Skip to content
Fritz Gerald ZephirinMenu

Social

Run AI Like a Hire

4 min read

When an AI deployment goes wrong in someone else’s company, the postmortem always blames the model. When one goes wrong in mine, I’ve learned to ask the question I’d ask about a struggling employee: who was managing this, and against what job description? The answer, embarrassingly often, is nobody, against nothing.

Here’s the frame that fixed it for us. An AI agent in production is not software in the old sense — deterministic, spec-bounded, done when it compiles. It is a capable, tireless, occasionally overconfident worker with no context about your company beyond what you deliberately gave it. We have a century of hard-won knowledge about managing exactly that profile. Almost none of it is being applied, because the thing arrived through the IT budget instead of the HR one.

So we run AI like a hire. Concretely:

Write the job description before the integration ticket. One page: what this agent is for, what it may decide alone, what it must escalate, how its performance will be measured. If you can’t write the JD, you’re not ready to deploy — you’re hoping. The discipline sounds bureaucratic and is the opposite: a healthy share of proposed agent projects die at the JD stage — which is precisely the failure-prevention working. (The same gate, applied to tools generally, is why the MIT study’s successful five percent embedded narrow tools in defined workflows rather than “rolling out AI.”)

Onboard it. A new hire gets context: the policies, the tone, the clients who are exceptions to everything. For an agent, onboarding is the knowledge base, the prompt architecture, the retrieval scope. An unboarded agent improvises from internet priors — which is what a new hire would do too, if you gave them a desk and no information.

Probation, in shadow mode. No agent of ours touches a customer on day one. It drafts; humans send; we measure the gap between its drafts and what experienced people actually sent. Promotion out of shadow mode is earned with data — resolution parity on the eval set, recontact rates, zero policy violations across a defined volume — exactly like a probation review, except the evidence is complete.

Least privilege, expanded with trust. A new bookkeeper doesn’t get bank signing authority in week one. An agent doesn’t get write access to billing because the demo went well. Capability horizons are doubling on a measured schedule — which means the temptation to over-grant runs ahead of the verification practices that make granting safe. Scope grows when review history justifies it, per task, never “root because it’s easier.”

Reviews, on a calendar. Humans get performance reviews; our agents get them quarterly. Eval scores against the test set built from real tickets, sampled transcript reads (the same audit a human queue gets), error taxonomy, a written verdict: expand scope, hold, retrain, or retire. Model upgrades from vendors are treated as new team members wearing a familiar face — re-evaled before they inherit the predecessor’s permissions.

Offboarding exists. Retiring an agent means revoking credentials, archiving transcripts for audit, documenting what replaced it. The forgotten agent with live API keys is the new forgotten contractor badge.

One honest caution about the metaphor, because it has a failure mode: anthropomorphizing breeds over-trust. The frame is governance, not psychology — the agent doesn’t have intentions, won’t flag its own confusion the way a person might, and fails in alien ways no employee would (confidently, identically, ten thousand times). Run the management structure like it’s a hire; calibrate your trust like it’s a machine. The combination — human-grade supervision, machine-grade skepticism — is the whole trick.

What you get for the overhead is the thing the era’s disappointing statistics say most companies don’t get: an AI deployment whose performance you can actually state, defend, and improve. Management turns out to be the complement that cheap intelligence was waiting for — which is the thesis of this site, shrunk to a checklist.

Sources & further reading