Producing the Work: Decks, Analysis, and Proposals

Deck production consumes an enormous share of consulting time and almost none of the value, which makes it the obvious target and worth being precise about.

What compresses well. Structuring an argument into a slide flow once the argument exists. Drafting the body text. Writing the executive summary from the detail. Adapting one analysis for a board, a working team and a steering group. Producing the appendix nobody reads and everybody requires.

What does not. Deciding what the deck should say. The storyline is the analysis, and a model asked to produce a storyline will produce a conventional one, which is the synthesis trap from the previous lesson in its most common form.

The practical division that works. The consultant writes the argument as a sequence of assertions, one per slide, in plain sentences. That is the thinking, and it takes as long as it takes. The model then builds the slides from those assertions, drafts the supporting text, and produces the audience variants.

Two cautions. Generated slide text tends toward the generically consultative, so it needs editing for specificity, and specificity is what makes a deck persuasive. And check the variants agree: an executive version generated from a detailed one tends to smooth away the caveats, which is the same failure the project management cursus identifies.

The honest observation for the profession. If a deck's value was its production quality, that value has fallen. If it was the argument, nothing changed.

This is the strongest genuine capability in consulting work and the one most under-exploited.

The problem it solves. A diagnostic engagement runs thirty interviews. Each produces pages of notes. No individual reads all of them carefully, so the synthesis is assembled from what the interviewers remember, which is dominated by the most recent, the most articulate and the most senior.

What AI does well here. Transcription, so interviewers stop taking notes and listen. Clustering across all thirty transcripts into themes. Surfacing which concerns recur and which are held by one person. Identifying where accounts of the same event differ, which is frequently the most valuable finding in a diagnostic.

That last capability is worth dwelling on. When two departments describe the same process differently, the discrepancy is the finding, and it is exactly what gets lost when synthesis happens from memory.

The cautions. A thematic summary under-represents the view held by two people, and in an organisational diagnostic the minority view is frequently the accurate one, since the majority may share an institutional blind spot. So read the outliers deliberately.

And interview material is unusually sensitive. People say things about colleagues and about the organisation on the understanding that it will be aggregated, so a searchable transcript store is a different artefact from a synthesis report, and it should be treated accordingly and deleted when the engagement closes.

Tell interviewees it is being recorded, and mean the confidentiality you promise.

The discipline that keeps a deliverable defensible, drawn as the path every claim must survive.

A claim enters the deliverable from one of three places. From a source the model surfaced, from the client's own data, or from the consultant's judgement.

Model-surfaced claims must pass the hardest gate: the consultant opens the source and confirms it says what was attributed to it. A model citing a real report does not mean the report contains the figure, and this is the check that is routinely skipped because the citation looks legitimate.

Client-data claims must trace to a query someone can rerun, so the number is reproducible rather than asserted.

Judgement claims must be labelled as judgement. A recommendation presented with the same visual weight as a sourced fact is claiming an authority it does not have, and clients are entitled to know which is which.

What reaches the deliverable carries its provenance. What cannot be traced does not go in.

The rule is simple and the practice is not, because the pressure to include a persuasive number you have not verified is real and constant.

Quantitative analysis compresses less than expected, and understanding why prevents disappointment.

The assumption is that the hard part is the modelling. In most engagements it is not. The hard part is getting the data into a state where it can be analysed at all: reconciling inconsistent definitions between systems, deciding what a field actually means, handling the three years where the categorisation changed, and establishing which of two conflicting sources to trust.

That work is judgement about a specific organisation's data, and a model has no access to the institutional knowledge required. Asked to reconcile two figures, it will produce a plausible reconciliation, which is worse than no reconciliation because it looks resolved.

Where it does help. Writing the analysis code once the data is understood. Explaining what a technique does and whether it suits the question. Checking an approach for the assumption you forgot. Producing the sensitivity analysis you would have skipped for time. And drafting the explanation of a result for a non-technical audience.

That last one is genuinely valuable, since explaining a regression to a board is a skill and the draft saves real effort.

The caution that matters most. A model asked to interpret a result will produce a causal story, and the marketing cursus makes the general point. In a consulting deliverable a causal claim carries more weight, because the client will act on it. Correlation described as driver is the most common analytical error in consulting decks and generation makes it easier to produce fluently.

Proposal writing is high-volume, largely templated, and mostly unbilled, which makes it an attractive target with a specific risk.

