Key Takeaways
- MIT's GenAI Divide study, built on 150 executive interviews, 350 employee surveys, and 300 analysed deployments, found that 95% of enterprise generative AI pilots delivered no measurable P&L impact across an estimated $30 to 40 billion of investment. The lead author traced the failure to enterprise integration, with model quality rarely the constraint.
- Gartner predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data. In its survey of 248 data management leaders, 63% said they either lack, or are unsure they have, the data management practices AI requires.
- McKinsey's State of AI research shows adoption is no longer the question: 88% of organisations use AI in at least one function, yet only around a third have begun to scale it, and just 7% report AI fully scaled across the enterprise.
- AI is a multiplier. It amplifies whatever the data already contains. On a governed ecosystem it amplifies analytical advantage. On a fragmented one it amplifies errors, faster than humans can catch them.
- Architecture sets the ceiling on intelligence. AI confined to one system answers narrow questions. AI running on unified commercial and operational data answers a different category of question entirely, and that gap compounds over time.
The pilot that impressed the board and then disappeared
Ask a senior executive at any mid-to-large enterprise whether the organisation is investing in AI, and the answer is yes. The pilot has run. The proof of concept landed well in a boardroom presentation. The vendor roadshows have been attended. The AI strategy document exists.
Twelve months later, the pilot has not become a production system. The business outcomes remain theoretical. The board is asking, again, why the spend has not turned into impact.
The narrative that usually accompanies this moment blames the technology. The models are immature. The use cases were too ambitious. The organisation moved too fast, or its industry is uniquely complex. This narrative is comfortable, widely repeated, and wrong, and it is damaging because it points investment away from the actual cause of failure.
The models that data scientists build in a proof of concept are genuinely capable. The demos land. The accuracy metrics are real. The failure happens when those models meet the organisation’s actual data ecosystem: production data that is fragmented, inconsistently labelled, and governed by no one; workflows that were never redesigned to act on the model’s output; and no infrastructure to monitor the model once it is live. The bottleneck is never the model. It is always the foundation underneath it.
What the research actually shows
The failure rate of enterprise AI stopped being anecdotal in 2025, when three independent research efforts converged on the same conclusion from three different directions.
The MIT NANDA initiative’s GenAI Divide: State of AI in Business 2025, covered in depth by Fortune, found that 95% of enterprise generative AI pilots delivered no measurable P&L impact, across $30 to 40 billion of estimated investment. Lead author Aditya Challapally was specific about the cause: executives tend to blame regulation or model performance, but the research points to flawed enterprise integration, a learning gap in which tools never adapt to actual workflows and organisations never adapt their workflows to the tools. Two of the study’s secondary findings deserve more attention than they get. Budgets concentrate in sales and marketing pilots, where measured returns are lowest. And organisations that bought from specialised vendors and built partnerships reached deployment around 67% of the time, while purely internal builds succeeded roughly a third as often, a finding that says experienced delivery patterns beat in-house enthusiasm.
Gartner reached the same destination from the data side. Its February 2025 analysis predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data. The supporting survey of 248 data management leaders found that 63% either lack, or are unsure they have, the data management practices AI requires. Gartner’s warning is structural: organisations that treat AI data requirements as an extension of traditional, report-oriented data management endanger their entire AI effort.
McKinsey’s State of AI research completes the picture from the adoption side. Based on 1,993 respondents across 105 countries, it found that 88% of organisations now use AI in at least one business function, up from 78% a year earlier. Yet most remain in experimentation or piloting, only around a third have begun to scale, and just 7% report AI fully scaled across the enterprise. Usage is everywhere. Value at scale is rare, and McKinsey’s own conclusion points to workflow redesign as the missing ingredient.
Three research houses, three methodologies, one finding. The constraint sits beneath the model, in the data and the operating model around it.
Three ways enterprise AI dies between pilot and production
Across manufacturing, pharmaceutical, and supply chain enterprises, the same three failure modes recur, and naming them precisely matters because each one demands a different fix.
Data quality and governance failure. The pilot model was trained on a hand-curated dataset that does not resemble the messiness of production. Deployed against real data, accuracy degrades, and there is no systematic process for resolving the quality issues because governance was never established as part of the programme. In regulated industries such as pharmaceuticals, this goes beyond accuracy: ungoverned AI inputs and outputs are a compliance exposure that can halt a programme entirely.
Operationalisation failure. The model produces an output and the organisation has no mechanism for acting on it. A demand forecasting model whose recommendation a planner must manually review, re-enter, and override in a separate system has automated nothing. It has added a step to an already complex process, and the AI sits technically live and practically unused. This is MIT’s learning gap made concrete: the tool never entered the workflow where the decision actually happens.
