The Best AI Diagnostic May Be the One That Proves Itself Wrong

Consulting is racing to turn diagnostics into software. The bigger opportunity may be to use AI differently: not to explain the client faster, but to discover where our explanation stops working.

A consulting idea is becoming almost irresistible.

Take an analysis that once required a small team, package it into software, connect a few data sources, add an AI layer, and give the client something useful before they have bought anything. It can be an inventory diagnostic, a pricing assessment, a working-capital tool, an AI maturity review, a resilience scan, or almost any other piece of structured analysis that consulting firms have spent years turning into methodologies.

The commercial logic is excellent. What previously required weeks can increasingly be produced in days and consumed in minutes. Instead of arriving with slides explaining what we could do, we can arrive with software that already does some of it.

I understand why everybody is moving in this direction.

I am just not convinced that we are automating the right part of consulting.

If every firm can create a respectable diagnostic at very low cost, the diagnostic itself cannot remain the differentiator for very long. More importantly, the software risks making a bigger mistake: it can produce an increasingly convincing explanation of a company before it has built a sufficiently rich model of how that company actually works.

That creates three problems, and I think the interesting opportunity for AI starts when we reverse each of them.

The first problem: client compression

Every strategic diagnostic begins by reducing the company to something manageable to be analyzed.

That is unavoidable. Consultants do it as well. We select variables, define categories, normalize data, choose a level of granularity, and decide which relationships matter enough to model. Without compression, there is no analysis.

The problem starts when the compressed representation becomes the company.

Imagine a working-capital diagnostic that receives inventory, forecast accuracy, service levels, lead times, and supplier performance. It identifies €80 million of apparent excess inventory and finds a strong relationship between forecast error and stock levels.

The conclusion almost writes itself: improve forecasting and reduce safety stock.

But perhaps the products with the highest inventory are also supporting launches. Perhaps they use a CMO with fixed campaign sizes. Perhaps release lead times are unusually long in three markets. Perhaps a customer contract requires a service level that makes the stock economically rational. Perhaps the business intentionally built inventory ahead of a manufacturing transfer.

None of those facts makes the original analysis mathematically wrong.

That is what makes the problem interesting.

The numbers may be perfectly correct inside a strategically incomplete model.

A conventional diagnostic tends to hide that incompleteness because its purpose is to produce an answer. An AI-generated one can make the problem worse because it can turn an incomplete representation into a beautifully coherent explanation at extraordinary speed.

So I would reverse the objective.

Instead of asking AI to explain the €80 million, ask it to find where its own explanation fails.

If forecast error explains the inventory, why do other products with the same forecast error behave differently? If long lead time causes the issue, why does the pattern disappear in another plant? What characteristic separates the observations that follow the model from those that refuse to follow it?

That is a much more useful question.

The purpose of the diagnostic should not be to conceal the compression. It should be to expose its boundary.

The second problem: convergence

The second issue is more commercial.

AI is very good at reproducing patterns that exist repeatedly across industries, companies, and bodies of management knowledge. Ask several systems to create an inventory diagnostic for a pharmaceutical business, and I would expect most of them will converge on familiar variables: service level, forecast accuracy, safety stock, lead time, supplier reliability, obsolescence and segmentation.

That is not a criticism of the technology. It is one of its strengths.

But it creates an uncomfortable question for consulting firms.

If everybody can encode the same frameworks into software, what exactly is proprietary?

A clever diagnostic may create differentiation for a while, but the half-life of that advantage is collapsing. A tool that might once have taken a consulting firm two years to build can increasingly be reproduced after somebody has seen what it does.

The wrong response is to search desperately for a more complicated framework.

The better response is to use convergence deliberately.

If AI is good at constructing the expected pattern, then let it.

Use the generic model as a reference case and concentrate analytical effort on the points where the client departs from it.

Suppose the normal relationship highlights that greater demand variability should require more inventory. Most of the client’s portfolio behaves exactly as expected, but one group of products does not. They carry substantially more inventory than comparable products even after accounting for demand variability, service targets, and lead times.

That residual is more interesting than another benchmark.

Perhaps those products share a packaging configuration. Perhaps they came from the same acquisition. Perhaps different allocation rules govern them. Perhaps they pass through a particular release process that nobody considered important when the original model was designed.

The strategic signal is not simply that two variables are related. Companies contain thousands of relationships.

The interesting moment is when the current model needs a new concept to explain what it sees.

That is a very different use of AI.

The generic pattern is no longer the insight. It becomes the control against which the client’s specificity becomes visible.

The third problem: verification

The third problem is the one I would worry about most if these tools are used as consulting entry points.

An AI system can generate more findings than a consulting team can realistically verify.

Imagine that a diagnostic produces twenty recommendations. Fifteen are sensible, four are obvious, and one fundamentally misunderstands the business. In an executive discussion, that one recommendation can destroy confidence in the other nineteen because the client is not evaluating the model statistically. They are evaluating whether we understand their company.

