The prompt is no substitute for method — why formalising tacit expertise has become the real work
Sixth text in the cognition / doctrine block. After true agentics, a question that confronts every organisation that has invested eighteen months of prompt engineering training without seeing the return: what that investment should have produced to reach the promised productivity.
The audit that found nothing
In a large French consulting house, in April 2026, management decides to run an internal audit on the use of AI on engagements. The scope is narrow, thirty consultants, selected for their regular exposure to the tools. All have been trained in prompt engineering for eighteen months, some since the beginning. Several hold external certifications. Management wants to know, after this volume of investment, what AI has actually changed in the production of their engagements.
The report runs to three pages. No measurable time savings on engagement duration. No measurable improvement in deliverable quality. No productivity gap between the trained consultants and those who were not. A few anecdotal accounts of one-off gains, never converted into an organisational effect.
Management commissions a second reading of the report. It finds no methodological error. It then commissions a third reading, by an outside firm, on the assumption that a self-justification bias had skewed the measurement. The outside firm confirms it. Prompt engineering training, measured against operational productivity, produced nothing visible.
This result is not a local anomaly. The surveys published in 2025 by McKinsey, BCG and the BCG Henderson Institute, for all their officially enthusiastic tone, converge on the same unsettling figure: fewer than fifteen percent of organisations that deployed internal prompt engineering training saw an improvement in their overall productivity. The industry average hovers around zero. And no one, neither the vendors nor the training firms nor the trained leadership, states clearly why.
This article holds to one hypothesis. Prompt engineering, as taught since 2024, optimised the wrong variable. It taught operators to phrase better, when what they lacked was not phrasing but a layer beneath it, more fundamental, that no training could deliver.
Four layers of a single gesture
To understand where the work has shifted, you have to look at what a prompt has become over three years. The word has come to cover four different objects that training programmes continue, wrongly, to treat as one.
The first layer is that of direct instruction. Summarise this text. Translate this paragraph. List the main arguments. This layer, which made up the bulk of usage in 2023, calls for no particular skill. Any model above a certain threshold now produces what is expected, and the margin for improvement stopped being significant eighteen months ago.
The second layer is that of the resume. You tell the agent who it is. You are a senior strategy consultant. You have fifteen years of experience on industrial transformation matters. You command the Porter, McKinsey 7S, and value-chain analysis frameworks. This layer, popularised by the prompt engineering guides of 2024, produced a real jump in quality. The agent becomes more precise in its vocabulary, more cautious in its assertions, more aligned with a professional register.
But the resume has a limit few have named. Enumerating competencies does not produce a decision. A senior consultant is not distinguished from a junior by the frameworks they know, since both command them. They are distinguished by the sequence of trade-offs they make silently when they approach a new situation. Which elements to look at first, which signals to set aside, at what moment to switch from a diagnostic reading to a prescriptive one. None of these gestures appear in any resume.
The third layer, which took hold from late 2024 onward, is that of the operating procedure. Here you describe what the agent must do, in what order, setting the conditions for moving from one step to the next. First, analyse the history. Then, identify the three critical variables. Next, cross-reference with the sector context. Finally, propose three scenarios. This layer produces another jump, deeper than the previous one, because it begins to transfer no longer an identity but a sequence — that is, a fragment of process.
And yet the operating procedure remains insufficient. It describes the steps, not the trade-offs. It does not say, for instance, at what signal the analysis must be interrupted and the framing changed because the initial question was badly posed. It does not say how to resolve the tension that arises when two constraints of the brief compete. It says nothing of the operation most typical of expert judgement, which consists precisely in knowing when to step out of the plan.
The fourth layer, which the best teams began to experiment with in 2025 but which we still hesitate to name, is that of the hierarchy of principles. Here you inscribe no longer what the agent does, but what it gives priority to when two objectives come into tension. Absolute priority to contractual rigour over rhetorical elegance. Priority to traceability over concision. If two options are equal on the substance, keep the one that best withstands public contradiction. It is this fourth layer, and it alone, that produces an agent capable of judging.
These four layers do not stack linearly. The fourth presupposes the previous three, but it also demands something none of the others required. A tacit expertise previously formalised by the one who holds it.
