What Consulting Sells, and Which Parts Compress

A consulting engagement bundles several distinct things, and clients pay for the bundle without necessarily distinguishing them.

Research: finding out what is known about a market, a technology, a set of competitors, a regulatory position.

Analysis: doing something with data, which ranges from arithmetic to genuine modelling.

Synthesis: turning a large quantity of input into a small number of defensible claims. This is the part that most resembles what a consultant is thought to do.

Judgement: deciding which claims matter, what to recommend, and what the client can actually execute given who they are.

Communication: the deck, the document, the workshop, the delivery.

And, unstated but frequently the real product, cover. An external party saying it, so an internal decision has an outside endorsement.

AI compresses these at very different rates. Research and communication compress substantially. Analysis compresses where the data is clean and structured. Synthesis compresses partially, and unreliably in a way examined later. Judgement does not compress at all. And cover is unaffected, because its value was never about the labour.

That uneven compression is the whole story for the profession. A firm whose value was concentrated in the first two has a problem. One whose value was in judgement and cover has a pricing problem rather than a value problem, which is different and still uncomfortable.

Mapping the bundle against how much AI actually changes it.

Desk research compresses most. Summarising public material, competitor positions, regulatory landscapes, and academic literature was a large share of junior time and is now substantially faster.

Deck and document production compresses next. Structuring, drafting, formatting, adapting one analysis into three audiences. The work was real and it was never the value.

Analysis compresses conditionally. Where the data is structured and the question is well posed, considerably. Where the data needs assembling from inconsistent sources, which is most client engagements, far less, because the difficulty was never the arithmetic.

Synthesis compresses partially and misleadingly, which is the trap examined in the next lesson.

Judgement does not compress. Knowing which of three defensible recommendations this client can actually execute depends on reading the organisation, and none of that is in any document.

And cover does not compress at all, since its value is that an outside party said it.

The firm-level implication: the pyramid model assumed junior time on the top two rows was billable and necessary. Both assumptions weakened at once.

Desk research is the clearest win and the one with a specific failure mode that matters more here than elsewhere.

What works. Rapid orientation in an unfamiliar sector. Summarising a large volume of public material. Identifying who the players are and what has been written. Producing the structured briefing that used to take a junior consultant two days.

The failure mode. A consultant's output is a claim to a client who will act on it, and a fabricated statistic in a client deliverable is a professional problem rather than an inconvenience. The hallucination cursus applies directly, and the specific risk in consulting is that the fabricated figure is exactly the kind that makes a deck persuasive: a market size, a growth rate, an adoption percentage.

So the discipline is non-negotiable. Every number in a client deliverable traces to a named source the consultant has actually opened. Not a source the model cited, which may not say what the model claimed and may not exist.

That last point deserves emphasis. A model citing a real report does not mean the report contains the figure attributed to it. The most dangerous output is a genuine source attached to an invented number, because a reviewer checking that the source exists will pass it.

The practical rule: the model finds candidate sources, the consultant reads them, and the citation is to what the consultant read. That is slower than it sounds and still much faster than the old process.

Synthesis is where consulting believes its value lives, and it is where AI is most deceptively capable.

What a model does well. Given a large quantity of material, it will produce a structured summary with themes, a narrative, and a set of implications. It reads like synthesis. It is fluent, organised and confident.

What it is actually doing. Producing the most typical structure for material of this shape, which means the themes are the ones that usually appear, the framing is the conventional one, and the implications are the standard implications.

That is not synthesis, it is convention. And it is hard to detect precisely because conventional analysis reads as competent. A client receiving it will find it reasonable, because it is reasonable, and reasonable is what they could have got anywhere.

The distinction that matters. Real synthesis produces the claim that is true and not obvious, which usually comes from noticing that two things in the material contradict each other, or that everyone in the sector believes something the evidence does not support, or that the client's actual constraint is different from the one they named. Those come from a person holding the whole picture and being surprised by something.

A model is not surprised. It has no expectation to violate.

So the practical use is inverted: use it to establish the conventional view quickly, so you know what the obvious answer is, and then spend your time on where the obvious answer is wrong. That is a genuine acceleration of the valuable work rather than a substitute for it.

