If a Client Plus Claude Plus One Engineer Can Do It, the Work Isn't Yours Anymore

There's a diagnostic in CMap's 2026 Boutique Consulting Operating Playbook, attributed to Florian Heinrichs, that I think is the most useful single question in the document:

For each of your top three service lines, ask whether a competent client plus Claude plus a forward-deployed engineer could deliver 80% of it. If yes, that service is going to be repriced by the market – or removed from it – within 18 months.

That's a good test. It's specific, it's runnable this afternoon, and it has a failure condition, which most strategic advice doesn't. I want to take it seriously and then argue with the answer the playbook gives.

Why the test is well-built

Notice what it isn't asking. It isn't asking whether AI can do your job, which is an unanswerable question that generates a lot of content and no decisions. It's asking something narrower and checkable: could a specific, realistic team – your actual client, a good model, and one capable engineer who shows up and sits with them – get to 80% of a specific deliverable you currently sell.

Eighty percent is the right threshold too, because it's the point where the remaining 20% stops supporting your price. A client who can get most of the way there will pay for the gap, and the gap is worth a fraction of the whole.

The market is already showing this in the place you'd expect. McKinsey is reducing headcount by about 10%, roughly 3,000 to 4,000 roles staggered over 18 to 24 months, concentrated in back-office and junior research functions. Its CEO has described the split directly: client-facing roles growing about 25% while non-client-facing roles shrink about 25%, with output from the shrinking side up 10%. That's not a firm in trouble. That's a firm shedding the decomposable layer on purpose and keeping the layer that isn't.

Worth a small correction, since it circulated at the conference in a bigger form: the number is around 3,000 to 4,000, not 5,000, and the firm's own statement frames it as improving the efficiency of its support functions during a period of rapid AI advance, alongside five years of flat revenue and pressure in a couple of major markets. The direction of the story holds up. The magnitude and the attribution both got rounder in retelling, which is its own small lesson about numbers that support a thesis you already like.

Where I part company with the answer

The playbook's answer to its own test is pillar four: point of view plus proprietary data plus consistency is the only moat that strengthens as AI strengthens. Lead with something AI demonstrably cannot copy.

That's close, and it's the answer most people in this conversation give, and it's soft in a specific way. A point of view is an output. Proprietary data is an asset. Both are copyable on a delay – the point of view gets absorbed into the next model within a year, and the data advantage lasts exactly as long as nobody else assembles a comparable set. Neither of them is a reason to hire a person.

What actually survives the test isn't a thing you produce. It's that someone is answerable for the result.

I'd put it in three parts. A system that shows the work, so the client can check it rather than trust it. Judgment from someone who has made this call before and will say so when the math doesn't work. And a name on the recommendation. A free tool has no reputation at risk, which is precisely why its confident yes is worth so little – it costs the tool nothing to be wrong, and it costs you a year.

The third one is the load-bearing leg and it's the one that can't be decomposed. You can generate a point of view. You cannot generate accountability, because accountability isn't a property of the output, it's a property of there being somebody who loses something if it's wrong.

The test the playbook gets exactly right

There's a second diagnostic in the same document, from Joe O'Mahoney, on whether a niche is any good. Three properties, and the first one is the whole argument:

The client can't solve it themselves. If they could, you're competing with their own internal team – and AI is making that easier.

The other two are that the buyer has real budget, and that you can charge what the expertise is worth rather than body-shop rates. But the first property is the same test as the Claude question, stated as a positive rather than a negative, and it's the cleanest statement of what advisory work is for that I've read this year. If they could answer it themselves, they would have. You're there because it's hard and because getting it wrong is expensive.

Running it honestly is harder than it sounds

The test is easy to state and easy to fudge, and the fudge has a shape. You picture your best client, on your best engagement, and you ask whether they could have done it with a model. Of course not – that engagement had three judgment calls in it that you can still remember.

Run it the other way. Picture your most capable client, the one with a sharp internal team and a head of data who's been asking pointed questions, on your most routine engagement. The one you could scope in a phone call because you've done it eleven times. That's the honest test, and it's the engagement most likely to be half your revenue.

The other common dodge is counting the relationship as part of the work. It isn't, for this purpose. The question isn't whether the client would prefer to work with you, it's whether the deliverable is reproducible without you. Preference erodes under a large enough price difference, and the price difference here is not small.

One more thing worth checking, because it cuts the other way: a service line can pass the test today and fail it on the deliverable rather than the thinking. If the expensive part of your engagement is producing the artifact – the deck, the model, the written synthesis – and the valuable part is the twenty minutes where you tell them what it means, then the work passes on substance and fails on cost structure. That one doesn't get repriced away. It gets unbundled, and you're better off unbundling it yourself.

What to do with a service line that fails

Most firms reading this will have one, and the instinct will be to defend it by adding quality. Resist that. A service line that fails the test doesn't fail because your version isn't good enough – it fails because the client's version is now good enough, and improving yours from an 8 to a 9 doesn't change the arithmetic.

Three honest options. Retire it, and stop selling something whose price is about to be set by a model. Fold it in as an input to work that passes the test, delivered as part of a larger engagement and priced accordingly rather than sold on its own. Or keep selling it deliberately at the new price, as capacity rather than expertise, with your eyes open about the margin.

What doesn't work is holding the old price on the old positioning and waiting to see whether clients notice. They notice. Increasingly they notice by not deciding, which looks like a slow pipeline rather than a lost argument, and it's much harder to learn from.

Run the test on your top three this week. If all three pass, you're in better shape than most firms and you should write down why, because that reasoning is your actual positioning. If one fails, you have roughly eighteen months and the useful part is that you found out from a question rather than from a renewal conversation.