From resume to operating procedure
The shift from the second to the third layer can be measured in the concrete case of bid management. A senior bid manager asked about their method, in 2024, enumerates what they know. The typical structure of a public tender. The usual scoring ratios. The classic pitfalls of a badly read CCTP. Their resume can be written in a few pages, and an agent instructed this way produces, on a simple file, an acceptable first response.
The same bid manager, asked in 2025, senses that their resume is not enough. They begin to describe a sequence. First, I read the consultation rules to identify the criteria and their weighting. Then, I read the CCTP looking for requirements hidden in the prose. Next, I cross-reference with what I know of the client to infer what the published scoring does not say. This operating procedure, far richer than the resume, lets the agent structure its work, no longer produce responses that look like copy-pasted marketing.
But the operating procedure quickly shows its own limits. On a complex file, the agent instructed this way conscientiously runs through the steps and produces, in the end, a deliverable that misses the essential. Why. Because it missed the weak signal that the senior bid manager would have caught on the third page, that would have led them to abandon the initial plan and rebuild the framing. The agent follows the plan. The human expert, for their part, knows when to step out of the plan. And that when fits nowhere in an operating procedure.
This limit shows particularly in what we will here call the steering mode. For bid managers who step in mid-engagement, after an AI agent has already produced a first version on which a call must now be made, the agent's operating procedure is never enough. At each trade-off, one has to make explicit on which principle the call is made, because the agent cannot inherit that intuition.
The hierarchy of principles
It is here that the fourth layer, that of the hierarchy of principles, begins to produce a qualitatively different effect. Rather than prescribing to the agent what it must do, you transmit to it what it must arbitrate on when two competing directions present themselves. When the CCTP is ambiguous on a critical point, never fill the gap with a favourable interpretation; state the ambiguity as an explicit assumption. When the commercial brief comes into tension with the contractual requirement, the contractual one prevails, unless the sponsor explicitly endorses the deviation.
These formulations, which look more like doctrine than instruction, demand of the one who writes them an operation that had never been required before. To identify, in one's own practice, the moments where one decided, and to reconstruct in words the trade-off one made. This operation is the opposite of what is asked in a resume. The resume lists what one knows. The hierarchy of principles formalises what one does without saying it.
And it is there, in this passage from knowing to tacit-doing-made-explicit, that AI productivity tips over. The rare teams that produce spectacular results with AI in 2026 are not those with the best prompts. They are those whose experts spent a considerable number of hours, upstream, formalising their own method. The prompt is only the final transcription of an explicitation work that came before.
The Polanyi paradox
This work of explicitation runs into a difficulty that the philosophy of knowledge identified sixty years ago. Michael Polanyi, in his 1966 work The Tacit Dimension, defended the thesis that the essence of what an expert knows cannot be put into words. His famous formula, we know more than we can tell, held that tacit knowledge structurally resists verbalisation, and that it is this resistance that constitutes it. An expert who had verbalised the entirety of their method would have ceased to be an expert and become a manual.
Hubert Dreyfus, following Polanyi, extended the thesis by applying it to the limits of the artificial intelligence of the 1980s. His arguments against the algorithmic formalisation of expertise were so strong that they long seemed to bar AI from any access to the professions of judgement. LLMs have rendered these arguments partly obsolete, but on one point only. They have shown that a system trained on massive quantities of practice can produce outputs that imitate judgement, without having formalised anything at all. The method remains tacit, in the model as in the human expert.
The paradox is therefore reformulated. For AI to produce expert-quality work on a given engagement, it must inherit a method that was never written down. The vendor trained the model on the tacit. The user, in turn, must explicitly transmit what, in their own practice, distinguishes their craft from a sector average. And this operation of transmission demands precisely what Polanyi held to be impossible: the formalisation of the tacit.
Where Polanyi was right, and where he was wrong. He was right that tacit knowledge resists complete formalisation. No expert can write the entirety of their method without losing something of it. But he underestimated one empirical fact. Partial formalisation, even imperfect, is enough to transform an agent's output. It is not necessary to say everything. One must say what has never been said, and what the agent could not deduce on its own from its training.
This effort of partial formalisation is what separates, in 2026, the organisations that draw a real return from AI from those that have plateaued. The former accepted to pay the cost of a new, heavy, thankless work. The latter believed that prompt engineering training could spare them from it.
