I normally try to make sure I don’t go off on too much of a rant when I write a blog, but I’m confused, and a little confounded, by something I keep seeing on repeat.

As a by-product of what my company does, we speak to a lot of people across industry sectors and all shapes and sizes, and every one of them is doing something with AI. We’re running a pilot for this. This team is piloting that swanky new AI capability. The business wants us to “do AI”. And almost everyone says the same thing: our data is holding us back, we don’t have the right controls in place, and we don’t trust our data to give us a good outcome. That is what I’m hearing, and it’s backed up by almost every survey and piece of research you read.

So, ultimately, we all know this to be true.

We spoke to one firm recently that was worried that, once AI wasn’t the focal point for the board, they wouldn’t keep getting budget for their governance programmes. So why is it such a struggle for these companies to get buy-in, or budget, from the business for governance? And why do they think the board won’t give them the support they need to run these programmes?

The whole thing seems mad to me. A bit of a rant to start with, but I think it’s something worth digging into.

It isn’t just what I’m hearing

This isn’t just hearsay or anecdotal.

The research tells the same story, at an enormous scale. Dun & Bradstreet’s 2026 survey of 10,000 businesses found that 97% have active AI initiatives (Pulse). Yet IDC, working with Lenovo, found that for every 33 AI proofs of concept a company launches, only four make it into production (CIO) – that’s 12%.

MIT’s Project NANDA reported that 95% of the organisations it studied saw no measurable return from their generative AI spending (Computing). That is a small sample, but the direction of travel is hard to argue with. And Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, because of rising costs, unclear value and weak risk controls (Gartner).

All of that costs money, people’s time, and the opportunity to have built something that would have worked. That is a lot of money spent on pilots that never get going. And, as the rest of this piece shows, we already know why, and probably knew why before we even started.

Measure twice, cut once

There’s a common saying from manual trades like carpentry and tailoring: measure twice, cut once. You measure, you check, you measure again. Only then do you cut. It feels slow. It is not. You do that because a wrong cut can’t be un-cut, and the timber or fabric is the most expensive part of the job, so you don’t want to waste it.

Most AI pilots are the opposite. Cut first, measure later, and hope the offcut is usable.

In AI, the measuring is the unglamorous bit. Do we know what data we have? Do we know where it came from, who owns it and whether we are allowed to use it? Do we know what success looks like, and who is accountable for it? If we can’t answer those questions, the pilot is not an experiment. It is a very expensive guess.

This is AI pilots in a nutshell. Companies are cutting without measuring; that’s the part people seem to skip, and in the long run it could prove costly – whether that’s just wasted budget and resources on pilot projects, or worse, a data breach caused by lack of governance and control.

Why the pilot gets funded and the foundation does not

Back to where I started: why is it easy to find money/time/resource for the pilot and so hard to find it for the data work underneath? To me it comes down to few different things:

  1. It is not exciting. Nobody gets a standing ovation for tidying up data – it’s genuinely the least sexy thing in the entire process. A pilot has a demo, a screen and a story. Data governance has a policy, a catalogue and a meeting. One of those gets approved in a week.
  2. It still sounds like a cost. Governance has spent years being sold as the thing that keeps you out of trouble. Avoid the fine. Pass the audit. That is a fair job, but it frames governance as insurance, and insurance is the first thing to be trimmed when budgets get tight.
  3. The benefit is hard to put a number on. If governance works, nothing bad happens, and it is difficult to show a chart of the disaster that didn’t happen. Governance teams rarely get to say “this saved us X”. The AI team gets to say “this will save us Y”, even when Y is a hope rather than a measurement.
  4. The biggest corporate FOMO out there. The pressure to be seen doing something with AI is enormous. “Fail fast” has become the default answer to every hard question. That works in a sandbox. It does not work in a regulated business, and regulators have been clear that they do not want experiments running loose on customer data.
  5. The project itself is not clear. This one is uncomfortable. Many pilots start without a named owner, a defined outcome, or an agreed way of knowing whether it worked. If nobody can say what good looks like, nobody can say what data it needs, and nobody can justify investing in it.

Put those together and the result is predictable. The visible, exciting work gets funded. The foundation, which decides whether it ever works, goes begging. It also explains the worry that firm shared with us. When governance funding rides on how exciting AI looks to the board, it is only ever one change of focus away from being cut.

The reasons are not a secret

This is the bit that gets me. The evidence is not hidden, and almost every major analyst says the same thing in slightly different words.

  • Gartner predicts that through 2026, organisations will abandon 60% of AI projects that are not supported by AI-ready data. Its own survey found that 63% of organisations either lack the right data management practices for AI or are not sure whether they have them (Gartner).
  • Gartner also expects at least 30% of generative AI projects to be dropped after the proof of concept stage. The reasons it gives are poor data quality, inadequate risk controls, rising costs and unclear business value (Gartner guidance, via Brightfin).
  • In the same Dun & Bradstreet survey, only 5% of businesses said their data is ready to support AI beyond pilots (Pulse).
  • Even the optimistic story has a catch. Most leaders expect a year or more before AI returns go beyond basic efficiency, which is a long time to run on shaky foundations (PwC, via Let’s Data Science).

Look at the reasons behind the failures. Data quality. Risk controls. Unclear value. Not one of them is “the model was not clever enough”. These are measuring problems, not cutting problems. And underneath most of them sits the same root cause: we cannot properly trust, secure and govern the data these systems run on, or the models themselves.

How to close the gap

None of this means slowing down. It means measuring in a way that makes the cut faster, more accurate and better quality. Here is what I have seen work.

  1. Start with an owner and an outcome. Before a pilot begins, write one sentence: what will change, by how much, and who is accountable. If you can’t write that sentence, the project isn’t ready, and no amount of data will fix it.
  2. Check the data first. For the use case in question, ask four plain questions. What data does it need? Is it good enough? Who owns it? Are we allowed to use it this way? This is a few weeks of work, not a programme, and it saves months later.
  3. Talk about value, not fines. Stop selling governance as protection from penalties. Sell it as the thing that lets AI get into production. Then put numbers on it. How many pilots did we fund last year? How many are live? How much time do our teams spend hunting for and cleaning data? That is the business case (the hardest part here might be getting answers to those questions in the first place!!).
  4. Build the pilot on the production path. Run it on a properly governed slice of real data, with the right access and controls from day one. A pilot that only works on tidy sample data proves very little.
  5. Fail fast, but inside walls. Experiments are healthy. Experiments on customer data with no guardrails are not. Agree what is safe to try quickly, and what needs checking first. Done well, this makes teams faster, because they are not stopped halfway by a question nobody planned for.

The cut is the easy part

The measuring is not the delay. It is the reason you cut where you wanted to. Getting that wrong is ultimately more expensive in every aspect than doing it right in the first place!

If you are looking at your own pilots and wondering which ones will make it, ask yourself this: do we know why we’re doing it? Do we know what success looks like? Do we know what data is feeding the pilot, and do we trust it?

If you can’t answer those questions with confidence, you’re not alone, and it doesn’t mean your pilots are doomed. It means the measuring needs doing first. At Nephos, that’s exactly what we help organisations with: getting their data ready for AI so pilots make it into production. If you’d like to talk through where your own pilots stand, get in touch with our team. We’re always happy to have the conversation.

Lee Biggenden

Lee, the Co-founder and CCO of Nephos, brings a wealth of experience and a pioneering spirit to the forefront of data system integration. Lee's thought leadership content offers invaluable insights into transforming data storage, processing, governance, and protection. Through his writings, Lee shares the latest trends, challenges and advancements in the data technology landscape - helping organisations to not only adapt but thrive in the digital era.