Between 60 and 95 per cent of enterprise AI initiatives produce no measurable business value, depending on how strictly value is defined. The Groundwork Advantage is our new handbook on what has to be true inside an organisation before AI produces anything worth having, and on which of those decisions belong to the C-suite rather than to IT. This article sets out the argument behind it and what you’ll find in it.
Most of what has been published about AI in the past two years describes what the technology can do. Very little of it describes what has to be true inside a company before that technology produces value.
That second question decides whether an AI initiative survives its first budget review. It’s also the one we keep running into with the leadership teams we work with, usually after a pilot has stalled and nobody can say precisely why. So we wrote our answer down. The Groundwork Advantage is a handbook for the C-suite, vendor-neutral by design, and the reasoning in it holds regardless of who executes the work.
Why enterprise AI initiatives fail
Enterprise AI fails most often because of the conditions it’s asked to run on: business processes, data and governance that were never built to support it. Most companies already have access to models that are more than capable of the work being asked of them.
RAND’s interviews with 65 data scientists and engineers found the most cited causes were leadership specifying the wrong problem, and data that wasn’t good enough to train on. MIT’s Project NANDA found that 95 per cent of generative AI pilots showed no measurable impact on profit and loss. PwC’s 2026 research puts the technology at roughly 20 per cent of an initiative’s value, with the other 80 per cent coming from redesigning the work around it.
Organisations that skip that redesign and go straight to deployment defer the work rather than avoid it, and they pay more for it later, once they’re rebuilding data architecture and retrofitting governance around a system that’s already in production.
What has to be true before AI produces value
Four things determine whether an organisation is ready: whether it owns its data, whether regulation is designed in from the start, whether existing systems of record are treated as assets, and whether adoption is building one foundation or fragmenting into dozens of disconnected tools.
Data ownership determines whether you control your own AI trajectory or depend on a vendor’s pricing and roadmap. Owning your data means knowing where it lives, who can access it, how it flows into any system that touches it, and being able to move that system elsewhere without losing the value already built into it.
Regulation belongs in the design. Deloitte’s 2026 Tech Trends research found that the gap between organisations piloting agentic AI, at 38 per cent, and those running it in production, at 11 per cent, comes down to governance and integration: a named owner, an audit trail, and a defined process for what happens when the system gets something wrong. The requirements of GDPR and the AI Act, explainability, data lineage and human oversight, are the same properties that make an AI system trustworthy enough to scale.
Existing systems of record can usually stay where they are. What most organisations are short of is a way to connect the data inside the ERP and the CRM to something that can reason over it. Replacement is the most expensive assumption in AI adoption, and it usually starts driving the budget before anyone has tested it.
That leaves the foundation itself, which has to run somewhere. Sinas, an open-source AI runtime platform we’ve developed at WeAreBrain, is our own answer to that requirement. It indexes enterprise knowledge, enforces governance and orchestrates retrieval across use cases, and because it’s open source it belongs to the organisation running it. We include it in the handbook as a concrete illustration of what a foundation layer has to provide, once, so that no later use case has to rebuild it.
How to choose a first use case, and know what it’s worth
Score candidate use cases on volume, friction and data readiness together, then estimate the expected value in three separate categories before committing any budget.
Volume asks how often the process happens, friction asks how much of the current effort is retrieval and coordination rather than genuine judgement, and data readiness asks how much work is required before the use case can run on governed data. Scoring well on one factor and badly on the others is a trap: high volume running on unready data is how pilots fail in public.
The value estimate then splits three ways, and blending them into a single ROI figure is how the most important one disappears.
| Impact category | What it covers | When it becomes visible | Why it gets lost |
|---|---|---|---|
| Direct efficiency | Hours and cost saved on the process itself | Almost immediately | It doesn’t, which is why it crowds out the other two |
| Quality improvement | Error reduction, consistency, customer and employee experience | With a lag of months | Hard to attribute once other changes land |
| Strategic option value | What becomes possible afterwards that wasn’t possible before | Later still, often after the second use case | Hardest to quantify, so it disappears into the other two when the estimate is blended |
BCG’s research across 1,803 senior executives found that 60 per cent of companies have yet to define or monitor any financial KPIs for the value AI creates. Much of that is a measurement gap: the impact was never estimated with this kind of structure beforehand, so there was nothing concrete to measure against afterwards.
Deciding, building and scaling
The work happens in three phases, each with its own owner, its own decisions and its own completion test. We’ve run the sequence often enough that it’s become a standing methodology inside WeAreBrain, which we call the Future Catalyst.
The Idea Forge is where you decide what to build first, and it’s complete when you have a validated, quantified case for that decision. The Reality Engine is where the use case and the foundation underneath it get built, and it’s complete when the outcome has been measured directly against the estimate written beforehand. The Growth Driver is where the next use cases get built. It’s working when each one takes less time and money than the last.
That last phase is where the foundation argument pays off. Gartner’s 2026 survey of 1,303 organisations found that only 22 per cent have successfully scaled AI across multiple business units, while 85 per cent of functional leaders plan to increase AI spending. Build the first use case on governed, reusable infrastructure and the second inherits the access controls, the auditability and the data indexing already in place. Build it standalone and the second one starts the foundation over again.
What’s in the handbook
The Groundwork Advantage runs to seven chapters. Anyone working on AI adoption will get something out of it, and for whoever signs off the budget it’s a must read.
It covers the principles that separate the organisations that scale AI from those that don’t, the Groundwork Model for prioritising where AI creates measurable value, the three-phase sequence above with the ownership questions each phase demands, and what AI-assisted software development does and doesn’t change about any of it. It closes with a five-question self-assessment and a 17-point checklist you can work through with your own leadership team.
Key takeaways
- Between 60 and 95 per cent of enterprise AI initiatives produce no measurable business value, and the research traces that back to data, process and governance rather than to model capability.
- The real investment decision is whether your data, process and governance are in a state where any AI tool could succeed, which makes the choice of model or vendor a comparatively minor one.
- A first use case should be scored on volume, friction and data readiness together, and chosen for the quality of its data rather than the scale of its ambition.
- Expected impact should be estimated in writing across direct efficiency, quality improvement and strategic option value, because a single blended ROI figure hides the category with the largest long-term value.
- Build the foundation once and each new use case starts from the application layer, cheaper and faster than the last.
Get the handbook
The Groundwork Advantage is free to download. The 17-point groundwork checklist sits at the back of it, if you’d rather start there.
Download The Groundwork Advantage
If you want to pressure-test any of this against your own situation, we run the same assessment with organisations preparing to scale AI. Get in touch and we’ll look at it with you.