The commercial consequence is the one the profession finds hardest, and it arrives whether or not a firm chooses to engage with it.

The billable hour prices inputs. If a deliverable that took three weeks now takes one, honest billing produces a third of the revenue for the same output and the same value to the client.

That leaves four positions, and firms are visibly occupying all of them.

Bill the same and deliver faster, absorbing the margin as profit. Works until a client notices, or until a competitor prices on the new cost base, and it is not stable.

Bill less and do more engagements. Honest, and it requires the demand to exist, which is an assumption rather than a fact.

Move to value or outcome pricing. The theoretically correct answer, harder to sell, and it requires the firm to be confident about the value it delivers in a way that hourly billing let it avoid.

And deliver more depth for the same fee. Use the recovered time for the work that was previously cut for budget: more interviews, more scenarios, more validation. This is the position most defensible to a client and the least discussed internally, because it converts efficiency into quality rather than into margin.

The underlying observation. Clients were never buying hours; they were buying an answer they could act on, and hours were the available proxy. AI removes the proxy's usefulness and forces the conversation the profession has deferred for a long time.

Consultants handle other organisations' most sensitive material under obligations that are usually contractual as well as professional, and this constrains tooling more tightly than in most functions.

What is typically in an engagement. Unannounced strategy. Financial data not yet public. Personnel information. Commercial terms with third parties. Material that would move a share price. And frequently material belonging to the client's own clients.

Three specific issues.

The engagement letter usually restricts disclosure to named individuals and prohibits onward transmission. Sending client material to a third-party AI provider is a disclosure, and whether it is permitted depends on terms most consultants have not re-read with this in mind.

Cross-client contamination. A firm serving competing clients has ethical walls, and a shared AI system retrieving across engagements breaches them silently. This is the permission problem from the company brain cursus with a professional-conduct dimension attached, and it is the most serious risk in this lesson.

And retention. Prompts and outputs held by a provider are copies of client material outside the boundary the engagement letter contemplated.

The practical positions. Check the engagement terms before adopting, not after. Prefer arrangements with no retention and no training. Segregate by engagement rather than pooling. And tell the client, because a client who learns later that their strategy went through an external system will not accept that it was technically permitted.

The client side of this is moving, and consultants are better served anticipating the questions than reacting to them.

What clients increasingly ask. Was AI used in producing this, and how. What happened to our data. Are we paying senior rates for work a model did. And, more pointedly, why are we paying you for research we could now do ourselves.

That last question is the real one, and it deserves a real answer rather than a defensive one.

The honest answer for a good firm. You are not paying for the research; you were never paying for the research. You are paying for someone who has seen this problem in eleven other organisations, who can tell you which of the obvious answers will fail in your specific context, who will say the thing your internal team cannot say, and who is accountable for the recommendation.

A firm that cannot give that answer has learned something important about its own value proposition, and the discovery is better made internally than in a procurement conversation.

The practical consequence for how engagements are sold. Lead with the judgement and the experience rather than the deliverable volume. A proposal promising forty pages of analysis is competing on a dimension that is collapsing in price. One promising a defensible recommendation from someone who has done this before is not.

And expect procurement to start asking for the AI disclosure explicitly, since it is already appearing in professional services tenders.

An adoption order for a consultant or a firm.

First, desk research with the sourcing discipline. Rapid orientation, landscape summaries, literature scans, with every number traced to a document someone opened. This is the largest immediate time recovery and the discipline is what keeps it safe.

Second, document and deck production. Structuring, drafting, adapting one analysis for three audiences. Real time, no judgement involved, low risk.

Third, interview and workshop synthesis. Transcription and clustering across many conversations, which is genuinely hard by hand and where the volume defeats people. Read the outliers rather than trusting the clusters.

Fourth, the conventional-view exercise. Deliberately generate the obvious analysis so you know what it is, then spend your effort on where it is wrong. This is the inversion from the synthesis step and it is the most sophisticated use in the list.

And before any of it: check the engagement terms, segregate by client, and decide the disclosure position.

What not to do. Present generated synthesis as your analysis. Put an unverified figure in a client deliverable. Pool client material across engagements. Or assume the pricing question will not arrive, because it will arrive from a client who has done the arithmetic.