comments into a themes summary for a department head |
| What good looks like | Themes named in the employees’ words, each with the number of comments behind it and one example |
| What I check before it goes out | That no comment can be traced to one person, and that a theme isn’t one loud voice repeated |
| Where the tool is weak | Tone, and small teams where a theme is really one person |
| What it must not do | Name individuals, rank managers or suggest who should be dismissed |
| Owner and review | Me, reviewed with the head of HR every quarter |
A completed example. The row people leave blank is what it must not do. A Future Factors working template, not a research-validated model.
You can build all of this on your own, and many people should start that way. When a whole team needs the same standard, working with someone who has seen where these programmes stall shortens the learning curve and builds shared capability across the group rather than in one enthusiastic person. It also helps people apply AI to their real work instead of generic examples, which is what our corporate workshops are built around.
This approach has limits. In fast-moving fields a checklist ages quickly, so date it and review it. Some expertise is relational or physical and doesn’t reduce to a page. And nothing here promises a raise, a promotion or job security, because those depend on your employer, your market and decisions outside your control. What it does is make your judgment more visible and more useful, which is the part you can influence.
Judge whether your expertise is compounding by what changed in the work, not by how much you wrote down.
Use, persistence and impact give you a quick test. Run it once a month on each asset.
| Question | Where to look | Yes means |
|---|---|---|
| Use. Did you or anyone else run the checklist or brief this month? | Your own work log, or a note from a colleague | It’s in the workflow |
| Persistence. Did you use it without being reminded, and did you update it? | Edit date and your habits | It’s becoming a habit, not a project |
| Impact. Did it catch an error, save a rewrite or change a decision? | One specific instance you can name | It is worth the maintenance |
Score one point for each yes. Three: share it more widely. Two: fix the weakest part. One or none: pick a different piece of expertise before you decide the approach doesn’t work.
A month is enough to build and test your first two assets, and the plan assumes about two hours a week.
| Week | Do this | Done when |
|---|---|---|
| 1 | Run the audit on your last month. Pick two rows to encode. Anonymise one example of each | Two scarce-part sentences and two clean examples |
| 2 | Run the interview prompt on both. Correct every [INFERRED] line | Two checklists you’d be willing to put your name to |
| 3 | Use one on real work. Fill in the AI does, human owns, how checked row for it | One piece of work reviewed against the checklist, with a named checker |
| 4 | Share it with one colleague. Run the monthly check. Decide what to keep, fix or drop | A scored check and a date to review again |
A Future Factors working plan. Stretch the weeks if your work moves slowly.
This week, pick the one task where you most often fix the draft, and write down the first thing you check. That sentence is the start of your first checklist.
Date every checklist and review it on a schedule, monthly in fast-moving areas. Write down the reasons behind each check as well as the check itself, because the reasons outlast the tools. If a reason stops being true, you’ll see it quickly.
Often your employer, depending on your contract and policy, so ask before you share them outside the company. Inside it, tell your manager what you’re building. Material you write about your own craft, without client or company detail, is easier to keep portable.
No. A prompt tells a tool what to do once. A checklist records how you decide, and it works with or without any tool, because it also trains the colleague who reads it. The prompt in this guide is only the way of extracting it.
Nobody can promise otherwise, and the evidence on this is limited. In the one study here, the assistant spread the working practices of stronger workers whether or not they asked for it. Writing it yourself keeps the standard, the corrections and the credit with you.
Anything involving confidential material you’re not cleared to share, and judgments about individual people. Keep decisions you’d struggle to explain in your own words. If you couldn’t check the answer yourself, don’t hand over the question.
I read two primary sources for this guide on October 26, 2026: the NBER working paper on generative AI in customer support (Brynjolfsson, Li and Raymond) and the Harvard Business School working paper on the jagged technological frontier (Dell’Acqua and colleagues). Both are single studies in particular settings, one in customer support and one in consulting tasks, and I report their findings as research, not as laws. I left out salary and job-market claims, because I couldn’t confirm any to a primary source for this piece. I did not test any tool hands-on for this article. The expertise audit, the extraction prompt, the ownership split, the asset template and the 30-day plan are Future Factors working methods, not research-validated models, and the scenes are illustrations, not client stories.