Closing the AI ROI Gap: Where Projects Struggle, and What Works

Drawing on research from 2026, a consistent pattern shows up: most organisations investing in AI are finding it harder than expected to turn that investment into measurable value. McKinsey found that close to 90% of organisations now use AI in at least one business function, yet only a small fraction report real financial impact from it. RAND’s analysis puts the gap starkly: more than 80% of AI projects fall short of their intended value — roughly double the rate for comparable non-AI IT projects. MIT’s research on enterprise generative AI found a similar pattern — 95% of pilots hadn’t yet produced a measurable return. And S&P Global Market Intelligence found the number of organisations pausing or abandoning AI initiatives nearly two and a half times over in a single year.

It’s tempting to read that as evidence the technology isn’t ready. The gap is in the organisation around technology, not the models themselves. The fix isn’t a better model, it’s a more considered approach to everything surrounding it. Read on as we outline our approach to solution.

Alignment

Business and technical leadership often start from different pictures of the same business. Ask a sponsor what’s needed and you’ll hear priorities shaped by what they’ve seen elsewhere; ask the people using the tool day to day and it looks different again; ask engineering what today’s systems can actually support, and you get a third picture. None of these views is wrong — they’re just rarely reconciled before the work starts, so feasibility, readiness, and even what “success” means end up meaning different things to different people in the same room.

At Work Perfect, we run a structured discovery process before any use case is chosen: business decision-makers and technical leaders in the same room, agreeing on the current state, the vision, the gap between them, and the real constraints. It takes a little longer on day one, and saves considerably more across the months that follow.

Foundation, governance, and security

Most organisations’ data was built to be read by a person, not a machine — a reasonable choice at the time. Years of growth add their own layer: a CRM here, a spreadsheet workaround there, a system kept alive after an acquisition — redundant, duplicated, quietly costing more than anyone’s counted. What a person works around, an AI repeats as fact; ask two teams the same question and you’ll often get two different answers, each pulling from a different dashboard’s version of the truth. Underneath, governance — ownership, lineage, security, sovereignty, audit — is usually treated as something to revisit later rather than a day-one decision, and an agent with broad, undifferentiated access to everything only raises the stakes. The cost shows up fast once scaling starts: every new use case doesn’t just add effort, it adds risk, and the whole thing gets more brittle with each addition.

At Work Perfect, we build a governed, unified view of the data first — access control, lineage, audit and data residency designed from the start — by connecting to the systems a business already depends on, not replacing them. Role-based access and proper segmentation make a production rollout genuinely safe, for a twenty-person business as much as an enterprise. We treat the foundation as something that evolves with the business, instead of quietly becoming next year’s version of the same problem.

Vocabulary and adoption

The vocabulary for what “AI” means hasn’t caught up: a nightly report gets called automation, a churn model gets called AI, a scripted chatbot gets called agentic — three different things under one word, so the eventual ROI conversation ends up comparing categories that were never the same. And even good technology stalls if adoption isn’t planned: a new assistant launches with one training session and a page of FAQs, and within a month the team’s quietly back on the spreadsheet they already knew. Recent workplace research found nearly a third of employees admit to working around new AI tools rather than adopting them — usually a sign of the support people had, not a verdict on the tools.

At Work Perfect, we ensure that the room agrees on what automation, machine learning, and genuinely agentic AI each mean before any use case is chosen, and we assess, deliberately, where AI actually adds value and in what form — sometimes the honest answer is better automation, not an agent. Adoption is planned and resourced from the start, because early participation is what turns an investment into ongoing value rather than a shelved pilot.

Success criteria

Success criteria often get defined after the spending has already begun. Organisations getting real results are, by McKinsey’s own research, two and a half times more likely to have a documented process for deciding which projects to pursue. Most don’t: a project gets funded, someone builds something promising, and only afterwards does the conversation turn to what “worked” was meant to mean — by which point there’s no baseline, and no one clearly holds the outcome.

Every Work Perfect engagement starts with the business defining the value, naming a sponsor, and agreeing what success looks like — cost reduction, faster insights, less toil, wider participation from non-technical users, fewer dependencies on specific teams or people — before any technical build begins. It’s a mechanism, not a workshop exercise, and it doesn’t stop at the start line: an evaluation framework built into the engagement checks value along the way, enabling course correction before spend is sunk.

Production mindset

A working demo isn’t a production system, and the distance between the two is easy to underestimate — proper grounding, a retrieval layer that holds up at scale, test automation, the same SDLC and DevOps discipline as any other production asset. Skipping this either leaves a promising pilot stuck indefinitely, or — worse — sees the proof-of-concept waved through as if it were production. A PoC was never built for real load or real failure modes, so when it’s promoted anyway, the expensive disappointment arrives all at once.

At Work Perfect, we don’t stop at a demo, and we don’t ask a client to commit to a large build on faith. Every engagement begins as a short, evidence-based starting point on a client’s own data, scoped with the production path already in mind from day one — version control, shared ownership across teams, incremental rollout, faster troubleshooting — the same “prove it, then scale” discipline, just treated as the starting point rather than an aspiration.

We understand why most organisations hesitate to start

Most of the businesses we talk to aren’t short on ambition, but on confidence — and usually for good reason. Some have misconceptions about what AI can realistically do today. Some have absorbed a steady diet of stories about expensive disappointments. Some have already tried something that went nowhere. Others simply aren’t sure where to begin. None of that reflects poorly on them, and it isn’t solved by a more persuasive pitch — it’s solved by trying it, cheaply, before anything larger is on the table: real familiarity and confidence, on your own data, before a significant investment decision has to be made.

Funding is the other common hesitation, and a fair one — committing budget to something this uncertain is genuinely hard to justify upfront. We treat that as part of the job: before asking anyone to find new money, we look at where existing spend may already be going to waste — redundant tools, duplicated licences, projects quietly delivering little — because addressing that first can often materially offset, or even fund, the work that follows. It’s a way of de-risking the funding decision itself, not only the delivery.

Our unique position

At Work Perfect, we bring genuine depth on Google Cloud — recognised by Gartner as a Leader across analytics, conversational AI, and AI infrastructure — alongside proven expertise on monday.com, the system where a great deal of our clients’ actual work already lives. That means we’re rarely starting from zero in understanding how a business operates.

We chose Google Cloud deliberately, because its semantic and governance tooling reinforces this discipline — it doesn’t replace it. More on why we bet on Google Cloud — in our next piece.

Talk to us, we’re here to help

None of this is an exhaustive list, and it isn’t a diagnosis that applies equally to every organisation — not every business will recognise all five patterns above, and that’s exactly the point.

If any of the patterns above sound familiar — leadership not quite on the same page, a data foundation nobody fully trusts, governance and security still catching up, or a pilot that hasn’t made the leap to production — that’s a genuinely common place to be. No two teams are alike, let alone two projects or two organisations, and treating them as if they were is a risk of its own. So we take every engagement through a structured journey before big decisions get made. We call it Inception — stay tuned, coming soon.

That’s the conversation we’d like to have in the meantime. Get in touch, and we’ll show you what a short, evidence-based starting engagement looks like on your own data.

See you in the cloud!