PwC Asked 767 Companies Why Their AI Demos Failed. The 87% Answer.
PwC asked 767 operations and supply chain leaders whether their AI investments were delivering. 89% said no, and 87% named the same culprit: their own data. This special works through why the demo holds and the deployment doesn't, what distribution shift does to a vendor's numbers, and how to test a proposal before the money is committed rather than after.
The Promise
- AI in production works. The companies running it well are getting measurable gains, and the failure point is procurement, which is a process you control.
- The 30-day vendor test is cheap: run the demo against 30 consecutive days of your own uncleaned production data, measure the gap, and tie the contract to it.
- The gap between demo accuracy and accuracy on your data is a real risk number. If the test lands at 62%, you know you need 33 points of human supervision, before you sign.
The Risk
- PwC's 2026 Digital Trends in Operations survey: 89% of 767 operations leaders say their technology investments haven't delivered, 87% blame data quality, and only 4% report success across all four metrics PwC measured.
- Deloitte put the abandonment rate at 42% of companies, up from 17% the prior year, with an average sunk cost of $4.2 million per abandoned project.
- Distribution shift is how machine learning works. A model tuned on the vendor's curated slice can lose 30 to 60 points on data it has never seen, and the demo is built so you never see that.
- An ROI model built on demo accuracy is a vendor's marketing number in a finance spreadsheet, and the gap lands on your operating review before the deployment does.
767 companies, one answer
PwC’s 2026 Digital Trends in Operations survey asked 767 operations and supply chain leaders at US enterprises one variation on one question: is the AI delivering?
89% said their technology investments haven’t delivered the expected results. 87% named the reason: poor data quality. Their own data was messy, contradictory and full of edge cases the vendor’s demo never saw, and it broke the deployment. Only 27% have fully embedded an AI strategy across business units. 4% report success across all four metrics PwC measured.
PwC is not an outlier. MIT Sloan’s 2025 research found 95% of generative AI pilots never reach production. RAND tracked 2,400 enterprise AI projects and reported 80% failed to deliver their intended value. McKinsey’s 2026 global AI survey put 73% of deployments short of projected ROI. Different samples, different methods, same pattern. The vendor demo predicts the procurement decision. The procurement decision does not predict the production outcome.
After 25 years in cybersecurity, I’ve watched this script run before: endpoint security in the early 2000s, identity in 2010, cloud security in 2015. The demo is always spectacular. The production deployment on your data is always a different number.
Distribution shift
There’s a name for what happens underneath: distribution shift. A demo runs on curated data. The vendor picks the dataset, cleans it and knows which questions the model handles well. Production runs on your transaction records, your customer notes and your 16 years of edge cases. When the production distribution doesn’t match the one the model learned from, performance collapses, often by 30 points and sometimes by 60.
That isn’t a bug, and it isn’t a vendor’s lie. It’s a vendor’s silence.
Who does what
If you’re an executive: was the ROI model built on demo accuracy or production accuracy? Almost certainly the first.
If you’re an IT leader: stop running pilots on the vendor’s data. Run them on your messiest production slice, and get the original deployment terms in writing before the accuracy drops and the vendor blames the data.
If you’re a board director: is there evidence the demo translated to production? Ask for it in the minutes.
The 30-day vendor test
One. Make the vendor run their demo against 30 consecutive days of your production inputs. No data prep, no curation. If they refuse, that’s the answer.
Two. Measure the gap between demo accuracy and accuracy on your data. That gap is your real risk number.
Three. Tie the contract to the gap. If the demo promised 95% and production lands at 62%, you pay for 62%, with a milestone path back and a penalty clause if it doesn’t close.
Demo accuracy minus your data accuracy. Get both numbers on the table before you sign.