Why It Feels Like You Know: "Surface Expertise" and the Siren Call of AI for Strategy

Most owners I work with are now using AI to develop thinking they would have asked me to develop in the past. I've contributed to AI-generated plans and watched them get discussed in leadership meetings, and it's gotten me wondering when AI-instead-of-expert works, and when it doesn't.

I've already written about Cognitive Runaway and Logic Engineering. The new term here is Surface Expertise: thinking that looks like expertise on the surface, but is missing critical substance that expertise itself holds.

AI is enabling confidence to precede competence

My clients feel like they have a whole range of newfound competence. They'll tell me about AI's limitations in their own field of expertise, then rely on it for a plan in an area where they have little. The feeling makes sense. In our own domains, work that finished like this meant we knew the field. AI now delivers that finish in fields where we don’t know what good looks like…and the “this is expert” signal fires anyway.

Most of what an owner does with AI is fine. The shift happens on a narrow band of decisions: the ones where the answer depends on context that requires expertise to put into the prompt. Strategy decisions are often complex, important, expensive, uncertain, and ambiguous - and those are conditions that are not well-suited to AI’s expertise. That “narrow band” is a lot wider when you’re working on strategic questions.

Sometimes good enough is good enough. Sometimes it isn't.

Expertise is eight jobs, and AI does a partial version of most of them

When you are the expert, eight things are happening at once.

1. Making the work.
2. Judging the work against a standard (what good looks like).
3. Building the frame the situation needs.
4. Judging the question - whether it is well formed and whether it is the right question at this point.
5. Understanding the standard itself, including the parts nobody in the field has written down.
6. Knowing the edge of your own competence and feeling when a problem sits outside it.
7. Carrying authority, so that people reason inside the frame you endorse.
8. Paying the cost (being accountable) if the expertise is wrong.

AI does a version of most of these. Ask it to judge the work and it will. Ask it for a frame and it will help you create one. But AI’s insight is only partial, and the partial part is always the same part: what depends on knowing my particular situation rather than situations a lot like mine. And understanding what “my particular situation” means in all its fullness, complexity, ambiguity, and uncertainty. The difficulty with AI is that partial expertise arrives with the same finish as the complete version.

Only the last job - being accountable - is out of reach entirely. AI cannot be wrong in a way that costs it anything. An expert can.

The risk isn't AI or the user - the risk is the pairing

AI does what it does everywhere: a partial version of a job, at full finish. We do what we do everywhere: read the finish as a signal about the work. Put them together and the tool's thin spot lands where our evaluation is weakest. The two gaps line up.

That is why the same document is safe in an expert's hands and quietly unreliable in a non-expert's. It also explains why improvements aimed at one side miss: a better model produces a better partial version at the same finish, and a more careful user runs the checks they know inside the frame they brought. Neither touches the pairing, and the pairing is where the risk lives.

Experts get more out of tools than novices. When is the novice’s work-with-AI good enough?

Surface Expertise doesn’t look off to the non-expert

Four of the eight jobs are where the thinness of AI’s expertise concentrates.

The standard. Neither party holds the complete standard. AI holds the generic version; the non-expert takes the output as the standard. So AI is checking itself.

The frame. AI builds inside the frame it's given. Ask for options and you get options; ask for a different angle and you get a different angle. But that is likely not the complete set of options or the right angle. (If you haven't discovered this in working with AI, you will.) Worse, if the frame isn't right, nothing in the output says so.

The question. Ask whether it was the right question and you get a different question and a different answer. Ask again and again and you can have five sets of questions and answers to sort through. At that point, how efficient is the tool compared to asking an expert?

The edge. An expert locates a new problem against a stock of cases and can feel when it sits outside their range. A non-expert has no cases. And I have never had AI tell me a question was outside its range. Ever.

The team accepts Surface Expertise using a rule that used to work

People have long used a reasonable rule: whoever produced work that looks like this knows what they’re talking about. It held because making expert-looking work required being an expert. Now the finish no longer ties to the expertise.

There are two problems with this. First, unless a group consciously fights it, the team grants the presenter more expertise than they have, and the frame in the document becomes the frame the group reasons inside. In a strategic decision that is complex, important, expensive, uncertain, and ambiguous, the chance that a non-expert got the frame right is a coin-toss. Second, when someone arrives with a fully-baked answer, it gets harder to question if the premise of the decision is the right one.

Worst of all, both problems actually feel like a meeting going well.

What this looks like in a strategy discussion

One owner I worked with built a careful pricing model for a new service line and brought it to his team as a decision to approve. AI answered "what should we charge." Anyone who had priced a service like that before would have started a step earlier: which customers do we want to keep, which are we ok losing, and what drives the buying in each segment. The second question changes the first. The model answered only the question asked, and the leadership discussion started at the wrong step. At best, that gets recovered by pushing the AI output aside. At worst, the team reaches a fast decision on the wrong basis, and we all know that errors caught early are the least costly.

What Surface Expertise is, and what to do with the feeling of competence AI gives us

Surface Expertise is the artificial expertise AI enables by doing the production side of expertise, at the finish and confidence of expert work, without the judging, framing, standard-holding, and calibrating substance that real experts have. The surface is complete. The substance underneath is incomplete or absent.

