You're not hiring AI.
You're hiring judgement.

The interesting question isn't which LLM — it's which decisions you're willing to delegate, and to whom.

The wrong question.

Every enterprise "AI strategy" meeting starts with the wrong question. Which model? Which vendor? On-prem or cloud? Fine-tune or prompt? These are procurement questions dressed up as strategy. They are decided in the last five minutes of the project, not the first.

The right question is older and harder. Which decisions, in which workflows, are you willing to delegate — and what quality of judgement are you willing to accept in exchange for what quantity of throughput?

Judgement is the unit.

A support agent that replies to 400 tickets a day is not "using AI." It is exercising 400 instances of delegated judgement. Each one: what is this person asking, how urgent is it, what is the precedent, what is the tone, when to escalate. A human does this implicitly. An agent does it explicitly — and the quality of the agent is the quality of the explicit reasoning you've encoded.

The hiring test.

When I scope an agent workflow, I run what I call the hiring test: describe the agent's role as if you were writing the job posting for a human. If the job posting is vague — "handles AI-powered automation" — the agent will be vague and the results will be bad. If the job posting is precise — "reads inbound tickets, classifies them against a 14-row taxonomy, drafts a reply using the brand voice doc, flags anything outside the taxonomy for human review" — the agent has a fighting chance.

This is why the model matters less than the people assume. GPT, Claude, Qwen, Gemini, Llama — any of them can do the job if the job is specified well. None of them can do it if the job is specified poorly. The hiring process is where the work is.

Four kinds of judgement, four deployment patterns.

• 01
  Low-stakes, high-volume, internal. Ticket triage, content classification, form extraction. Let the agent run, because nothing it does leaves the building. Sample a share of its work for human review.
• 02
  Medium-stakes, medium-volume. Drafting replies, writing first-pass reports, summarising meetings. Deploy in co-pilot mode. Human approves before sending.
• 03
  High-stakes, low-volume. Hiring decisions, pricing changes, legal interpretations. Don't deploy. Use the agent to prepare the brief for a human, never to replace the human.
• 04
  Adversarial. Anything where a user might try to game the agent — refunds, fraud flags, contract generation. Needs a separate adversarial eval loop, and a human approving the outcome. It is the step teams are most tempted to skip.

Evals are the interview.

You would not hire a senior operator without a working interview. You would not promote one without performance reviews. AI agents are the same. An eval suite is not a nice-to-have — it is the interview loop and the quarterly review, encoded as software.

If you are about to deploy an agent and your team can't show you 50 labelled examples of the kinds of decisions it will face — 25 easy, 20 hard, 5 adversarial — you are not ready to deploy. Build the evals first. They will tell you more about the problem than the model ever will.

The quiet shift.

The teams winning with AI right now are not the ones with the most models in production. They are the ones who have become disciplined about delegating judgement. Small eval suites, narrow job descriptions, shadow-mode deployment, supervised rollout, regular review. It is the least glamorous part of the AI wave, and it is the part that separates teams that ship from teams that demo.

Stop asking which model. Start asking which decision. The answer to the first will follow from the answer to the second, every single time.