MLOps absence. The model performs at deployment and degrades as the underlying data distribution shifts. With no automated monitoring, no drift detection, and no retraining pipeline, the degradation goes undetected until the business notices the recommendations have become unreliable. By then trust in the system is damaged, and rebuilding trust costs far more than building the monitoring infrastructure would have.
Notice what the three share. Every one of them is decided before the first line of model code is written, by the state of the data ecosystem and the design of the operating model around it.
Impressive AI pilot, but no production date? Book a 30-minute review to pinpoint which failure mode is responsible.
The ceiling nobody prices in: fragmented ecosystems and partial intelligence
There is a second, quieter way the foundation limits AI, and it operates even in programmes that avoid outright failure. It determines what the AI can see.
Ask any Chief Data Officer how the current data ecosystem came to look the way it does, and the answer is almost always the same: it wasn’t planned. It evolved. A warehouse built for financial reporting. A lake added during cloud adoption. A separate analytics environment stood up for one supply chain project. An integration layer connecting CRM and ERP, built by a different team, to a different standard, on a different budget. Fifteen years later, the ecosystem is something nobody designed and nobody can fully map, with data that exists in three systems and agrees in none of them.
For AI, this fragmentation has a structural consequence beyond bad data quality: each model only knows its own silo.
AI applied to CRM data alone can tell an organisation about customer behaviour. Likely next purchase, churn risk, next best action in a sales conversation. Genuinely useful, and narrow. AI applied to unified commercial, operational, and financial data answers a different category of question. Which customer commitments are at risk because of a supplier disruption that happened this morning? What pricing decision protects margin without breaking a delivery promise? Where in the product portfolio is margin being eroded by supply cost trends the pricing team cannot see?
No amount of model sophistication produces that second category from a single system’s partial view. The sales system holds what was promised. The ERP holds what operations can actually deliver, at what cost, by which date. In most organisations both are well maintained, both are trusted within their own function, and both are thoroughly disconnected from each other. The gap between commercial intent and operational reality is one of the most expensive inefficiencies in enterprise operations, and it is entirely architectural in origin.
The architecture sets the ceiling. An organisation whose AI strategy runs on CRM data alone has a low one. Connecting commerce to operations removes it.
What the 5% build first
The organisations that move AI from pilot to production, and from production to advantage, share a consistent pattern, and it inverts the sequence most programmes follow.
They build the data foundation before any model. The data the model will depend on is audited, governed, and established as a production asset first. InspireXT delivers this on the Databricks Lakehouse with the Medallion architecture: Bronze retaining raw, auditable source data; Silver applying transformation and business logic; Gold delivering governed, business-ready data products, with lineage end to end. In manufacturing environments, the Connected Shopfloor AI capability integrates PLC, SCADA, IoT sensor, and ERP data into this same foundation, which means the data that trains the model is the same data structure that operates in production. Pilot accuracy becomes production accuracy because nothing changes between the two.
They design for the decision, before the model exists. The opening question is never whether a model can make a prediction. It is how the model’s output will change a specific business decision, and who owns that workflow redesign. The bi-directional Salesforce, Oracle ERP, and Databricks integration that InspireXT’s Connected Intelligence practice builds makes the answer concrete. A sales manager looking at an account in Salesforce sees current inventory position, supplier lead times, and cost of goods at the current margin, without leaving the CRM, so the commitment made is informed by operational reality. A supply chain planner building the monthly production plan sees the CRM pipeline for the next quarter, so the plan reflects commercial intent instead of historical run rates alone. A CFO reviewing monthly accounts sees margin by product, customer, and channel, calculated from the same governed source the commercial and operations teams use, so the meeting starts from one number.
They treat MLOps as a prerequisite, never a retrofit. Automated performance monitoring, drift detection, retraining pipelines, and version control for both models and data are in place at deployment, with full audit trails for regulated environments. Predictive maintenance models, Golden Batch modelling for process industries, and OEE intelligence applications go live with this infrastructure from day one, which is why they are still performing twelve months later when the operational environment has moved.
This sequence is unglamorous in exactly the way MIT’s and Gartner’s research predicts the successful 5% would be. The foundation work makes no boardroom demo. It is also the entire difference between AI that works in a presentation and AI that runs the business.
If your AI roadmap ignores the data foundation, book a 30-minute review to map the gap between pilot and reality.
What this means for AI sponsors in 2026 and beyond
Gartner’s 60% abandonment prediction is a forecast about foundations, and it splits the market in two. The organisations funding data readiness now will own the projects that survive, and each surviving project compounds: a governed foundation built for the first use case accelerates the second and the third, because the data products, the governance, and the MLOps infrastructure are already in place. The organisations bolting copilots onto fragmented ecosystems will keep generating wrong answers faster, creating compliance exposure, and burning the internal trust that every future AI initiative will need.
