A founder told me recently — with genuine pride — that his firm had gone "all in on AI." Licences for the whole team. A prompt-engineering workshop on a Friday afternoon. A Slack channel full of screenshots and small victories. The enthusiasm was real, and I don't mock it. Enthusiasm is where everything starts.
I asked him one question: what changed in your P&L?
The pause told me what six months of experiments hadn't. A few thousand dollars a month in subscriptions — S$2,840, when we finally added it up — and not a single line in the management accounts that had moved because of AI. Not revenue. Not cost. Not margin. Not one hire avoided or deadline beaten.
He isn't unusual. He's the norm. And the uncomfortable part is that nothing about the tools failed. The failure was structural — and it was installed on day one.
The Pilot Was Never Designed to Land
When I unpick a stalled AI pilot — and I've unpicked several this year — the same three absences appear every time.
No baseline. Nobody measured the process before AI touched it. How many days does month-end close actually take? How many hours does one proposal consume, across how many people? What does a week of status reporting cost in labour? If you can't state what a process cost on the day you started, you can't state what AI saved. Every claim after that is a feeling.
No target. "Let's see what the team does with it" is curiosity, not deployment. A target is specific: close the books in four days instead of eight; first-draft proposals in two hours instead of two days. Without a target, the pilot can't fail — which is another way of saying it can't succeed.
No owner. The subscription sits with IT, or with whoever's corporate card was nearest. Usage belongs to the enthusiastic few. Results belong to nobody. And when the novelty fades — it always fades — there's no name against the number, so the number quietly stops being asked for.
A pilot without a baseline, a target, and an owner isn't an experiment. It's a demo with a monthly invoice.
Tools Are Not Capability
There's a distinction most AI conversations skip entirely.
A tool is something you pay for. A capability is something your business can reliably do — documented, tested against real work, usable by more than one person, and still there when the person who discovered it resigns.
Buying Excel never gave anyone a finance function. Buying a CRM never gave anyone a sales process. Yet somewhere along the way we collectively decided that buying AI subscriptions gives a business AI capability. It doesn't. It gives the business access — and access without process is how an entire team of capable people ends up using a frontier technology to reword emails.
Demo-Driven Adoption, Measured Deployment
The standard SME adoption pattern is demo-driven. Someone sees an impressive demonstration — at a conference, on LinkedIn, from a nephew — and buys in. Usage spikes in week one, halves by week six, and settles into a long tail of light editing. Then the renewal invoice arrives and the question gets asked backwards: are we getting value from this? Nobody knows. Nobody could know. Nothing was measured on the way in.
Measured deployment inverts the sequence. Choose the process first — ideally one that's contained, repetitive, and expensive. Baseline it. Set the target. Deploy against it. Measure monthly. Adoption follows results, instead of results being hoped for from adoption.
The CFO Fix
The fix isn't technical, which is precisely why it works. It's the discipline a CFO applies to any deployment of capital: baseline, target, owner, monthly measurement.
Here's what that looks like in practice. Take a mid-sized Singapore services business weighing an AI transformation programme. Before a single skill gets built, you baseline the operation: month-end close at 8 days across hundreds of monthly transactions. A steady cadence of pitches, each one swallowing a senior team for the better part of a week. Dozens of concurrent projects on a team already running flat out.
Then the targets, written down before any deployment: close compressed from 8 days to 2–4. Pitch cycles cut by around 40%. Capacity modelled as the equivalent of 10–14 additional hires — without hiring. Every target gets a named owner. Every month, actuals are read against the model, exactly the way you'd review budget variance.
Those are targets, not results — the numbers a programme like this commits to before it starts, and I've framed them that way deliberately. I have no interest in the consultant's habit of declaring victory early. But notice what's already different from every stalled pilot I've examined: the moment such a programme moves the P&L, you know precisely by how much. And if a skill doesn't move it, you know that too — by month two, not month fourteen.
Stop Buying Pilots. Start Building a Library.
The deeper shift is from pilots to assets.
Treat each AI workflow as a skill: a documented, tested, transferable unit of capability — costed honestly at roughly S$1,000 each when built properly, owned by your team rather than rented from a vendor, and trainable into every new hire who walks through the door. Ten skills make a toolkit. A hundred make a library — and a library compounds, because a brief-writing skill feeding a research skill feeding a reporting skill creates chains of capability no single subscription ever will. Building that library, measured monthly against a published model, is the entire premise of the AI Transformation Programme.
That's the real difference between the firms whose AI spend shows up in the P&L and the firms still forwarding each other screenshots. It was never the tools. It was never the team, either.
Tools depreciate the day you buy them. Capability compounds the day you build it.