The Cheapest AI Advantage Nobody's Talking About: Saying No to Most of It

I've noticed a pattern in almost every AI budget review I've sat through this year: the conversation is always about what to add next. A better model, a new agent platform, another integration. Almost nobody in the room is asking what should have been cancelled six months ago.

There's some genuinely useful research doing the rounds on running AI models locally, and one throwaway line in it has stuck with me more than the technical detail around it: local AI has no per-message subscription meter, but it can quietly develop another bill — storage, complexity, tool sprawl, the accumulated weight of things nobody remembers signing up for. That's not really a statement about local AI. It's a statement about every AI decision a company makes right now, cloud or otherwise.

Bigger Isn't Better, It's Just Bigger

There's a principle buried in this research that applies directly to enterprise AI spending, even though it was written about choosing which model to download onto a laptop: a model that's marginally better on a benchmark but twice the size is often a poor choice, because everything about running it costs more — time, memory, money — for a capability gain nobody actually needed for the task at hand.

Enterprise AI procurement makes this mistake constantly, just with bigger numbers attached. Teams default to the most capable, most expensive tier of a platform because it feels like the safe choice, when the actual workload — drafting routine correspondence, summarising internal reports, classifying support tickets — would run just as well, and far more cheaply, on something several tiers down. The research calls this "quality per gigabyte." In a budget conversation, I'd call it the discipline of buying for the job, not for the spec sheet.

The smartest toolkit isn't the largest one. It's the smallest set that actually covers your real workloads.

Tool Sprawl Is the New SaaS Sprawl

Every CFO I know has already lived through one wave of this: a company that signed up for a dozen SaaS tools solving overlapping problems, none of them fully used, all of them still on the books eighteen months later because cancelling anything felt riskier than leaving it running. AI subscriptions are heading toward exactly the same pattern, faster, because the tools are cheaper individually and easier to adopt without anyone senior noticing.

The research is blunt about the discipline that prevents this, aimed at an individual managing their own model library: adopt a monthly audit, keep a simple note of what you're running and why, and delete anything superseded rather than keeping every version "just in case." That habit scales up almost unchanged. A company running an honest quarterly audit of every AI tool actually in use — not licensed, *used* — will usually find real money sitting in tools nobody remembers approving.

The Question That Should Gate Every New Tool

If a tool does not save time, improve output, or reduce errors, it's entertainment, not infrastructure. That's a hard line, and it's the right one. Before any new AI tool gets budget at Times Network, the question I actually want answered isn't "is this impressive" — most demos are impressive — it's whether it clears that specific bar, with a real number attached, not a hopeful one.

The tools that survive that question tend to be unglamorous. They handle one recurring, well-defined job extremely well. They're rarely the ones anyone brags about in a conference talk. They're also the ones still delivering value a year later, instead of quietly renewing themselves into next year's cost base.

What This Has Actually Meant for Us

Across the AI transformation we've run at Times Network — eleven products launched in under two years, real EBITDA overachievement alongside it — the discipline that's held up hasn't been about spending more aggressively on AI. It's been about applying the same scrutiny to an AI tool that we'd apply to any other recurring cost: does this earn its place, measured against what it actually replaces, not against how capable it sounds in a pitch.

The Boring Advantage

Nobody writes a case study about the AI subscriptions a company cancelled. It doesn't make for an exciting board slide. But in a market where every vendor is incentivised to sell you more capability than most of your workloads need, the company that's simply more disciplined about saying no to the 80% of AI spend that isn't earning its place will end up with a real cost advantage over the company that said yes to all of it.

That's not a technology strategy. It's the same operating discipline that's always separated well-run companies from the rest — just applied to a category of spending that's new enough that most boards haven't started asking the question yet.