The competitive gap this opens is architectural, which makes it durable. A rival can match your model in a quarter; the frontier models are available to everyone. Matching a unified, governed data ecosystem that connects commerce to operations takes years, and the organisation that has one is building AI on a richer foundation, optimising decisions its competitors still make manually, and compounding the advantage with every new use case it ships.
The six challenges explored across the Data That Decides series, the investment gap, fragmented architecture, AI readiness, data debt, visibility gaps, and commercial intelligence silos, are not separate problems. They are symptoms of one condition: organisations that invested in data systems without investing in data coherence. The discipline that resolves them is consistent. Strategy before platforms. Outcomes before tools. Governance before scale.
How InspireXT approaches this problem
InspireXT works backwards from the decision. The Data & AI Blueprint identifies which business decisions the AI programme exists to improve, sequences use cases by business value and data readiness, and defines governance before any platform commitment is finalised. Platform Modernisation then turns the fragmented ecosystem into a governed Lakehouse foundation, with the XT Value Chain Lakehouse accelerator providing pre-built, deployment-validated data models for Supply Chain, Commerce, Finance, and Operations that compress delivery from years to weeks.
On that foundation, Connected Shopfloor AI and Managed MLOps take models into production and keep them there, with monitoring, retraining, and audit built in from the first deployment. Connected Intelligence closes the commercial loop, integrating Salesforce, Oracle ERP, and Databricks so that the intelligence reaching each decision-maker reflects the full operational and commercial picture. The outcome is the inversion of the industry’s failure statistics: AI that reaches production, stays in production, and continues to perform as the environment around it changes.
Key Questions Leaders Are Asking
Why do most enterprise AI pilots fail to reach production?
The model is rarely the cause. MIT’s 2025 research traced the 95% failure rate to enterprise integration, and Gartner’s to the absence of AI-ready data. In practice four conditions kill the pilot: it was trained on manually cleaned data while production data is fragmented and ungoverned, the workflow was never redesigned to act on the output, no one owns the model after deployment, and no MLOps infrastructure exists to detect degradation. All four are set before the model is built.
How do I know if my organisation is ready for the AI programme we are funding?
Ask one question of your data team: define the top five KPIs, show where each comes from, and demonstrate that every function agrees on the definition. If that takes more than an afternoon or surfaces disagreement, the estate is fragmented, and AI deployed on it will amplify the fragmentation. Gartner’s survey found 63% of organisations in exactly this position, so discovering it early puts you ahead of most.
Can a modern planning system coexist with a legacy ERP without creating integration risk?
Ask one question of your data team: define the top five KPIs, show where each comes from, and demonstrate that every function agrees on the definition. If that takes more than an afternoon or surfaces disagreement, the estate is fragmented, and AI deployed on it will amplify the fragmentation. Gartner’s survey found 63% of organisations in exactly this position, so discovering it early puts you ahead of most.
What does AI-ready data actually mean?
Data aligned to specific use cases, governed at the asset level with named owners, supported by automated pipelines with quality gates, and continuously monitored rather than audited quarterly. The Medallion architecture operationalises this: raw data preserved and auditable in Bronze, business logic applied in Silver, governed data products in Gold, with lineage running end to end so every AI output traces back to its source.
Can we run AI on our CRM data while the wider ecosystem gets fixed?
You can, and it will produce narrow value: churn prediction, next best action. The strategic cost is the ceiling. Intelligence that spans pricing, delivery risk, and margin requires commercial and operational data unified in one governed platform. Organisations that stop at the CRM concede that category of intelligence to competitors who connect the two, and the gap widens with every model the connected organisation ships.
Who should own a model after it is deployed?
Someone with production accountability and MLOps infrastructure behind them. If the answer is the data science team that built it, with no monitoring or retraining pipeline, the model will degrade undetected as data distributions shift. In regulated industries ownership also covers audit trails for AI outputs, which makes it a board-level compliance question rather than a technical one.
How should we define success for an AI programme?
In business outcomes at twelve months, never in model accuracy at deployment. A forecasting model succeeds when planners act on it without manual override, inventory positions improve, and the decision it informs is measurably faster or better. Accuracy is an input. The changed decision is the outcome, and programmes that optimise for accuracy alone are optimising for the demo.
Does fixing the foundation mean replacing our existing systems?
No. Platform modernisation restructures how data flows between existing systems and imposes governance on what was previously ungoverned. Salesforce stays the system of record for commerce, the ERP for operations. The Lakehouse unifies their data into one governed foundation, and AI runs on that foundation rather than on any single silo, which is why the work takes weeks with pre-built domain models rather than the years a replacement programme would.
If your organisation has an impressive pilot, a fragmented ecosystem underneath it, and a board asking when the AI investment will show up in the numbers, InspireXT would like that conversation.