The Asset First, Then the Demo

Twenty hand-picked PDFs. That's all the demo needed. In the board meeting, an AI assistant answers questions about maintenance manuals, cleanly, quickly, with citations. Applause, budget for the rollout. Six months later there are 40,000 documents, many manuals exist in three versions, the bills of materials live in Excel, the service history in the heads of two technicians close to retirement. The answers go wrong, the trust is gone. The project isn't cancelled. It just falls asleep.

A different company, the same idea. For the first few months nothing happens that you could show anyone. The team links fault codes to spare parts, cleans up manuals, writes down which number has to move for the whole thing to pay off, and who pays for it. The first demo comes late and looks unspectacular. A year later the system is running in operations.

Both had the same model. The first project didn't fail because of the AI. It failed because of what should have happened before the AI.

The misconception: it's the model

When an AI project gets stuck, the first reflex is almost always the same: the model isn't good enough yet. So you wait for the next generation, switch vendors, start the next pilot. And the next pilot looks great in the demo again.

That leads nowhere. The research on this is fairly clear. In 2026 Gartner surveyed infrastructure and operations leaders and concluded that ROI from AI is not driven by the sophistication of the model. The researchers behind MIT's NANDA initiative reached the same conclusion in 2025: the divide between successful and failed projects isn't driven by model quality, but by approach. And BCG boils it down to a rule of thumb: ten percent of the effort belongs to algorithms, twenty to technology, seventy to people and processes.

The model is the part anyone can buy today. That's exactly why it so rarely decides the outcome.

The numbers: a gap between pilot and P&L

At the end of 2024, S&P Global Market Intelligence surveyed more than a thousand IT and business leaders in North America and Europe. On average, a company scraps 46 percent of its AI proofs of concept before they reach production. The share of companies abandoning most of their AI initiatives before production rose from 17 to 42 percent within a year.

Gartner paints a similar picture in its own survey: only 48 percent of AI projects make it into production at all, and getting there from the prototype takes eight months on average. For generative AI, Gartner had already predicted in 2024 that at least 30 percent of projects would be abandoned after proof of concept by the end of 2025.

And even where AI runs, it rarely shows up in the income statement. In McKinsey's 2026 survey, 37 percent of respondents say AI contributes to their organisation's EBIT at all. That figure has barely moved from the year before.

The first cause: there is no business case

Ask what's holding things up, and one point almost always comes first. In Gartner's survey, the biggest obstacle isn't technology, talent or data, but the difficulty of estimating and demonstrating the value of an AI project. 49 percent name exactly that. The RAND Corporation interviewed 65 experienced data scientists and engineers. The most common reason for failed projects: the people involved misunderstood, or never clearly stated, what problem the AI was supposed to solve.

In 2026 BCG asked CEOs of large companies whether they had defined the P&L impact of their AI initiatives. 14 percent can say so for all of them. Gartner itself warns against organisations that "chase flashy demos".

That's the core of the problem. A demo proves that a model can do a task. It doesn't prove that anyone will pay for it. Build the pilot before the business case and you get a nice show and no answer to the only question that counts in the steering committee at the end: what does this get us?

The second cause: there is no asset

The second cause sits one level deeper, in the data. According to Gartner, 63 percent of organisations either don't have the right data management practices for AI or aren't sure whether they do. Gartner expects 60 percent of AI projects without AI-ready data to be abandoned through 2026. In BCG's 2025 study, 68 percent name lack of access to high-quality data as a key hurdle. 79 percent say they lack the expertise to handle unstructured data.

That last number deserves a second look. Unstructured data means exactly the manuals, bills of materials, maintenance reports, wiring diagrams and service tickets from the demo. In industry and the Mittelstand, that's where the real knowledge sits. And that's exactly where it's useless to an AI at first.

An AI that's supposed to answer from this material first needs a queryable model of it: which document is valid, which part belongs to which fault, which version is current. That isn't an IT side project. It's the step everything else depends on.

The concession: AI is delivering

To be fair, the story of the eternally failing AI project isn't true any more. In September 2026 BCG reported that almost half of companies now generate real value with AI. A year earlier, by BCG's own count, it was five percent. In a Wharton study, roughly three in four leaders at large US companies report positive ROI. Those are self-reports, but the trend is unmistakable.

That doesn't refute the thesis. It sharpens it. Because the companies that deliver differ in exactly these two places. At McKinsey, almost three quarters of high performers have fundamentally redesigned their workflows because of AI, against a quarter of everyone else. BCG describes the winners as companies that rebuild workflows end to end and measure the effect directly in their P&L.

It isn't AI that fails. It's the projects that were built as demos.

The insight: the asset first, then the demo

The order is simple, and it still gets reversed almost every time. First the business case: which number has to move, for whom, and who pays. Then the asset: the data and knowledge the AI needs to move that number, prepared and reliable. And only then the model and the demo.

Without a business case you get a show with no buyer. Without the asset you get a pilot that shines in the lab and falls apart in operations. Both together are the precondition. The model is just the tool.

