Stop Asking Which AI Agent Is Best. Ask What Job You're Hiring It For

Every vendor pitch I sit through this year opens the same way: their agent is the best one. Best at research, best at coding, best at customer support, best at everything if you let the deck run long enough. I've stopped finding that persuasive, and I think most CEOs should.

My own team at Digit just published a piece of research aimed at ordinary users trying to build a personal AI toolkit, and buried in it is a line I'd put in front of any procurement committee: don't ask which AI agent is best, ask what job you're hiring this agent to do. That's not a consumer tip. That's the entire discipline a lot of enterprise AI spending is currently missing.

There Is No Single Best Agent, at Any Budget

The research lays out something worth internalising: AI agents aren't one category of tool, they're at least six — general-purpose assistants, productivity agents living inside your existing office suite, automation agents connecting apps together, coding agents, business-platform agents inside your CRM or support system, and local or private agents for anything that shouldn't leave your control. A tool that's genuinely excellent in one of those categories is often mediocre or irrelevant in the others.

Most enterprise AI budgets I see get spent as though this weren't true — one flagship platform license, applied everywhere, on the theory that consolidation equals simplicity. Sometimes it does. More often, it means a company is running a general-purpose assistant on a job that actually needed deep CRM integration, and calling the gap a training problem instead of a category mismatch.

The Five Questions That Matter More Than the Demo

Before any agent gets budget, my own filter has converged on something close to what this research recommends for individuals: what is the actual task, what data will it need to touch, what actions should it be allowed to take on its own, what absolutely requires a human's approval, and — the one people skip — what happens if it's simply wrong.

That last question is the one that should set your risk tolerance, not the vendor's confidence in their own demo. A tool that occasionally produces a mediocre first draft is a low-stakes bet. A tool that can independently send a message, approve a refund, or touch customer records is a different category of decision entirely, and it deserves a different level of scrutiny before it gets anywhere near production.

The right agent isn't the most powerful one you can afford. It's the one that solves your recurring problem safely, affordably, and repeatedly — which is a much narrower bar than most vendor pitches suggest.

Treat Every Tool Like It Might Have to Leave

The part of this research I've adopted most directly for how we evaluate tools at Times Network is the least glamorous: before making any tool central to a workflow, ask whether you could get your data, prompts, and processes back out of it if you had to.

A surprising number of companies build real operational dependence on a platform without ever answering that question, and only discover the answer is no during a renewal negotiation with very little leverage left. I've started treating that exit-plan check the same way I'd treat due diligence on an acquisition target — because functionally, giving a vendor deep access to your workflows and data is a decision with the same shape as one, even when the invoice looks nothing like it.

What This Has Meant for Us

Across the AI transformation we've run at Times Network — eleven products launched in under two years, real AI embedded across the group rather than centralised in one department — the discipline that's actually held up isn't picking one dominant platform and forcing every use case through it. It's closer to what a sensible individual toolkit looks like in this research: one reliable general-purpose tool for the broad stuff, specific integrations where the ecosystem already lives, and narrow, purpose-built tools only where they've proven they save real time or generate real revenue.

If a tool doesn't clear that bar, it's not infrastructure. It's an expensive experiment still waiting to be cancelled.

The Discipline, Not the Logo

AI tooling will keep changing faster than any procurement cycle can track cleanly — new names, new pricing, features that merge or vanish inside a bigger platform's next release. Building organisational skill around a specific vendor's logo is a losing bet no matter how good that vendor is today.

What's worth building instead is the habit underneath it: define the task precisely, choose the right category of tool for it, limit what it's allowed to do until it's earned more, and always know how you'd walk away. That's a duller discipline than chasing the best agent on the market. It's also the one that's still going to be right in three years, regardless of which vendor is winning the demo circuit by then.