Forty-two per cent of companies now abandon the majority of their AI initiatives before they reach production, up from seventeen per cent the year before, according to S&P Global Market Intelligence's 2025 Voice of the Enterprise survey. MIT's NANDA initiative put the number higher still: ninety-five per cent of enterprise generative AI pilots deliver no measurable P&L impact. The problem is rarely the model. It is that organisations commit full build budgets before they have evidence the thing will work on their data, with their users, inside their systems. A four-week prototype is the cheapest way to buy that evidence.
A prototype is not a demo
Most teams use the word "prototype" to mean a working PowerPoint with screenshots of a fictional interface. That is a demo. A prototype is working software, deployed to real infrastructure, processing real data, that answers three questions:
- Does the model work on your data? Not the vendor demo data. Yours, with the missing fields and bad encodings. Gartner's 2024 forecast attributed the bulk of generative AI abandonment to "poor data quality, inadequate risk controls, escalating costs or unclear business value". Three of those four are discoverable in week one if you actually touch the data.
- Does anyone want to use it? Put it in front of three real users. Watch what they do. MIT's 2025 study found the biggest predictor of pilot failure was not model quality but the absence of learning loops with real users.
- Does it integrate with the systems it needs to? Production AI projects die not because the model is wrong but because nobody can get the data into it or the predictions out of it. A retail client of ours had a working recommendation model in a fortnight. It took four more months to discover their order management system could not accept the payload format. Better to find that out on day ten than day two hundred.
Why AI projects die in the middle
The distance between a clever notebook and a production system is mostly plumbing. McKinsey's 2025 State of AI report puts the average time from enterprise AI initiation to production at around seventeen months. Long-running data prep surveys consistently find data scientists spend seventy to eighty per cent of their time on data collection, cleaning and integration, not modelling.
That tells you where the risk lives. If most of the work is plumbing, most of the risk is plumbing. Yet most AI pitches still lead with the model. A four-week prototype inverts the order of attack: do the plumbing first, on a small scale, so that by the end of week one you know whether the rest of the project is viable.
One financial services client came to us with a twelve-month plan for a credit decisioning model. We proposed a four-week prototype instead. By the end of week one we had found that forty per cent of the historical decision records were missing the outcome field entirely. That single finding, which would have surfaced in month six of a traditional build, reshaped the programme. The business still invested. It invested in the right thing.
What a four-week prototype looks like
Week one: get the data. Real, anonymised, in production-shaped format. If the data cannot be extracted in a week, that is itself a finding and the programme needs a different shape.
Week two: build the smallest thing that could possibly work. One model, one screen, one workflow. Connected to a real backend. Deployed somewhere your users can reach it. The goal is not elegance, it is contact with reality.
Week three: put it in front of users. Watch them. Iterate fast. Three users is usually enough to surface the things that will kill the build version.
Week four: write the report. What worked, what did not, and the recommendation: build it for real, kill it, or pivot.
A public sector client used this sequence on a document classification project. Week one exposed that the source PDFs were a mix of scanned images and native text, which nobody had flagged in scoping. Week three showed that caseworkers did not trust model confidence scores unless paired with the exact passage the model had read. Week four produced a costed recommendation with an explicit OCR carve-out. The production system shipped on time because the surprises had already been absorbed.
What to look for at the end
If the prototype lands well, three things will be true: you will have a clear, evidence-based recommendation on whether to invest the rest of the budget; the team will know exactly what to build next, so build phase starts at speed; and your users will already have seen something that helps them, making them advocates rather than sceptics.
If it lands badly, the most expensive thing you will have learned is that you should not have spent the rest of the budget. MIT's 2025 study found organisations that buy or partner with specialist AI vendors succeed around sixty-seven per cent of the time, while pure internal builds succeed about a third as often. The prototype is the cheapest way to find out which side of that line your use case sits on.
The principle is simple. Spend a small amount of money to find out whether to spend the rest. We have run this playbook dozens of times across financial services, retail and the public sector. The projects that clear the four-week gate tend to ship. The ones that do not, should not have been built.
Sources
- S&P Global Market Intelligence, Voice of the Enterprise: AI & Machine Learning, Use Cases 2025 (abandonment rose from 17% to 42% year over year): https://www.spglobal.com/market-intelligence/en/news-insights/research/ai-experiences-rapid-adoption-but-with-mixed-outcomes-highlights-from-vote-ai-machine-learning
- CIO Dive summary of the S&P Global findings: https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/
- MIT NANDA initiative, The GenAI Divide: State of AI in Business 2025 (95% of pilots deliver no measurable P&L impact), via Fortune: https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- Gartner press release, July 2024: 30% of generative AI projects abandoned after proof of concept by end of 2025: https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025
- McKinsey, The state of AI 2025: how organisations are rewiring to capture value: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- HPCwire / BigDATAwire on data preparation surveys (data scientists spend 70-80% of time on data prep): https://www.hpcwire.com/bigdatawire/2020/07/06/data-prep-still-dominates-data-scientists-time-survey-finds/
