AI doesn't need to be smarter. It needs a boss.

About 95% of enterprise AI pilots return nothing, per MIT's NANDA project, after 30 to 40 billion dollars of enterprise investment. And here's the part I keep coming back to: the report itself says that divide is "not...driven by model quality or regulation." It's "determined by approach." The same models writing poetry and acing benchmarks are sitting in those dead pilots. The capability showed up. The return didn't.

So whatever is failing, it isn't intelligence. As I see it, it's a leadership problem wearing a technology costume.

Think about your AI the way you'd think about any team, and stop making the distinction that it's AI. When you deployed it, you basically hired a brilliant junior team. Endless energy, reads everything, types faster than your whole department. Also: no memory of yesterday, no idea what matters to your business, and an overwhelming urge to hand you something plausible and call it done.

Now the question that explains the 95%...who did you assign to RUN that team?

The standard playbook assigns nobody. Buy the tool, send the email, hope it works out. No onboarding, no supervision, no feedback loop. Then the pilot gets graded and the "intern" takes the fall for a job nobody was managing.

I see the same thing at the scale of one shop. I pulled five misses from one working stretch in early September (hand-picked, so it's a tally, not a rate). In three of the five, the catch started with me, not with the AI and not with my automated checks. And there's research on why the seat matters: Kelley and Riedl (2026) found the same models push back when you set them up as an advisor and cave when you set them up as a peer. The role isn't fixed by the model. It's set by whoever sets up the work. That's a boss's job.

Would you turn a new employee loose with zero supervision and then judge the whole department on the results? That's the experiment most pilots are running.

And don't wait on the labs to fill that seat. They can tell you what the next model will BE. They can't tell you who runs it in your shop, because they can't see your shop.

What the job actually looks like is the next post.

So your AI needs a boss. Here's what that job actually looks like, from doing it every day for the last six months.

Recap: the 95% isn't a model problem. Everybody bought the brilliant junior team and assigned nobody to run it. The fix isn't another tool. It's a job, and as far as I can tell it comes down to three duties.

1. Onboarding. Every. Single. Time. Your AI wakes up brand new every morning, so whatever it needs to know has to be LOADED before the work starts: what the task is, what the constraints are, what right looks like, what it got wrong last time. Not a wiki it's supposed to remember to go read. In my shop, rules that get loaded before the AI says a word actually fire. Rules it has to go looking for...don't. Of the three, this one matters most.

2. Supervision, before consequences. The work gets checked at the moment it wants to be done, by someone or something that didn't write it. Here's what surprised me: the popular agent frameworks will let you require a human to approve an ACTION, like issuing a refund. But in OpenAI's and Anthropic's agent SDKs, nothing built in lets you require a human to confirm the work was actually DONE right. (CrewAI has a switch for a person to review the final answer. It's off by default.) That gap is exactly where a boss sits.

3. Feedback that becomes structure. This is the one nobody does. When I correct my team, the correction becomes part of how the next session starts, or a check that refuses that same mistake automatically. A lesson that lives in a document is a lesson the next session never met...because, again, born this morning.

Load it, check it, turn every miss into structure. It's the same leadership you'd give any talented team that's new to your business, which is exactly why this gap is so fixable, and why waiting for a smarter model won't close it. A smarter intern with nobody leading it is still an intern nobody's leading :)

So assign the owner. One person whose actual job is to run the AI like a team. The labs can't do that part for you. It's yours.