Why Most AI Projects Fail Before They Start

‍

Most AI projects do not fail in deployment. They fail much earlier, in the assumptions made before a single line of code is written. The business decides it needs to adopt AI, a budget gets approved, a vendor gets selected and the project begins. What rarely happens first is the question that determines whether any of it will work: what outcome are we trying to achieve, and do we have the data and the organisational design to support it?

‍

That question sounds straightforward. In practice, most organisations skip it entirely, and the consequences tend to surface at exactly the moment the business can least afford them.

‍

Efficiency Is Not a Strategy

‍

The most common starting point for an AI project in a growing business is a cost or efficiency target. The board has seen what the technology can do, has been told what competitors are doing and wants to see progress. The result is a project scoped around reducing headcount in a particular function or automating a process that was previously manual and measured against the time saved.

‍

There is nothing wrong with efficiency as an outcome. The problem is that efficiency projects tend to be designed around existing processes rather than around the outcomes those processes are meant to serve. The organisation ends up automating something that probably should have been redesigned and the AI layer inherits all the structural limitations of what it was built on top of.

‍

"If you've got an organisational model that is still designed for a linear process, what you're going to get by putting AI on top of it is some efficiency. The design premise has to be what outcome do you want, not what process do you currently run. That shift is really important." - Kirsty Rutter, Head of Venture Capital, Lloyds Banking Group

‍

This is the distinction that separates AI projects that compound in value over time from those that deliver a one-time efficiency gain and then plateau. Outcome-led design requires a different kind of conversation at the start, one that most project briefs never get to.

‍

The Data Problem Nobody Wants to Admit

‍

The second reason AI projects fail before they start is data. Specifically, the gap between the data a business believes it has and the data it actually has in a state that is useful for the purpose being proposed.

‍

This gap is more common than most leadership teams realise, and it tends to be discovered at the worst possible moment: after the project has been scoped, the vendor has been contracted and the first sprint has revealed that the underlying data is inconsistent, incomplete or stored in systems that cannot talk to each other cleanly.

‍

Lee Provoost, CTO of Flagstone, made a deliberate decision to hold back from aggressive AI adoption in 2025 and focus instead on the organisational foundations that would make that adoption genuinely useful when it happened. His reasoning was direct: businesses that have not addressed their structural and data foundations will not get the return from AI that they are expecting, and the cost of discovering that mid-project is significantly higher than the cost of addressing it first.

‍

"There are so many basics that companies tend to forget. If you actually get your house in order, a lot of companies could easily get a 20 to 30 percent productivity boost without even touching AI or GitHub Copilot or any of those products." - Lee Provoost, CTO, Flagstone

‍

The Governance Gap

‍

The third failure mode is one that tends to get the least attention at the project scoping stage and the most attention from regulators and auditors after the fact. AI systems that are deployed without a clear governance framework, without explainability built into the design and without a defined process for auditing outputs, create a category of risk that is very difficult and very expensive to address retrospectively.

‍

Kirsty Rutter, whose career has spanned innovation, risk and regulation across some of the UK's largest financial institutions, is direct about where she sees the market failing on this.

‍

“The real winners are going to be either the tools that are written with risk and control built in, or services that serve that need. If you haven’t got it, it is going to cause you so much trouble. That is still what I believe is where, in a couple of years time, it will be what stops all the really bad stuff happening.” – Kirsty Rutter, Head of Venture Capital, Lloyds Banking Group

‍

For PE-backed businesses operating in financial services, healthcare, logistics or any other regulated sector, this is not a theoretical concern. It is a live risk with direct implications for regulatory standing, valuation and the ability to scale.

‍

What a Well-Designed AI Project Actually Looks Like

‍

The businesses that get genuine, compounding value from AI investment share a small number of characteristics. They start with a clearly defined outcome rather than a technology. They audit their data before they scope the solution, they build governance into the design rather than appending it afterwards and they have technical leadership in place that can hold the business and the technology in mind simultaneously, asking the right questions at the point where the answers are still cheap.

‍

At Gathered and Found, we work with leadership teams at the start of that process, before the brief is fixed and before the vendor is contracted. The questions that determine whether an AI project will deliver value are almost always easier and less expensive to answer at that stage than at any point afterwards.

‍

‍

At Gathered and Found, we help PE-backed businesses design AI programmes that are built around outcomes, supported by clean data foundations and governed from the start. If you are scoping an AI project and want to pressure-test the brief before you commit, get in touch.

‍