It is the individual version of Cognitive Runaway, and it is how a capable person, with no lapse in judgment, presents what looks like expert knowledge without expert understanding.

Fixing one side of the pair does not reach it. That's the point of Logic Engineering: it changes the pairing.

How do we fight it? Take an honest inventory of the eight jobs and ask which ones you hold in this field. If the answer is the first, plus a borrowed version of the fifth and seventh, your belief likely precedes your competence. Then make it concrete: could I tell a weak version of this analysis from a strong one, without AI pointing the way? If not, the feeling of expertise cannot tell you which parts of the work are right. With strategic decisions that are complex, important, expensive, uncertain, and ambiguous, good enough is probably not good enough, and the ROI of an expert is likely strong.

Frequently asked questions

What is surface expertise?

Surface expertise is expert-looking work without the substance that expertise holds. It happens when a non-expert uses AI to produce work at expert finish, and the judging, framing, standard-holding, and self-calibrating parts of expertise are thin with the AI and the user. The person feels competent because the output matches what competence looks like in their own field, and the people around them read the finish the same way. The term names how capable people present, and accept, AI work they cannot evaluate. It pairs with Cognitive Runaway, which names the organizational result, and Logic Engineering, the organizational answer.

What are the eight jobs of an expert?

An expert does eight things at once, and AI does a partial version of seven of them.

1. Making the work.
2. Judging the work against a standard, what good looks like.
3. Building the frame the situation needs.
4. Judging the question: whether it is well formed, and whether it is the right question at this point.
5. Understanding the standard itself, including the parts nobody in the field has written down.
6. Knowing the edge of their own competence and feeling when a problem sits outside it.
7. Carrying authority, so that people reason inside the frame they endorse.
8. Paying the cost if the expertise is wrong.

AI's version of each is thin on the same part every time: what depends on this particular situation rather than situations like this one. The eighth job AI cannot do at all. A leader can use the list as an inventory: which of these do I hold in this field myself, as opposed to which ones the output did a version of for me.

How is surface expertise different from overconfidence?

Overconfidence is an estimate of your ability running ahead of your actual ability. Surface Expertise is that estimate being right about the wrong thing. The feeling reports, correctly, that expert-finish work was produced. It cannot report on whether the question was right, whether the frame fit, whether the standard was met, or whether the problem sat outside anyone's range, because those readings come from jobs you hold in your own field and do not hold in this one. The two travel together. Surface Expertise often carries overconfidence with it, since AI often enables belief of competence preceding actual competence.

How can a leader tell when AI output is beyond their ability to judge?

Ask whether you could tell a weak version of this analysis from a strong one without AI's help. If you can name what a weak version would get wrong, you hold enough of the standard to judge it. If you cannot, your confidence is reporting on the production, and the work needs a second reader who holds the standard in that field. The eight-jobs inventory is the longer form: which of the jobs do you hold yourself, as opposed to which ones the AI output did a version of for you. If the answer is that you used AI for production, and to know the standard and carry authority, your belief is likely running ahead of competence.

When is AI good enough in place of an expert?

Much of the time. Retrieval, drafting, summarizing, and working inside a framework you already hold are places where the tool outruns a competent person and your bar is a good bar. The band where it is not is narrow: decisions whose answer depends on context that takes expertise to put into the prompt. A fact about people. A thing nobody has admitted. A comparison to a case you have never seen. On those questions the output arrives in the same voice as everywhere else, so the crossing does not announce itself. When good enough is enough, and when it is not, is the subject of a later piece in this series.

How does Surface Expertise affect strategy discussions and strategic planning?

Strategy is where Surface Expertise is most problematic. Strategic questions are complex, ambiguous, important, uncertain, and expensive, and the answer depends on context about people and situation that is often difficult to put into an AI prompt. That is where AI's thin part concentrates, and where a leader's evaluation has the fewest cases to check against. In a planning discussion, the effect shows up in three ways. (1) The conversation starts at the wrong step, because the AI’s plan answered the question asked rather than the one an expert would have started with. (2) The AI’s framing becomes the team's framing, so people are forced to reason inside it. (3) The decision comes fast, because a finished-looking plan invites approval rather than argument. The danger is that the cost surfaces late, since strategic errors take quarters to show.

How do Surface Expertise, Cognitive Runaway, and Logic Engineering fit together?

Surface expertise is the person-level mechanism: a leader presents and accepts AI work they cannot evaluate because it carries expert finish. Cognitive runaway is the organizational result: mediocre output gets accepted as the standard and the standard quietly moves. Logic engineering is the organizational answer: designing the reasoning so the standard is built into the work instead of checked for afterward. Because the risk lives in the pairing of AI and user, improving either side alone does not reach it. Logic engineering is the fix that changes the pairing.

When the answer looked right and you could not say why

If a piece of AI work felt finished and you could not say what would make it wrong, that is a conversation I have with owners most weeks now. I help Second Stage companies treat AI as part of business strategy. Let's talk.