What is the difference between an AI pilot and an AI demo?

A pilot can state, in one sentence, the baseline number it beat and where that number came from. A demo cannot. If the owner of a project says the team feels it is faster, or that accuracy is strong, or starts describing the model, that is a demo. It proves the technology functions. It does not prove it changed anything, because there is no recorded before to compare the after against. Baselines reconstructed after the fact are arguments, not measurements. This distinction is why so many pilots that 'went well' are quietly cancelled at the next budget review.

Why can't we prove AI ROI even though the pilot seemed to work?

Almost always because no baseline was recorded before the pilot started. PwC's 2026 CEO Survey found 56% of CEOs report neither increased revenue nor decreased costs from AI over the previous twelve months, and only 12% report both. The projects behind those figures were not technically broken. They ran against processes nobody had documented, with no owner accountable for the outcome, and no measured starting point. When finance asks what changed, there is nothing to compare against, so the spend cannot be defended and the project is cut.

What should we do before funding another AI pilot?

Three things, none of which require buying software. Write down the baseline: a specific number, with a date and a note on where it came from. Name one owner who can be asked in a board meeting what the pilot returned and who has authority to change the process it touches. Document the process the pilot will automate, and check it by asking three people to complete the same task independently and comparing the results. If their approaches differ significantly, the pilot will work for one of them and fail for the others. Fix that first.

14 September 2026

What percentage of AI pilots actually reach production?

Q: What percentage of AI pilots actually reach production?

S&P Global Market Intelligence found the average organisation scrapped 46% of its proofs of concept before production, and that 42% of firms abandoned most of their AI initiatives in 2025, up from 17% the year before. Reaching production is not the same as succeeding, though. MIT's 2025 research found 95% of AI pilots deliver zero measurable P&L impact, which includes pilots that did go live. A reasonable working assumption for a mid-sized business is that roughly half of pilots never reach production, and of those that do, only a minority can be shown to have changed a financial number.

The 2026 numbers on AI pilot-to-production rates, with named sources and dates, and what separates the pilots that survive from the ones that quietly stop.

By Carl Chessum

Somewhere between 5% and 58%, depending on what you count as production and who is doing the counting. The headline number everyone quotes is MIT’s: 95% of AI pilots deliver zero measurable P&L impact. The more useful number, because it describes the decision rather than the outcome, comes from S&P Global Market Intelligence, which found the average organisation scrapped 46% of its proofs of concept before production, and that the share of firms abandoning most of their AI initiatives rose to 42% in 2025 from 17% the year before.

So: roughly half of pilots die before anyone tries to put them into production, and of the ones that do go live, most cannot be shown to have changed a number on the P&L. The reason is almost never that the model underperformed. It is that nobody wrote down what the model was supposed to beat. A pilot that finishes without a stated baseline, a named owner and a documented process has produced a demo, and demos do not survive a budget cycle.

The 2026 numbers in one place

These are the figures being cited in board papers this year. Each one measures something slightly different, which is exactly why they appear to contradict each other.

Finding	Number	Source
AI pilots delivering zero measurable P&L impact	95%	MIT NANDA, 2025
Proofs of concept scrapped before production, average organisation	46%	S&P Global Market Intelligence, 2025
Firms abandoning most of their AI initiatives	42% in 2025, up from 17% in 2024	S&P Global Market Intelligence
CEOs reporting neither increased revenue nor decreased costs from AI in the last 12 months	56%	PwC 2026 CEO Survey, via Forbes, Jan 2026
CEOs reporting both increased revenue and decreased costs	12%	PwC 2026 CEO Survey, via Forbes, Jan 2026
Executives saying their company’s AI strategy is ‘more for show’	75%	Writer / Workplace Intelligence, 2026
Executives calling AI adoption a massive disappointment	48%, up from 34% the previous year	Writer / Workplace Intelligence, 2026
Chief strategy officers who do not own the top strategic decisions in their own function	65%	Deloitte 2026 Global CSO Survey

Two things are worth saying about this table before anyone screenshots it.

First, the MIT 95% and the S&P 46% are not the same measurement. MIT is counting pilots that never moved a financial number. S&P is counting pilots that were killed before the question was even asked. A pilot can go into production and still land in MIT’s 95%, which is the bit most people miss. Reaching production is not the finish line. It is the point at which you start being able to prove whether it worked, assuming you set yourself up to.

Second, the PwC figure is the one that matters most to the people reading this. Fifty-six percent of CEOs report no revenue increase and no cost decrease over a full year of AI spending. Only 12% report both. That is the shareholder-facing version of the same story: the money went out, and the finance function cannot find where it came back.

