Let AI Inventory Your Own Expertise── Hand over past work and let AI find the thread you cannot see from the inside

01Three bodies of work went in; one thread came out

The operator selected three sections of a site, each covering a different subject, and asked an AI to verbalise the expertise they contained and organise it into twenty slides. The three sections had been written at different times for different purposes.

The AI returned a twenty-slide outline. At the top it stated that the three sections were not separate fields but a single line: read norms through history, structure the risks that regulation leaves uncovered, and translate that structure into a form machines can enforce. The axiom running through all three was information asymmetry.

The operator had not seen this line. Each section had been written on its own, with its own goal. The AI read all three from start to finish, pulled out the concepts that recurred, and compressed them into one proposition.

02The longer you practise, the harder it becomes to name your own knowledge

There is a well-documented explanation for this. As experience accumulates, knowledge shifts from conscious reasoning to automatic judgement. Ask an experienced reviewer why a certain claim is problematic and the answer is often "it just is". Management scholar Ikujiro Nonaka called this kind of hard-to-articulate knowledge tacit knowledge, and named the process of converting it into explicit words externalisation.

Externalisation matters to organisations. As long as tacit knowledge stays inside one person's head, nobody else can use it. When that person leaves, the knowledge leaves too. Once it is in words, it can be shared, taught and built into systems.

The difficulty is that the person who holds tacit knowledge rarely has a reason to put it into words. It feels obvious. It takes an outside questioner, someone who does not share the same assumptions, to draw it out.

03Invisible expertise leaves you unable to sort tasks between yourself and AI

When working with AI, invisible expertise has a practical cost. At the idea stage, if you do not know what you know and what you lack, you cannot decide what to delegate and what to judge yourself.

A 2023 field experiment by Dell'Acqua and colleagues at Harvard Business School tested management consultants working with AI. On tasks within the AI's strength, performance rose by roughly 40 percent. On tasks outside it, performance dropped by about 19 percentage points. The researchers called this uneven boundary the jagged technological frontier. The boundary is not obvious from the surface of a task.

If your expertise is in words, the boundary becomes easier to estimate. Knowledge you hold deeply is the domain where you judge, not the AI. Knowledge you lack is the domain where you let the AI investigate. The inventory gives you the material for that sorting.

04An inventory means handing past writing to AI and getting structure back

The inventory described here is a specific action. You collect documents you wrote in the past, hand them to an AI, and ask it to return the structure of the expertise they contain.

Documents that work well for this are ones you spent time writing yourself. Meeting minutes or transcriptions of someone else's words carry little of your own expertise and yield thin results. In the operator's case, the material was three sections the operator had researched, organised and written.

05In material review, past comments reveal the reviewer's unstated criteria

Pharmaceutical material review is a setting where this method applies directly. A reviewer with years of experience holds a large body of past review comments. Hand those comments to an AI and ask it to extract the recurring criteria, and the reviewer's unstated standards become visible.

Patterns might emerge: expressions that extend beyond the approved indication, data citations that cannot be traced to a source, graphs whose comparison baseline is not stated. These can be shared as explicit rules. When a new reviewer joins, the criteria can be handed over as a document rather than absorbed over months of observation.

Research by Brynjolfsson and colleagues found that AI assistance raised the productivity of less experienced workers by 34 percent. One likely reason was that the knowledge of experienced workers reached newer colleagues through the AI. An inventory that turns a reviewer's tacit criteria into explicit rules serves the same function: it turns one person's experience into something the team can use.

06AI finds the thread because it reads everything at once, without context

When you re-read your own documents, you read them in the order and context in which you wrote them. Even if a common thread runs through documents written years apart, the context of each one makes it hard to see.

An AI has no context. It reads the material from beginning to end without knowing the order it was written in or the purpose behind each piece. That is why it can notice that the same concept appears in several places. In the operator's case, the concept of information asymmetry recurred across all three sections, but the operator had not been aware of it as a unifying idea.

This does not mean the thread AI finds is always right. Language models tend to return answers that match what the user appears to believe. Anthropic researchers have reported that this tendency toward agreement is common across current AI assistants. The structure AI returns needs to be checked against your own judgement. If you can say, "yes, this criterion does guide my decisions", the words are usable. If not, tell the AI where the description is wrong and have it revise.

07From tomorrow: pick three documents, ask for the thread, and verify it yourself

There are three steps.

1. Choose three or more documents you wrote. Different periods and different subjects work best. More material reveals more, but three are enough for a thread to appear.
2. Ask AI to extract the structure. A prompt like "from these documents, extract recurring concepts, shared assumptions, and the thread that runs through all of them" works. Ask for a structured format such as slides, chapter headings or a diagram, rather than a flat list, so the relationships between concepts become visible.
3. Check the thread against your own judgement. Ask whether the axis the AI found actually influences your decisions. If it does, use it as material for the next idea. If it does not, say where it is wrong and ask the AI to revise.

Anthropic's prompting guidance notes that giving AI background and context improves the quality of its output. The words you get from an inventory are that background and context. Telling an AI "my expertise is in structuring information asymmetry" before the next request helps it focus its response on the area where your knowledge is deepest.