Making Sense of the AI Frenzy

Here's what's actually driving the AI frenzy inside most organisations: everyone in the room has a different idea of what AI is for.

The CFO pictures cost savings. Managers picture more deliverables. The intern pictures ChatGPT quietly generating the output nobody wants to write.

None of them are wrong. They're just optimising for different things — which is exactly how a Head of Department ends up with a dozen AI proposals a quarter, and no shared way to tell a good one from a bad one.

Before deciding whether something should be built, it helps to ask a more boring question: is it sustainable, is it worth it, and is it safe to run.

Sustainable, first

Sustainability comes first, because most AI projects don't fail at launch. They fail quietly, eighteen months later — the model drifts, the pipeline breaks, the person who built it leaves, and nobody owns it.

AI adoption doesn't collapse in public. It dies quietly, the moment people stop trusting the system around them.

Real sustainability means planning for maintenance from day one. Giving every system a named owner. Fixing data quality before scaling the use case. Building AI literacy alongside the tooling, not after it.

Say no on purpose

Every filter needs teeth. These are the moments to use them:

• Say no when the problem isn't well-defined enough to solve.
• Say no when a simpler fix would do the job without the maintenance burden.
• Say no when the data isn't clean enough to trust, or errors would be costly with no real safeguard.
• Say no when the real motivation is "we're using AI" rather than solving an actual pain point.

Is it worth it?

Everything that survives that filter should still answer one more question: is this worth it.

Price what the problem costs today against what building and running the system will actually cost. Be honest about the use case, not the pitch. Pilot cheaply before committing, and think through the second-order effects — on workflows, on headcount, on trust.

How much oversight does it need?

Once something clears that bar, the last decision is usually the one that gets skipped: how much human oversight it needs.

• Low-stakes, reversible work needs occasional spot-checks.
• Medium-stakes work needs approval before action, tapering as trust builds.
• High-stakes, irreversible work needs mandatory sign-off and an audit trail, every time.

The oversight has to be real — a way to override, to flag errors, to feed corrections back in. A review step that just rubber-stamps isn't oversight. It's decoration.

None of this requires everyone to agree on why AI matters. It just requires agreeing on how you'll decide.

The organisations pulling ahead aren't moving fastest. They've built a shared filter — everyone evaluating proposals on the same basis, instead of a dozen private ones.

Frenzy is optional. A shared filter isn't.