Horizontal AI keeps stalling. Narrow AI keeps shipping.

There's a pattern showing up in every serious AI adoption study this year, and it's worth naming plainly: the big, horizontal AI deployments keep stalling, and the narrow, embedded ones keep quietly working.

MIT's NANDA report, covered in Forbes, pegged the failure rate of enterprise generative AI pilots at roughly 95%. That number got a lot of airtime. What got less airtime is *which* 5% worked. According to Healthcare IT News' reporting on the same study, the pilots that produced measurable ROI weren't the flashy horizontal ones. They were narrow, workflow-embedded use cases, usually with a human still in the loop.

The two shapes of an AI project

Horizontal AI is the shape most companies buy first. Roll out a copilot to everyone. Give the whole team a chatbot. Bolt an assistant onto the CRM. It demos well because it can do anything. It fails in production for the same reason: it doesn't do any one thing well enough to change a specific number on a specific report.

Narrow AI is the shape that ships. It reads the invoices coming into one shared inbox and routes them. It drafts the first version of a specific quote using a specific price sheet. It reconciles two systems that never talked to each other. It's boring. It's also measurable, which is why it survives past the pilot phase.

McKinsey's 2025 State of AI survey points in the same direction. The functions reporting real bottom-line impact are the ones where AI is embedded into a redesigned workflow, not laid on top of an existing one. "Redesigned" is doing a lot of work in that sentence. It usually means someone sat with the team, mapped the actual steps, and cut or rewired half of them.

Why the boring version wins

A horizontal copilot has to be judged against expectations that are basically infinite. If it can do anything, users will try everything, and remember the failures.

A narrow automation has a fixed job. Either the invoices got routed correctly or they didn't. Either the draft quote saved the salesperson twenty minutes or it didn't. The scoreboard is honest, which is uncomfortable at first and useful forever.

There's also a governance angle. Narrow tools have a narrow blast radius when they go wrong. A general assistant hallucinating a policy answer to a customer is a much worse Tuesday than a routing script sending an invoice to the wrong folder.

What to do with this

If you're staring at an AI budget and wondering where to point it, the honest advice is: pick one workflow that annoys a specific person on a specific day of the week. Automate that. Measure it for a month. Then pick the next one.

That's less exciting than a company-wide AI strategy deck. It's also, based on the data, the thing that actually works.

If you want a second set of eyes on which workflow to pick first, that's what our free discovery call is for. No pitch, just diagnosis.