I'm Not Worried About Automating the Workflow. I'm Worried We're Automating the Wrong One.

I'm not worried about the automation. I'm worried about the problem underneath it

Ask most engineers what keeps them up at night about an AI project and they'll tell you about latency, context windows, hallucination rates, orchestration cost. Ask me, and it's none of those. I'm more worried about solving the wrong problem than I am about how well we automate it. A beautifully engineered agent that ships fast and cheap is still a waste if it was pointed at the wrong target from the start.

That worry isn't hypothetical. In 2026, CEOs and CFOs at a lot of companies are saying, in effect, the same thing: we adopted AI, and we don't see the business grow. Not "the model is bad." Not "our team is slow." The performance just isn't showing up where it's supposed to.

The C-suite isn't imagining this

McKinsey's 2026 State of AI research found that only 6% of organizations qualify as "AI high performers" — companies attributing at least 5% of EBIT to AI and describing the impact as significant. The other side of that number: 63% report no measurable enterprise earnings impact at all, even as they expand deployment and see real gains in individual productivity.

MIT's widely-cited research on the "GenAI Divide" found something sharper still: 95% of generative AI pilots deliver no measurable impact on P&L. Not underwhelming — no measurable impact. Adoption climbed. The needle didn't move.

“Widespread deployment and flat earnings impact, at the same time, in the same companies. That's not a rounding error. That's a signal that something upstream of the automation is broken.”

They're not short-sighted. They're reading a different scoreboard than we are

It's tempting, if you're technical, to read those numbers as C-suite impatience — leadership wanting a quarter's worth of magic from a technology that takes years to mature. I don't think that's it. I think the CEO and CFO are asking a completely reasonable question — did this move the business — and a lot of us building the thing were answering a different question the whole time: did this work.

Those aren't the same question. A CTO or an engineer can be genuinely excited about what a new model or agent framework can do — and that excitement, on its own, says nothing about whether it's aimed at the constraint that's actually limiting the company's growth. Enthusiasm about the technology and clarity about the business problem are two entirely separate things, and it's easy to mistake one for the other when the tech is this interesting.

Two ways this goes wrong — one forgivable, one not

The first failure mode is almost sympathetic: teams use "AI initiative" budget and cover to finally fix things that badly needed fixing anyway — brittle pipelines, undocumented systems, years of shortcuts nobody had budget to unwind. That's real, valuable work. It's just not the thing the P&L was told to expect. It's technical debt repayment wearing an AI-transformation costume, and it partly explains why adoption keeps climbing while earnings impact stays flat — a lot of the "AI work" happening was never going to show up as revenue or margin, because it was never trying to.

The second failure mode is worse, and it's the one I actually lose sleep over: we build something technically impressive that automates a workflow nobody needed automated, while the actual constraint on the business — the thing genuinely capping growth — goes untouched. Fast, cheap, well-engineered, and irrelevant. That's a worse outcome than a clumsy solution to the right problem, because it consumes the same budget, the same trust, and the same one shot at proving the technology's worth internally.

And beneath both of those: FOMO

Even worse than either failure mode is the reason a lot of these projects get greenlit in the first place: competitors are doing something with AI, so we need to be doing something with AI, and the something gets picked before the problem does. That's not a leadership flaw or an engineering flaw — it's a shared one. Standing still long enough to ask "is this actually the right problem" feels riskier, in the moment, than shipping something. So everyone ships something.

“The gap between AI adoption and AI impact isn't a talent gap or a model-quality gap. It's a gap between the problem we found exciting and the problem the business actually had.”

The fix isn't distrust. It's asking the question out loud, together

None of this is an argument for CFOs to start micromanaging architecture decisions, or for engineers to stop being excited about what's newly possible. It's an argument for one question to get asked out loud, before scope gets written, by whoever is closest to both sides of the table: what business constraint does this actually remove — and how would we know if it didn't? If nobody can answer that in a sentence, the project needs a different sentence before it needs a better model.

This isn't only a C-suite decision. It's a One Person Company decision

Here's the part that applies whether or not you own the roadmap: you don't need a CEO title to make this call. Every task you choose to automate, every workflow you decide is worth an agent's time, is a bet you're personally accountable for — whether you're an engineer, a PM, a marketer, or running a company of one. Think like a One Person Company (OPC): the ticket closing isn't the win condition. The business constraint actually loosening is. That discipline — asking what this really moves, before asking how well it can be built — is available to you at your desk today, regardless of which track you're on or whether anyone above you is asking it yet.

That's the real leverage in this moment. Not being first to automate something. Being the person in the room — at any level — who insists on naming the right problem before anyone touches the workflow.