Every failed transformation burns internal credibility. AI is coming to already tired organizations. Without a real diagnosis before launching, the promise wears out faster than the model produces.
A few years ago, a large European group launched an ambitious SAFe transformation: mapping of value flows, identification of ARTs, team training, PI Planning. Six months later, the trains were running. Eighteen months later, during a PI Planning, 40% of the features were blocked by dependencies on three cross-functional teams outside the train. SAFe provides mechanisms for this: Shared Services, expansion of the scope of the ART. None were activated: no one in the room had the mandate to attach teams from other departments to the train.
Everyone knew it. The teams, the local managers, the architects. The program continued.
This case has been happening for twenty years. ERP in the 2000s, cloud in the 2010s, agile at scale from 2015. Each wave produces results below promises. The usual diagnosis cites lack of sponsorship, resistance to change, lack of implementation. Convenient readings that suggest the problem could have been avoided with more rigor. The real results of these waves point to a cost of another nature, rarely measured because it is absent from business cases: the cost in internal credibility.
The debt of credibility
Employees who have gone through two or three transformation cycles have learned to wait for it to pass. They adopt the vocabulary of the moment, participate in the ceremonies, fill the artifacts, and continue to function as before. This facade conformity is a rational adaptation: the environment has demonstrated, over time, that the announced transformations did not keep their promises.
A leader who launches a new program in this context is not starting from scratch. It starts from below zero. It must repay the debt of credibility accumulated by previous programs before building anything. This cost does not appear in any business case. Yet it is one of the most predictive factors of what will happen next, and the most systematically ignored.
AI compresses the disappointment cycle
Generative AI enters this already tired terrain with an aggravating factor. We are talking here about formal AI programs, those which go to COMEX with a budget and a promise of results, not about widespread adoption via ChatGPT or shadow IT POCs.
An ERP was deployed over several years. A cloud migration over eighteen to thirty-six months. AI compresses these delays. A COMEX that has seen a demo expects results in weeks. Disappointment happens more quickly, in front of a wider audience: AI affects customer relations, documentary production, decision support, subjects that business management measures directly.
Deploying an LLM requires accessible, governed data of sufficient quality. This prerequisite is rarely met. But the main problem is elsewhere: when an AI program asks teams to structure data, validate outputs, review their processes, it is aimed at employees who have already seen agile programs, cloud migrations, product reorganizations. The power of the model does not compensate for the accumulated fatigue.
Why the diagnosis does not take place
Why don’t organizations assess their actual state before launching? The answer lies in the design of the decision-making process itself.
The room where the adoption of a transformation program is decided typically brings together four actors whose incentives converge towards the same result. The publisher or integrator, whose economic model is based on adoption. The internal champion, who needs a visible program to obtain budget and legitimacy. The COMEX, which buys a promise of outcomes and has neither the means nor the interest to engage in a technical debt audit. The IT department, which knows the real state of the IS but knows that frontal opposition positions it as a brake.
Each actor benefits individually: a signed contract, a budgeted program, a narrative that can be presented to the board, an IT department that is not perceived as an obstacle. The collective result, a launch without a diagnosis, is no one’s decision. No one in the room has the mandate to say no.
Two dimensions to cross and a framework to do it
In a previous article published here, I explored how organizations locally optimize each area without diagnosing the constraint that limits the system as a whole. The problem of unverified prerequisites comes from the same logic: we evaluate each component separately without crossing the two dimensions which determine what a program can actually produce.
The first is technical: state of the IS, level of application coupling, data quality and governance, realistic decoupling capacity within three to five years.
The second is governance: effective level of delegation on priority decisions, ability to arbitrate between areas when value flows contradict each other, real tolerance to prioritization conflict, level of residual credibility of transformation programs with teams (rarely measured). SAFe addresses this dimension through several mechanisms: Lean-Agile Leadership, Lean Portfolio Management, Participatory Budgeting. The Team Topologies model assumes a self-service, consumable platform that cannot exist without clear investment trade-offs. AI requires data governance that affects the prerogatives of business management. Each approach documents the importance of this dimension. None, in practice, makes it a go/no-go criterion.
A program can progress with a partially coupled IS if the governance allows dependencies to be processed in real time. A well-architected IS produces nothing if priority decisions remain centralized by silo. The two dimensions interact. To evaluate them separately is to not evaluate them.
What is actionable
The following guidelines are aimed at organizations that have already gone through one or more cycles. Those where the debt of credibility exists and where at least one decision-maker has measured the cost.
Change the diagnostic deliverable: Value Stream Mapping already identifies dependencies, handoffs, bottlenecks. The problem lies in the use made of them: in the SAFe process, these observations serve as inputs to the design of the ARTs, not as a map of the constraints to be treated before cutting. Using the VSM as a diagnostic deliverable rather than as a step toward slicing would change the nature of the program. The same logic applies to AI: cross-statement of data quality and governance, before the catalog of use cases.
Minimum condition: a sponsor ready to receive a less attractive deliverable than the one he had planned to present in committee.
Make the foundations visible from a budgetary perspective: in almost all the programs observed, the prerequisites are supposed to be built immediately. They are never built, systematically arbitrated for the benefit of delivery. Without a distinct line, without clear ownership, without an independent roadmap, foundations do not survive the first prioritization review. What does not appear in the budget disappears from arbitration.
Minimum condition: a sponsor who agrees to protect a line without immediate business deliverable.
Introduce an actor whose mandate does not depend on the launch: if the four actors structurally converge towards “we launch”, the only way to produce a reliable diagnosis is to mandate someone whose interest is not aligned with adoption. An evaluator before the decision to go, with a simple deliverable: here is where you are on the two dimensions, here is what this positioning allows you to expect, here is what it cannot promise.
What is happening now
The next AI wave will produce its results in organizations that still bear the traces of previous waves. The model will be powerful. The use cases will be real. And the teams who will be asked to commit will have already learned, program after program, that transformation promises have a limited lifespan.
The question that will determine what happens next is not technical. It can be summed up in one sentence: does this organization still have the credit necessary for people to believe in it once again?