What TenderGraph calls the method layer
This analysis, conducted on the observation of the past eighteen months, shaped the design of TenderGraph TITAN. We made the choice, from the very first version shipped in 2025, not to ask the user to write their hierarchy of principles themselves. The bet was simple. If the formalisation of the tacit is what holds adoption back, then the method must be delivered already formalised, and the user left with the sole responsibility of supplying the material of the file.
This pre-designed method, which we internally call the method layer, encodes thirteen phases of tender analysis, a hierarchy of principles that decides on the typical trade-offs of the craft, and hard divergence rules for when a file presents an atypical configuration. It was written with senior bid managers who agreed, over months, to revisit their own decisions in order to extract the silent rules from them.
The user, then, no longer has to learn prompt engineering. They do not have to formalise their craft. They supply the DCE, follow the sequence, and decide at the moments where the method hands control back to them. This architectural choice, which seems obvious once stated, contradicts the dominant doctrine of AI vendors in 2025, which consists in delivering a generalist agent and asking the user to produce their own method by prompt.
We hold that this latter model cannot work at scale, because it underestimates what Polanyi had pointed out. The vast majority of practitioners are not capable, and have no vocation to become capable, of making their own method explicit. To ask this is to ask doctors to become epistemologists, lawyers to become grammarians of law, bid managers to become doctrinaires of public procurement. It is possible in theory. It is rare in practice. It is massively costly within organisations.
The right question at the wrong moment
Let us return to the consulting firm's audit with which this article opened. Why did prompt engineering training, delivered to thirty consultants over eighteen months, produce nothing measurable. Three readings of the report sought a methodological bias. None found one. The fourth reading, the one no one dared to conduct, would have been the right one.
Prompt engineering training taught the consultants to phrase. It could not, structurally, teach them to make their own method explicit. That operation calls for a work of professional introspection, followed by a laborious putting into words, which does not happen in two days of inter-company training. It happens over several weeks, with support, by revisiting real files to reconstruct the trade-offs made. No 2024 or 2025 training catalogue offers such a format. And for good reason. It is not prompt engineering.
The leaders who believed they were buying AI productivity by buying prompt engineering training asked the right question at the wrong moment. They should have started by asking whether their experts' method was formalised, and therefore transmissible to an agent. Since the answer is almost always no, the next investment should have gone toward explicitation, not phrasing. In hindsight it is obvious. It was not in 2024.
The lesson is not only budgetary. It is doctrinal. Prompt engineering, as a discipline, rests on the premise that the quality of the AI output depends on the quality of the phrasing. This premise was true on layers 1 and 2. It is false on layers 3 and 4. And it is on these last layers that real productivity now resides.
Coda — what formalising does to the expert
One last observation, perhaps the most interesting, deserves to be set down. The experts who agreed, by constraint or by curiosity, to formalise their own method in order to transmit it to an agent come out of it transformed, and this transformation had not been anticipated.
The work of explicitation, in effect, forces one to revisit trade-offs one used to make without thinking. To name the exceptions one made to one's own rules. To identify the biases that practice had sedimented. Several senior bid managers we worked with said, after the exercise, that they had become better at their own craft by having tried to teach it to a machine. They had discovered, in forcing themselves to formalise, inflections of their practice they would not have seen otherwise.
This is, perhaps, the principal benefit of the decade now opening. AI forces what organisations had never managed to do spontaneously, the systematic explicitation of the tacit knowledge of senior staff. It forces it not through pedagogy nor through a culture of knowledge management, but through a simple operational necessity. Without this explicitation, the agent produces nothing useful. And where pedagogy failed for thirty years, technical constraint succeeds.
The prompt is no substitute for method. It is only its final writing. The method, for its part, lives in the heads of experts who never had the occasion nor the need to write it down. The work now opening is not to phrase better. It is to rediscover what one was doing, without knowing it, for years.
Bibliography
On tacit knowledge and its limits
- Michael Polanyi, The Tacit Dimension, 1966.
- Michael Polanyi, Personal Knowledge, 1958.
On the philosophical critique of classical AI
- Hubert L. Dreyfus, What Computers Still Can't Do, 1992.
On the 2025-2026 prompt engineering ROI surveys
- McKinsey Global Institute, The State of AI in 2025, annual report.
- BCG Henderson Institute, Generative AI in the Enterprise: Beyond the Hype, 2025.
On the formalisation of expertise in the professions of judgement
- Donald Schön, The Reflective Practitioner, 1983.
- Richard Sennett, The Craftsman, 2008.