What works well. Assembling the standard sections from prior proposals: credentials, methodology, team biographies, the boilerplate that must be present and that nobody reads. Tailoring the language to a sector. Producing the first draft of a response to a structured tender, which is largely a matching exercise. And checking a draft against the tender requirements to catch an unanswered question, which is a common and expensive omission.

Where the risk sits. A proposal is a contractual document in the making, and generated content can commit the firm to scope, timelines or outcomes nobody agreed. A methodology section describing an approach the team does not actually use, or a timeline the model produced by convention, becomes what the client expects and what the engagement letter may reference.

So the discipline is that scope, price, timeline and any commitment are written by a person and never generated. The boilerplate can be assembled; the promises cannot.

And the differentiation problem from the marketing cursus applies here directly. If every firm generates proposals from similar prompts, proposals converge, and the ones that win are the ones that say something specific about this client's situation. Generated proposals are systematically less specific, so the efficiency gain and the win rate can move in opposite directions.

Use the recovered time on the sections that differentiate rather than producing more proposals.

Consulting firms accumulate a large body of prior work and reuse almost none of it systematically, which is the company brain problem in a setting where it is unusually costly.

The waste is specific. A firm has done this analysis before, for a different client, in a different year. The consultant doing it now does not know that, cannot find it, and rebuilds it. Meanwhile the person who did it has left.

What retrieval over prior engagements offers. Finding the analysis that already exists. Surfacing the methodology that worked. Locating the person who has done this before, which is often more valuable than the document. And identifying that the firm has a view on a question, so a new proposal does not contradict a previous one.

The constraint that makes this hard, and it is severe. Client confidentiality means prior engagement material cannot simply be pooled. Ethical walls between competing clients must hold. And engagement letters frequently restrict internal reuse in ways the firm has not catalogued.

So the workable version is narrower than the ambition. Retrieve over sanitised methodology and approach rather than client-specific findings. Retrieve over who did what rather than what they found. And where client-specific material is included, enforce permissions per engagement rather than per firm, which the company brain cursus establishes as an early-binding access control problem.

Done within those limits it is still substantial, and it addresses a waste every consultant recognises.

The profession's development model is unusually exposed, because it was built on exactly the work that compressed.

How consultants used to learn. Do the desk research, which teaches how to find things and what sources are worth. Build the model, which teaches what the numbers mean. Produce the pages, which teaches how an argument is constructed. Sit in the interviews taking notes, which teaches what to listen for. Years of that, and gradually the judgement forms.

The compression removes most of it. A junior consultant with good tooling produces in a day what took a week, and the metrics look excellent. What does not develop is the underlying sense of when an analysis is wrong, because that came from doing analyses and being corrected.

And the profession's economics compound the problem. The pyramid depended on leveraged junior time being billable. If that time is worth less, firms hire fewer juniors, which narrows the pipeline that produces partners.

The partial mitigations are the ones the managing-teams cursus names. Move juniors up the stack deliberately rather than by default, which here means having them critique generated analysis against a standard rather than produce it. Preserve unassisted work explicitly for development, accepting the cost. And teach the judgement directly rather than expecting it to accumulate, since the accumulation mechanism has gone.

The honest position for a firm. This is a real problem, the cost lands in five years, and nobody currently has a convincing answer.

Pulling the production workflows into something a consultant can adopt.

Before anything: check the engagement terms, segregate material by client, and decide what you will tell the client.

On research. Model surfaces candidates, you read them, citations are to what you read. Every number in a deliverable traces to an opened source. No exceptions, because the one exception is the one that appears in a slide someone screenshots.

On interviews. Record with consent, transcribe, cluster across all of them, and read the outliers and the contradictions yourself. Delete the transcripts at engagement close.

On analysis. Do the data reconciliation yourself, because that is where the judgement is. Use the model for the code, the sensitivity analysis you would have skipped, and the explanation for a lay audience. Never let it interpret causally.

On decks. You write the assertions, the model builds the slides. Edit for specificity. Check the audience variants agree on facts.

On proposals. Assemble the boilerplate, write the commitments. Spend the recovered time on the differentiating sections rather than on more proposals.

And throughout, label judgement as judgement. A deliverable that distinguishes what is sourced, what is computed and what is the consultant's view is more useful to a client and more defensible to a reviewer, and it is what the profession's authority actually rests on.