Assign Models Like Staff

You would not hire one person to run sales, accounting, product, and legal.

So why do you keep asking one AI model to do everything?

Most teams land on a favorite model and hand it every assignment: drafting, editing, summarizing, critiquing. That is like asking your best salesperson to close the books. They might try. The result is still a mess.

Some follow instructions tightly and write in your voice on the first try. Others are better critics. Some are fast and cheap enough to run routine work all day. Others are slow and worth the wait when the stakes are high.

This is not a technology problem. It is a staffing problem.

You already know how to solve a staffing problem.

Part 1: Teach AI Your Voice Once, Use It Forever introduced the starter version of this idea. This is the full version: the scoring system, the assignment decisions, the economics of fast versus deep, and a retesting cadence that keeps your decisions honest as models change.

The 60-second version

1. Evaluate on real work, not demos. Run the same five assignments on each candidate model and score the results.
2. Assign by strength. Let the best drafter draft, the best critic criticize, the fastest model handle routine jobs.
3. Pair a drafter with a critic. Two models in sequence, with a human making the final calls.
4. Route by stakes. Fast model for routine transformations, deep model for consequential work.
5. Retest quarterly. Rankings age like produce, not stone.

Stop asking one model to do everything

You would not ask your CFO to write the ad campaign. You would not ask your top salesperson to audit the books. Each hire goes where the evidence says they belong.

One chat window is convenient. Every product defaults to a single model. But convenience is not a staffing strategy.

Test each model on the work you actually do. Keep a scorecard. Assign each job to the model that proved itself on that job.

Evaluate models on practical dimensions

Forget benchmark scores. Evaluate on what you can observe in real work:

• Instruction following: Does it obey the brief without reminders? Check word count, format, and constraints on the first attempt.
• Tone control: Does it match your voice without drifting into imitation? Compare the output to an approved example.
• Reasoning: Can it find the real decision inside a messy request? Give it a brief with a buried lede and see what it surfaces.
• Editing discipline: Can it improve a draft without rewriting everything? A good editor touches what is broken and leaves what works.
• Context handling: Can it use your project files without losing constraints? Feed it the voice profile, then check the draft.
• Tool access: Can it work with your approved directories and documents? Some models read connected files well; others ignore them.
• Speed: Does the response time fit the job? Slow output on routine work wastes hours across a team.
• Reliability: Is it consistently useful across several attempts? One brilliant draft and three bad ones is a bad hire.

No instruments needed. Watch each model do each job, and write down what you see.

Run the 5 assignment test set

The test set is the job interview. Five assignments from real weekly work:

1. Draft a 200 word post from a short brief.
2. Turn a page of meeting notes into an executive summary.
3. Revise a stiff paragraph in plain language without changing its meaning.
4. Explain a complex idea to a nontechnical leader in plain terms.
5. Critique a draft against your quality checklist.

Give every model the same brief, files, and instructions. Score each result from one to five on each dimension:

Model: ____________________   Date: ____________________

Score each assignment 1 to 5 on: accuracy, voice match, clarity,
usefulness, editing required (5 = almost none needed), constraint compliance.

1. Draft post (200 words):        ___ ___ ___ ___ ___ ___
2. Executive summary:             ___ ___ ___ ___ ___ ___
3. Revise paragraph:              ___ ___ ___ ___ ___ ___
4. Explain to nontechnical lead:  ___ ___ ___ ___ ___ ___
5. Critique against checklist:    ___ ___ ___ ___ ___ ___

Totals:
  Drafting skill (assignments 1, 3) .. ___ / 60
  Thinking skill (assignments 2, 4) .. ___ / 60
  Reviewing skill (assignment 5) ..... ___ / 30

The totals tell you who to hire for what. The best drafter may be a mediocre critic. That is why you keep a staff.

Use the two model drafter and critic workflow

For work that matters, never let one model finish the job alone.

Model A drafts

Give the first model the brief, your project files, and one extra instruction: list any factual gaps it could not fill from the material provided.

A drafter that admits what it does not know beats one that invents answers. Read the gaps list first.

Model B critiques

Give a different model the draft and this prompt:

Review this draft as a skeptical editor. Identify:
1. Unsupported claims, facts, or numbers that have no source.
2. Generic language that could appear in any company's content.
3. Voice mismatches: places where the tone breaks from the voice profile.
4. Missing business implications: facts stated with no explanation of what a
   decision maker should do about them.
Do not rewrite the draft yet. Return a numbered list, most serious first.

A model that wrote the draft is the worst person to judge it, just like a writer who falls in love with their own opening paragraph.

A human revises

Read the critique. Accept what is right. Reject what is wrong. A critic can detect problems and still write with the personality of wet cardboard.

The friction is the point: the drafter builds, the critic attacks, the human decides, and the result is a draft you can defend in a meeting.

Route by stakes, not convenience

Not every assignment deserves your best and most expensive model. This is a budget decision.

• Fast model: Routine transformations. Reformatting, summarizing internal notes, extracting action items, first pass drafts of low risk content. Work a human reviews anyway.
• Deep model: Consequential work. Anything a client, investor, or employee will read. Anything with numbers, claims, or commitments. Work where the cost of a bad draft exceeds the cost of a slow one.

Rough awareness is enough: the fast model costs pennies, the deep model costs meaningfully more, and running the deep model on routine work is like flying the company jet to a lunch meeting.

The relegation rule: if the fast model handles a job well three times in a row, it owns that job. If it starts fumbling, it loses the assignment. Assignments are earned, not inherited.

How often should you retest your models?

Models change under you. Your best drafter in January can be your third best by April. The test set is impartial; it does not remember which model you like.

So rerun the same five assignments every quarter. Same briefs, same files, same scorecard. When the results change, update the assignments. No announcement needed. The scorecard is the authority.

What this creates for a business

• Better output per dollar: Routine work runs on the fast model, consequential work gets the deep model, and the budget reflects the stakes.
• No single model dependency: When one model degrades or gets replaced, scored alternatives are ready. Your system survives the vendor's roadmap.
• Decisions based on evidence: Assignments come from the scorecard, not demos or habit. You can explain why each model does each job.
• Institutional judgment: The test set and scores become company knowledge. New team members inherit a working system, not a favorite chatbot.

Not more AI. Better staffed AI.

Frequently asked questions

Final checklist

Before you assign models like staff, confirm:

• The five assignments reflect your real weekly work, not hypothetical exercises.
• Every candidate model gets the same brief, files, and instructions.
• A human reads the actual output and fills out the scorecard.
• Assignments match proven strength, not habit or brand loyalty.
• The drafter and critic are different models for work that matters; a human reviews every gaps list and critique.
• Fast and deep routing reflects stakes, not convenience, and the relegation rule is posted where the team can see it.
• Quarterly retesting is on the calendar with a named owner.
• No client names, internal numbers, or deal terms appear in any test material.

Then start hiring.

Your first scorecard will surprise you. That surprise is the whole point. You were guessing before. Now you have evidence.