Most AI pilots don't die because the AI fails. They die because nobody defined success before building, the person with buying power wasn't in the room, and the team never bought in. A great demo can't fix any of that. Defining success, ownership and adoption up front can.
How often do AI pilots actually stall?
More often than most people expect. I mean both kinds of pilot here: the evaluation that stalls after a demo, and the build that goes live but never gets used.
Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025.1 A year later, it predicted that more than 40% of agentic AI projects will be canceled by the end of 2027.2 And McKinsey's 2026 State of AI survey found that only 37% of respondents attribute any EBIT impact to AI, about the same as the year before.3
Look at the reasons Gartner gives: poor data quality, inadequate risk controls, escalating costs and unclear business value.1 Almost none of that is about the model.
The AI usually isn't what fails.
What happens after a great demo?
Usually, nothing. And that's the problem.
The demo goes well. Everyone's excited. Then the decision takes too long. There's an AI committee now. There's a security review. Forrester found that the average B2B buying decision now involves 13 people.4
None of that is anyone's fault. Regulated businesses like banks and credit unions are careful for good reasons. But momentum is fragile, and a demo creates excitement without creating a plan.
What does "let me take it back to my team" really mean?
Sometimes it's completely real. They need to talk to their partners or their board.
And a lot of the time, it means something harder: the person who saw the demo has to explain it to people who didn't. Even with a notetaker, it's hard to explain something technical secondhand. The idea gets weaker every time it's retold.
So we try to make sure the people we talk to actually understand what we're proposing, in plain language, well enough to explain it themselves.
Why did we stop leading with demos?
Because a demo shows what AI can do in theory. Businesses buy what it will do for them in practice.
Now we lead with the work, which is what an AI-native engineering firm like Soren is built to do. For the first couple of weeks we're learning: one-on-one interviews with the people doing the work, how it actually runs, where the bottlenecks are. Then we deploy. It's the same reason we embed engineers instead of selling software.
That builds trust in a way a demo can't. By the time the first workflow is live, the client has seen us understand their business, and they know what success is supposed to look like. More on how we work.
What separates a pilot that dies from one that makes it?
Honestly, a lot of it is who you're talking to.
The pilots that died for us usually came down to timing, or to talking with someone who was enthusiastic and doing great work but didn't have the power to buy. The ones that made it had a buyer with real authority from the start, a specific problem they wanted solved, and enough time with us to trust that we understood it.
| Pilots that stall | Pilots that make it | |
|---|---|---|
| Who's in the room | A champion without buying power | A buyer with authority |
| Success | "Let's see what it can do" | Defined before building |
| The team | Hears about it secondhand | Involved from the first interviews |
| Starting point | The demo | One real workflow |
| Adoption | Assumed | Planned, with training |
Research points the same way. McKinsey's March 2025 State of AI survey found that CEO oversight of AI governance is one of the elements most correlated with higher bottom-line impact from generative AI, especially at larger companies.5
Who has to be in the room?
Two people, at minimum: whoever is buying, and whoever is doing the work day to day.
The buyer can make the decision. The person doing the work knows whether the thing will actually be used. If either one is missing, the pilot is running on hope.
How should you define success before building anything?
Start with buy-in. How many people at the firm actually want this?
It's really hard to get a pilot off the ground if people have already turned it off in their heads. Few teams measure buy-in, and many don't measure results either. The Thomson Reuters Institute found in 2025 that only 20% of professionals knew their organization was measuring the ROI of generative AI.6 In McKinsey's survey, fewer than one in five respondents said their organizations track KPIs for generative AI.5
So we define success as two things: the people who'll use it are using it, and there's a measurable result tied to one workflow. Hours saved, turnaround time, errors caught.
I've seen plenty of teams start out skeptical. What changed it was testing with them, then teaching and training as we went. BCG found in 2025 that only 36% of employees feel adequately trained on AI.7 That gap is fixable, but someone has to own it.
The AI is rarely the hard part. Getting people to trust it and use it is.
Define success before you build
Start with a 30-minute call with our team. We'll help you define what success looks like before anything gets built.
Frequently asked questions
Why do most AI pilots fail?
Most AI pilots fail for organizational reasons, not technical ones. Common causes are unclear success criteria, no executive owner, slow decisions across committees, and teams that don't adopt the tool. Gartner, for example, points to unclear business value, escalating costs, poor data quality and weak risk controls.
How long should an AI pilot take?
A focused pilot should show a measurable result within weeks, not quarters. In our work, the first workflow is typically live within two to three weeks of getting access to the needed systems. Long pilots without a clear success measure tend to lose sponsors and momentum before they reach production.
Who should be involved in an AI pilot?
At minimum, the person with authority to buy and the people who do the work day to day. The buyer can approve the rollout, and the team knows whether the tool will actually be used. Pilots missing either group often stall after the demo, even when the technology works.
How do you measure AI pilot success?
Pick one workflow and one or two measures before building, like hours saved, turnaround time or errors caught. Then track adoption: are the people it was built for actually using it every week? A pilot that works technically but isn't used hasn't succeeded yet.
Sources
- Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025, Gartner, July 2024
- Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner, June 2025
- The state of AI in 2026, McKinsey & Company, August 2026
- The State of Business Buying, 2024, Forrester, December 2024
- The state of AI: How organizations are rewiring to capture value, McKinsey & Company, March 2025
- 2025 Generative AI in Professional Services Report, Thomson Reuters Institute, April 2025
- AI at Work 2025: Momentum Builds, But Gaps Remain, BCG, June 2025
