Give Your A-Team a Team: An Interview on AI and Work

Watch the interview
Ryan York with Berg Atkinson · 42 minutes · on YouTube

1The question underneath everything

1:12 I opened and closed on the same question: how do we use this, really? Not which model is smartest, but how an organization actually puts it to work. It is the question behind a whole series I've been writing.

More on this: How Do We Use This?, the working paper, and the series built from it.

2What these systems are, and what they aren't

5:25 A lot of the stories about AI going rogue describe a system working exactly as it was built. What people call "the AI" is a model wrapped in layers of software, a harness, and there is less AI in it than people think.

More on this: the primers What is an LLM?, What is a harness? and What is “AI”?

3It needs leadership

6:55 Working with it as a team, what I found is that it needs leadership. I work with it the way I work with people, with a lot more handholding. It is your smartest, laziest employee, most of the time.

More on this: AI doesn't need to be smarter. It needs a boss. “Laziest” here means stopping at the first defensible answer (point 5).

4Don't hand a new hire the keys on day one

11:51 When someone tells me their AI ran wild and spent a fortune, my question is whether they would give a new employee the company card and the keys to everything on the first day. Trust gets earned in stages, with people and with this.

The evidence: Who Gets It First (section 5.5): a rollout that stages the checking works, but only when the checks are real. With half of them skipped, the simulated organization settled on a worse way of working 30% of the time. Choosing how much to automate, by type and level, is well studied (Parasuraman, Sheridan and Wickens, 2000).

5It stops at the first answer it can defend

13:13 By default the models stop looking when they reach their first defensible answer, and correctness has nothing to do with when they stop. We all know people who do the same thing, which is why most of the fixes already exist in how we manage people.

More on this: The model stops at the first answer it can confidently defend to you. Not when it's right. [Our position. Training does reward correctness where a checker can see the answer (math, code with tests); the stop in open-ended work is where none can.]

6Give your A-team a team, not a tool

17:17 My first advice to any company is to take its A-team, every company has one, and give them a setup like mine. Give them a team, don't give them a tool. And don't lay off your senior people, because they are the ones who can teach the juniors how to lead.

The evidence, in the model: Who Gets It First (sections 5.2 to 5.4), run after this interview. In a simulated 200-person organization, giving a new tool to the ten sharpest people first left it on a worse way of working 1.8% of the time, against 2.8% for a staged rollout (a difference within the noise), with about a sixth of the checking; ten people chosen at random, with the same checks, 12.8%; and the 25 team leads going first without checks, 57.8%, against 68.8% for everyone at once (a comparison that changes both the group and the checking). The reason is that a check is only as good as the person doing it: with five tries, the sharpest ten let 17% of worse options through, ten random members 33%. Built into the model: the sharpest are the ten who read results with the least noise, they can always be found, and everyone else waits 100 rounds (see the figure), so the plan pays off where careful checks are expensive. [Keep your senior people: practice view; the model has no seniority in it.]

7Build the capability, don't buy the product

18:19 I have bought very few tools, because a team set up this way can build the capability it needs in minutes. Products move too fast to be worth selling, and anyone can build the capability they need if they put their senior people on it.

More on this: Rent the pipes. Own the judgment., which is about renting infrastructure, not building every product. [From one practice.]

8Humans with experience, leading it

24:30 When MIT asked whether AI could help build a jet engine, the answer was yes, with experienced humans leading it. Ten or twenty years of experience is the part nobody can teach quickly, and it is what makes someone able to lead the machine.

The evidence: MIT News on the JARVIS Challenge (July 14, 2026), in which teams of MIT undergraduates designed, built and tested a small jet engine in four weeks with AI as their main engineering partner, and found engineering judgment to be the deciding factor. [Ten or twenty years of experience: practice view; the teams were undergraduates. Against it: AI sped up early design but struggled with physical constraints, and the winning team had resisted using it.]

9Rules don't work. Boundaries and incentives do.

26:09 People are surprised that their AI didn't follow their rules. Once you understand how it works, you know you can't just give it a list of rules. You give it guardrails, and you almost have to incentivize it to make good choices, the way you would an employee.

More on this: Rulebooks are theater. Boundaries and incentives are not. [Our position, from one practice. A model does not change while it works, so “incentivize” means harness design, such as accepting work only once a check passes.]

10Every handoff is a game of telephone

28:26 Context rot is a real problem, and when one agent hands work to another it hands over its own summary of what matters, not yours. The deeper the chain, the more the context mutates, which is one reason I stay in the middle of it.

More on this: the primer What is context? [Two separate effects: context rot, a model doing worse as one context grows (Hong, Troynikov and Huber, 2025), and handoff loss, what a summary drops.]

11Measure the return, starting before you begin

33:00 It is easy to measure what you spend on AI and hard to know what it returns. You can't give a real number if you aren't measuring anything, so measure before and after, and measure more than story points.

More on this: the proof page, which shows how this site's figures are modeled, not measured before and after, and the live totals on the home page, recompiled nightly.

12Is your human in the loop, or watching it?

38:27 Everyone says they have a human in the loop. Hardly anyone measures how good a job that human is doing, and I'd bet most of them are watching the loop rather than in it. Being in it takes experience, which is another reason to keep your senior people.

The evidence: Never let your AI grade its own homework and Your corrections are the most valuable data in your company. In Who Gets It First, a staged rollout whose checks turn into approvals was modeled as one with checks skipped: with half, 80% and 95% skipped it ended on a worse way of working 30%, 51.7% and 64.8% of the time, the last close to no plan at all (68.8%). Research on automation bias finds that experts and teams are not immune to it (Parasuraman and Manzey, 2010), which is why the checking has to be measured.

 The recording, checked

I recorded the interview on 20 August 2026. I went back through the recording in early October 2026 and checked it two ways: what I said from memory against the record as it stood on 20 August, and whether the two arguments I leaned on hold up when tested. The first is here. The second is points 4 and 6 above, who gets the tool first and earning trust in stages.

What I said, against the record

Seven things I said from memory, each checked against the record as it stood on 20 August 2026. Three matched and four did not.