Why the abandonment number tripled

The jump from 17% to 42% in a single year is usually read as a market correction. It is not. It is the first budget cycle in which pilots had to justify themselves against something other than novelty.

In 2024, an AI pilot could be funded on the grounds that it was an AI pilot. Board pressure was to be seen doing something. By 2025, the question changed to what it returned, and a large number of projects discovered that they had no defensible answer, because nobody had recorded where they started. The pilot ran, people said it felt faster, and there was no before to compare the after against.

That is why the abandonment figure rose rather than the failure figure. These projects did not fail in any technical sense. They were terminated because they could not be defended. The distinction matters, because the fix is administrative, not algorithmic.

What the survivors did differently

The 12% in PwC’s data who report both revenue up and costs down are not running better models. They had three things in place before the pilot started.

A stated baseline. A number, with a date, and a note on where it came from. Invoice processing took an average of 4.2 days across 1,100 invoices in Q3. Support tickets resolved on first contact ran at 61% over the last six months. It does not have to be sophisticated. It has to exist in writing before the pilot begins, because a baseline reconstructed afterwards is not a baseline, it is an argument.

A named owner with actual authority. Not a steering committee. One person who can be asked, in a board meeting, what the pilot returned, and who has the standing to change the process it touches. Deloitte’s 2026 CSO Survey found 65% of chief strategy officers do not own the top strategic decisions in their own function. If that is true of strategy officers in their home territory, ask yourself honestly who owns AI outcomes across finance, operations and IT in your business. Usually the answer is everyone and therefore no one, which is a subject worth working through properly.

A documented process underneath it. This is the one that quietly kills more pilots than the other two combined. If three people in your business complete the same task three different ways, an AI pilot will automate one of those three ways and produce inconsistent results for the other two. The pilot will look successful in the hands of the person who designed it and fall apart when it meets the rest of the department. That is not a model problem. That is a process documentation problem wearing a model’s coat, and it is one of the recurring reasons pilots never scale.

The one-sentence test before you fund the next one

Here is a test that costs nothing and takes about a minute.

Ask the pilot owner to state, in one sentence, the baseline number the pilot beat and where that number came from.

If they can say “manual reconciliation averaged 3.1 days across 480 transactions in April, measured from the finance system’s timestamps, and we brought it to 0.9 days,” you have a pilot. If they say “the team feels it’s much quicker,” or “the accuracy is really strong,” or they start talking about the model, you have a demo. Demos are not worthless. They tell you a technology functions. They just cannot be scaled, because there is nothing to scale against.

Run this on every pilot currently in flight in your business. My experience is that most organisations find that fewer than half of their active pilots pass, and that the ones that fail tend to be the ones with the most enthusiastic internal sponsorship.

The pilot-to-production gap is organisational, not technical

None of the three things the survivors had is a technology decision. Baseline, owner, documented process. You could put all three in place this quarter without buying anything.

This is awkward, because it is far easier to approve a tooling budget than to approve six weeks of writing down how the accounts payable process actually works. The tooling purchase feels like progress and appears in a board update. The process documentation feels like admin and appears nowhere. So companies buy the tool, pilot it against an undocumented process with no baseline, and then join the 46%.

The same logic applies to the data underneath. If your key systems cannot share data without someone exporting to a spreadsheet, your pilot’s path to production runs through a person doing manual exports, and that person is a single point of failure the production system will inherit. Dirty data is the cheapest thing to fix before you spend on AI, and it is the most common reason a pilot that worked in a clean sandbox stops working at scale.

There is a related trap in the governance column. Two-thirds of office professionals at companies with over $500m in revenue have used unauthorised AI tools at work, while only 18% of organisations report having formal AI security policies. That means many businesses already have AI in something close to production, informally, run by staff, with no baseline and no owner. It just does not appear on any budget line.

Where to start if the numbers above describe you

If you read the PwC figure and recognised your own last twelve months, the useful next step is not to pick a better use case. It is to find out which of the six underlying conditions is missing, because a pilot will fail on the weakest one regardless of how good the other five are.

That is what the free AI readiness audit is for. Thirty questions, 7 minutes, six pillars: Data, Process, People, Technology, Strategy, Governance. It gives you a score out of 120, a band for each pillar and a short synopsis. Every answer is self-reported, so it is a structured self-assessment rather than a systems inspection, and it does not produce an action plan at the free tier. What it does do is show you which pillar is going to sink the next pilot, before you fund it.

The first question in the Data pillar is the one most leaders get wrong, and it is deliberately blunt: if you needed a clean, accurate list of your best 500 customers right now, how confident are you in what you would get? Answer that honestly and you already know something about your odds of reaching production.