This is a subtitle for your new post
According to MIT’s The GenAI Divide: State of AI in Business 2025, 95% of organizations are generating no measurable return from their investments in generative AI, despite an estimated $30–40 billion having been committed.
The study’s methodology can be debated. The order of magnitude, however, is consistent with what we observe in practice. And the root cause is rarely the technology.
Failure begins long before the first model is built.
An instructive precedent
This phenomenon did not begin with AI.
Years before the current wave, we observed the rollout of a business intelligence tool within an industrial group. The objectives were never clearly defined. Roles were never assigned. Training and development were rushed.
The result was thousands of reports, each usable only by the person who had created it.
The tool itself worked perfectly. Yet the figures it produced inspired little confidence -not because they were wrong, but because poorly configured reports were often measuring something different from what users believed they were measuring.
Replace “report” with “model” and the anatomy of many current AI failures becomes immediately recognizable.
Only the budgets are larger, and the expectations higher.
A chain of failures, not an isolated incident
AI initiatives rarely fail for a single reason. Failures accumulate, and each one prepares the ground for the next.
The problem has not been defined.
The company wants to “accelerate on AI”-but which decision needs to improve? Which performance outcome is expected to change: margin, cash, inventory, lead time, quality?
A technological ambition is not a strategy.
The right question is not: Where can we use AI? It is: Which important decision are we currently making with insufficient precision, anticipation or discipline?
Use cases are selected for visibility, not value.
The MIT report documents this bias. Generative AI budgets are heavily concentrated in sales and marketing -the functions that are easiest to showcase to an executive committee -while the most tangible gains it identifies often sit in less visible areas: document automation, finance, procurement and reduced reliance on external providers.
A conversational interface is easy to demonstrate in a management meeting.
A use case focused on scrap reduction, working capital or forecast accuracy is less spectacular—and often more profitable.
This is a capital allocation problem, not a technology problem.
The organisation overestimates the readiness of its data.
A company can be rich in data and poor in usable data.
ERP systems, CRM platforms, historical records and thousands of spreadsheets may coexist with conflicting definitions of the same indicator, incomplete histories, untraceable manual corrections and no clear accountability for data quality at source.
Important data without a genuine owner remains fragile.
Data quality is not an absolute concept. It must be assessed against the decision the organization is trying to improve.
The pilot proves technical performance, not economic value.
Model accuracy says little about the value created.
Technical performance must be connected to the decisions changed, the cost of the actions triggered, false positives, and the costs of integration and maintenance.
A technically strong model may produce no return if it intervenes too late.
An imperfect model may be highly valuable if it improves an economically important decision sufficiently.
A successful demonstration is not the same as a sound capital decision.
The decision process remains unchanged.
A score appears on a dashboard. An alert is sent.
But who acts? Within what timeframe? According to which rule?
A credit-risk model does not collect a receivable. A predictive maintenance model does not repair a machine.
Value does not reside in the prediction itself. It resides in the decision and action that the prediction changes.
Without redesigning the process, AI becomes another layer of information—like the thousands of reports no one opened.
Nobody owns the outcome end to end.
The data team owns the model. The business expects a solution. IT secures the infrastructure. The vendor delivers a tool.
When everyone owns one component, no one owns the whole.
Scaling then exposes everything the pilot concealed: system diversity, recurring costs and actual user adoption.
The MIT report finds that initiatives delivered with specialized external partners reach deployment significantly more often than internally developed solutions. It also notes that this difference may reflect organizational maturity as much as the choice between building and buying.
Either way, the lesson is the same: the difficulty of scaling is consistently underestimated.
Five questions the board should ask
Before funding or expanding an AI initiative:
- Which decision or performance outcome are we trying to improve?
- What data is required, and what condition is it actually in?
- How will economic value be measured?
- Who owns the outcome end to end?
- What conditions will trigger termination, correction or deployment?
If these five questions do not have explicit answers, the project is not ready to be funded—regardless of the quality of the proposed technology.
Governing AI as an investment
AI should be neither treated as a routine IT tool nor elevated into an inevitable transformation.
It should be governed as an investment: with hypotheses, evidence, thresholds and the discipline to stop what fails to meet them.
That is precisely what a prioritisation diagnostic should determine before capital is committed: which decisions genuinely warrant AI, in what order, and on the basis of which actual data.
Everything else (the model, the platform, the vendor) comes afterwards.
Useful AI is not the AI that impresses, it is the AI that improves a decision.
Christian Cirino - Nobilys Group
Sources: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf