AI can do it. Should it?

Most organisations are asking the wrong question about AI. They're asking what it can do. The harder question, and the one that actually shapes what you become, is what you'll let it do.

Those are not the same question. AI can run a call centre. It can handle onboarding, generate the policy summary, answer the member at 2am. The capability is there or nearly there. But capability isn't the decision. A call centre run entirely by AI might be cheaper and faster and still be the wrong choice, because the call is where a person decides whether they trust you. Hand that to a machine to save money and you've automated away the thing the relationship was built on. That's not a technology question. It's a judgement about what kind of organisation you want to be, and it sits with the executive, not the tool.

This is the part most technology leaders skip. AI arrives framed as automation — what can we take off people's plates — and the whole conversation collapses into cost and efficiency. But an organisation built well around AI isn't defined by what it automated. It's defined by what it deliberately chose not to.

Despite $30–40 billion in enterprise investment, 95% of organisations are seeing no measurable return on AI.

MIT — State of AI in Business 2025

That number gets read as proof AI doesn't work. It isn't. The technology works. The organisations failed it — and the two experiments I ran show where.

I tested the capability side directly. I run a small IT team in an NFP, where the gap between what needs doing and the headcount to do it is permanent, so I wanted to know how far AI could close it. Two experiments. A tightly defined script worked first go and turned a developer job into half an hour of work — real capability a small team couldn't otherwise afford. A larger, looser build drifted as it grew, filled the gaps with its own assumptions, and produced something that looked finished and wasn't. It lied, and it held the line when challenged until I put the evidence in front of it.

What that told me is where the value actually moves. AI crushes the work once the thinking is done — the BA documentation, the admin, the tasks where you already understand the problem and it's the execution that eats the time. Give it that and, in my experience, you get real gains - around twenty percent in the right roles. But it needs the understanding supplied up front, and it needs relentless checking on the way out, because it will confidently hand you something wrong. The value doesn't leave the human. It moves to the two ends AI can't hold: defining the problem, and judging the result.

Which is why an AI-empowered organisation is a redesigned one, not a retrofitted one. You don't get there by handing the existing team a set of tools and running the same structure. The roles have to change. Some work that was a whole job becomes a slice of one. New work appears around defining, prompting, validating. And it only runs on clean, consolidated data — feed AI a fragmented mess and it produces a confident, fragmented mess. Data integrity isn't a phase. It's the precondition for any of it working.

None of that is comfortable, and that's the real barrier. Redesigning roles means hard conversations about what people do and whether they can do the new version. Most organisations won't. They'll buy the tools, leave the structure alone, declare themselves AI-enabled, and get worse outcomes delivered faster. The ones that pull ahead will be the ones willing to make the uncomfortable calls — redesign the work, hold the line on data, and decide, deliberately, what they will never hand to a machine no matter what it can do.

That last decision is the one that matters most, and it's the one AI can't make for you.