This creates a confidence debt.

The marginal cost of producing another recommendation is collapsing, while the cost of establishing that the recommendation deserves to be trusted is not.

Someone still has to reconcile the data, test the assumptions, understand the history, speak to the people running the process, and determine whether an observed relationship is causal, contingent, or simply accidental.

So here again I would reverse the design.

Do not ask the system to jump from observation to recommendation. Ask it to check and remove explanations that no longer fit.

Inventory is 18% above expectation. The increase is concentrated in externally manufactured products. Yet several externally manufactured products do not show the same behavior. The difference appears concentrated in products that experienced regulatory changes. Release-cycle data is missing.

At this point, the right output is not “reduce inventory.”

It is:

Did release lead time change before inventory policy changed?

That may look less impressive on a dashboard, but analytically it is far more mature. We have moved from one broad correlation to a narrower causal question without pretending that we have already answered it.

AI does not need to prove the explanation. It can create value by shrinking the number of plausible explanations.

The real opportunity is model emergence.

This is where I think these technologies become much more interesting for consulting.

Most current diagnostics are designed around answer density: more findings, more recommendations, more red and amber boxes, more things that appear to have been discovered.

I would design almost the opposite.

Start with a provisional model of the business. Let AI identify the relationships the model explains reasonably well, and focus on the residuals: the products, customers, plants, transactions or decisions that remain unexplained.

Those residuals should trigger questions.

A client executive may look at four unusual products and say, “Those all came from the same acquisition.”

That single sentence introduces a variable that did not exist in the original model.

Acquisition origin now becomes an entity worth testing. Perhaps it explains a different planning policy, a different supplier contract, or a different regulatory architecture. The model changes, the unexplained space becomes smaller, and a new anomaly appears somewhere else.

That is the part I find strategically interesting.

The system is no longer just running an analysis on the client. It contributes to the building of a better representation of the client.

And the consulting engagement becomes a process of model emergence. It tells what the current model can explain, where it fails, and the next piece of context that would most change our understanding.

That is much harder to commoditize because the value does not sit in the original framework. It accumulates as the model becomes increasingly specific to the company.

Which also changes the entry point

If this is right, then the commercial proposition changes too.

The traditional free diagnostic says:

“Look how much we already know about your business.”

That may work while the tool itself still feels novel, but novelty is not a durable consulting strategy.

A stronger proposition would be:

“Look at the parts of your business that the obvious explanation cannot account for.”

That creates a very different conversation.

The client does not have to trust that the software already understands the organization. They only need to recognize that something important has appeared that neither the generic model nor the current data can explain properly.

From there, every interaction can make the model better:

baseline → anomaly → question → context → revised model → new anomaly.

The application becomes less generic with use rather than pretending to be unique at the beginning.

And that may be the real opportunity.

AI will make diagnostics cheaper. It will make frameworks easier to encode and hypotheses almost free to generate. A meaningful part of what consulting firms once treated as intellectual property will probably become commodity work.

That is not necessarily bad news.

It migrates the scarce capability elsewhere.

The next generation of consulting AI should not compete on how quickly it can tell the client what is happening. It should compete on how quickly it can discover that its first explanation is insufficient, identify what the model is missing, and direct the next question toward the piece of information that matters most.

Sometimes the most valuable thing an AI diagnostic can discover is not an answer at all.

It is the point at which the current explanation stops working.

________________________________________________________________

Disclaimer: Views or opinions represented in this article are personal and belong solely to the article writer and do not represent those of people, institutions or organizations that the writer may or may not be associated with in professional or personal capacity, unless explicitly stated.

_________________________________________________________________

Frequently Asked Questions

What is the main problem with turning consulting diagnostics into AI software?

While automating diagnostics into software is commercially attractive, it risks producing convincing explanations of a company before building a sufficiently rich understanding of how that company actually works. This can lead to recommendations based on incomplete models that miss critical context.

What is 'client compression' in consulting diagnostics?

Client compression is when consultants reduce a company to manageable variables for analysis by selecting data, defining categories, and deciding which relationships to model. The problem occurs when this compressed representation becomes mistaken for the actual company.

Can a diagnostic be mathematically correct but strategically incomplete?

Yes. A working-capital diagnostic might correctly identify excess inventory mathematically, but miss crucial context like product launches, supplier constraints, customer contracts, or intentional strategic decisions that make the inventory levels economically rational.

Why can't diagnostics remain a differentiator if every consulting firm creates them?

If every firm can create respectable diagnostics at very low cost, the diagnostic itself cannot remain a competitive differentiator for long, making it a commodity offering.

What is the better opportunity for AI in consulting according to this article?

Rather than using AI to produce faster explanations, the real opportunity is using AI to discover where current explanations stop working and identify the limits of incomplete models.

What happens when software produces an increasingly convincing explanation without sufficient understanding?

It creates the risk of making bigger mistakes by providing confident-sounding recommendations based on incomplete data, causing clients to act on flawed strategic conclusions.