The unglamorous step used to be expensive because it took specialists and months. With AI it's getting much cheaper itself. It stays unglamorous all the same. And it stays the precondition for everything that follows.

A guardrail before anyone takes a wrong turn

"The asset first" doesn't mean: two years of data strategy first. Anyone who wants to clean up all of the company's data before building a single application has just started the next big project with no result.

The asset grows along a concrete business case. The data this one case needs gets cleaned up. Then the next case. That way the asset grows with every project, and every project pays its way as it goes.

What changes about the proof of concept

The classic PoC answers the question: can the model do it? Today the answer is almost always yes. That's exactly why it's worthless.

A PoC worth having answers different questions. Does the asset hold, meaning are the quality and structure of the data good enough in live operations and not just on twenty hand-picked documents? Does the number from the business case move? And is there a clear criterion for when to stop? A proof of concept without a kill criterion isn't a test. It's the start of a project nobody wants to stop any more.

The zoom-out: the knowledge is in the binders

European industry and the Mittelstand have an advantage they rarely see as one: decades of accumulated knowledge about their products, plants and customers. It sits in manuals, drawings, service histories and the heads of experienced people. Unstructured, scattered, hard to get at. In that state it's worth almost nothing to an AI.

Whoever structures it owns the asset every AI application in their field depends on. Any competitor can buy the model. Not the knowledge about your own installed base. Why exactly this asset suddenly makes business models pencil that were nonsense yesterday is the subject of What Didn't Pencil Yesterday Pencils Today. What it looks like in service, specifically, is in The Cost Centre That Out-Earns Your Core Business.

The principle: put the business case before the model

What would an approach look like that actually bridges the gap between pilot and P&L? It would have four properties.

First: the business case comes first. Which number should move, for whom, and who pays for it? Without that answer, there's no pilot.

Second: the asset is part of the project, not its precondition. Data and documentation are prepared for exactly this case, and the result stays in the company as a reusable asset.

Third: the PoC tests the business, not the model. Real data instead of hand-picked data, a target number instead of enthusiasm, a kill criterion instead of hope.

Fourth: build the workflow, not the tool. AI works where the workflow around it gets rebuilt, not where it's placed next to an old process as an assistant.

Coming full circle

The twenty-PDF demo wasn't bad. It proved what was obvious anyway: that the model can answer questions. The other project had nothing to show for a long time and in the end delivered the only thing that counts: an operation that pays off.

The demo proves the model works. Whether the business works is decided before that — by the asset and the business case.

Sources

• S&P Global Market Intelligence (451 Research), "Generative AI Shows Rapid Growth but Yields Mixed Results" (2025; survey Oct–Nov 2024, n = 1,006): 46% of PoCs scrapped before production, 42% of companies abandon most AI initiatives (prior year 17%) — spglobal.com
• Gartner, "Gartner Survey Finds Generative AI Is Now the Most Frequently Deployed AI Solution in Organizations" (7 May 2024): 48% of AI projects reach production, eight months from prototype; demonstrating value is the top barrier (49%) — gartner.com
• Gartner, "Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025" (29 Jul 2024): prediction — gartner.com
• Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk" (26 Feb 2025): 63% lack, or are unsure of, AI-ready data management; prediction of 60% abandonment through 2026 — gartner.com
• Gartner, "Gartner Says AI Projects in I&O Stall Ahead of Meaningful ROI Returns" (7 Apr 2026): ROI not driven by the sophistication of the model — gartner.com
• Gartner, "Why Half of GenAI Projects Fail: Avoid These 5 Common Mistakes" (2026): warning against "flashy demos" — gartner.com
• RAND Corporation (Ryseff, De Bruhl, Newberry), "The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed" (2024): misunderstood problem as the most common cause; lack of data and data infrastructure — rand.org
• Boston Consulting Group, "The Widening AI Value Gap: Build for the Future 2025" (2025, n = 1,250): 68% lack access to high-quality data, 79% lack expertise for unstructured data, 10-20-70 rule — bcg.com
• Boston Consulting Group, "Nearly Nine in Ten CEOs See Some Cost or Revenue Benefits from AI in Targeted Areas, But Most Are Struggling to Scale It" (22 Jul 2026, n = 152): 14% have defined the P&L impact of all AI initiatives — bcg.com
• Boston Consulting Group, "AI Is Starting to Pay Off. Almost 50% of Companies Now Generate Value with It." (30 Sep 2026, n = 1,330) — bcg.com
• McKinsey & Company, "The State of AI in 2026: On the Road to ROI" (2026, n = 1,719): 37% report EBIT contribution; almost three quarters of high performers fundamentally redesigned workflows — mckinsey.com
• MIT NANDA, "The GenAI Divide: State of AI in Business 2025" (2025, preliminary): divide "not driven by model quality … but by approach" — report PDF
• Wharton Human-AI Research / GBK Collective, "Accountable Acceleration: Gen AI Fast-Tracks Into the Enterprise" (2025): roughly three in four report positive ROI (self-reported) — wharton.upenn.edu