The Nine Positions: Who Owns What When AI Joins the Team

The refund went out for $4,200 on a $40 order. The AI agent read a decimal wrong, approved it, and moved on to the next ticket before anyone noticed. In the post-mortem, six people sat around a table and each one pointed at the machine. "The model did it." Everyone nodded. Nobody's name was on the incident.

That nod is the problem. Not the decimal.

The accountability vacuum

Here's the pattern, and it has a name: the accountability vacuum. It forms the moment you add AI to a workflow and assume the software absorbed the responsibility along with the task. It didn't. The machine took the typing. The accountability stayed exactly where it was, unassigned, floating over a table of people who all believe someone else has it.

When a human does a job, a stack of accountabilities rides along invisibly. Someone can say no. Someone knows why the rule exists. Someone smells when the output is wrong. You never wrote those down because one person carried the whole stack in their head. Automate the task and the stack doesn't vanish. It scatters. If you don't catch each piece and hand it to a named person, it hits the floor.

WISER calls those pieces the Nine Positions. They're not job titles. They're accountabilities, and each one has to be owned by a specific human the day AI joins the work. Let's run the refund agent through all nine and watch where the $4,200 actually came from.

Walking one agent through the nine

Start with the obvious one. Authority. Who can say yes, who can say no, and who can hit the stop button when the agent goes sideways. In our refund team, this was supposedly the ops manager, except she'd never been told she owned the kill switch, and there wasn't one. An agent approving refunds with no ceiling and no human veto is Authority left on the floor. That's incident number one, already waiting.

Next, the money. Stewardship owns the resources and the cost. Someone should have looked at the agent's spend the way they'd look at a junior employee's approval limit. Ours had no limit. A person owning Stewardship would have capped auto-approvals at the average order value and flagged anything above it. The $4,200 clears a $40-order threshold by two orders of magnitude. Stewardship, unowned, is why nobody caught it in flight.

Now the quiet one that kills more deployments than any other. Skepticism. This is the person whose job is to professionally distrust the output and hunt for the silent failure, the error that doesn't throw an error. The agent didn't crash. It felt confident. It was confidently wrong, which is the worst kind, and no one owned the task of not believing it. Skepticism is the Position leaders skip most, because trusting the shiny new system feels like progress. It isn't. It's how the decimal slips through.

Keep going. Translation turns messy reality into terms the system can act on, and back again. Customers don't write clean refund requests. They write "charge looks wrong, fix it," and someone has to own how that becomes a structured input the agent won't misread. Context holds the why and the history: this account is a known chargeback risk, that promo code changes the math, this policy exists because of a lawsuit in 2019. Strip Context out and the agent optimizes against rules it doesn't understand.

Then the two everyone forgets because they're about people, not process. Empathy represents the human on the other end, the customer who's now waiting three extra days because the whole queue froze during the cleanup. Safety guards against harm and the edge cases, the refund pattern that looks like fraud, the input crafted to trip the model. Neither shows up on a metrics dashboard. Both show up in the incident.

Execution is the one that usually does have a name, because someone has to run the thing day to day, watch the queue, restart it when it hangs. Teams point at Execution and think they've covered accountability. They've covered one Position out of nine.

Last, the one that decides whether you can trust the whole arrangement. Integrity keeps the system honest and makes sure it doesn't deceive, doesn't quietly paper over its own mistakes or present a guess as a fact. When the agent approved that refund, did it log the decision truthfully, or bury it in a batch? Integrity owns that answer.

Count the failures in our story. Authority, Stewardship, Skepticism: unowned. The other six, half-owned at best. This is the shape of nearly every deployment that blows up. Not a bad model. Three or four Positions with no name attached.

A person is accountable when AI causes harm, not the algorithm. Every deployment needs a named human behind the objective. That's not a compliance line. It's the whole reason the refund had an author and the algorithm never could.

Before you deploy anything

Put a name next to each of the nine. Not a team, not "ops," a person who knows they own it. Some names repeat; one human can hold three Positions on a small team, and that's fine as long as they know they're holding them. What's not fine is a blank.

Run the list before go-live, then run it again after, because Positions drift when people change roles and nobody re-hands the accountability. An unowned Position isn't a gap in a chart. It's the exact place your next incident is already sitting, waiting for a decimal.

The model didn't do it. Nobody did. That was the whole trouble.