The AI model routing ladder
Model pickers are getting smarter, but the human skill still matters: know when a task needs a quick answer, deeper reasoning, or an expert-grade review.
AI tools are becoming less like a single chatbot and more like a set of modes. OpenAI's model selector explains different choices for quick answers, deeper reasoning, and more advanced work. Its Enterprise model guidance describes Auto, Instant, Thinking, and Pro-style options. Anthropic's Claude model documentation makes a similar distinction across capability, speed, reasoning, and long-context work.
That creates a practical question for knowledge workers: which mode should you use for which job? The answer is not "always use the smartest model." The answer is to route the task based on risk, ambiguity, and consequence.
The skill
The AI model routing ladder is a simple way to decide how much reasoning to request before you start. It keeps lightweight tasks fast and makes higher-risk tasks slower, more deliberate, and easier to review.
AI model routing ladder
Level 1: Fast
Use for: rewriting, formatting, brainstorming, simple summaries, low-risk drafts
Prompt style: "Give me a quick version."
Level 2: Focused
Use for: structured analysis, comparison, planning, meeting prep, source review
Prompt style: "Work step by step and show assumptions."
Level 3: Thinking
Use for: ambiguous decisions, tradeoffs, technical tasks, financial or legal-adjacent review, strategy
Prompt style: "Think carefully, test alternatives, and state uncertainty."
Level 4: Expert review
Use for: high-consequence work, executive recommendations, public content, customer promises, irreversible actions
Prompt style: "Review like a senior expert. Look for flaws, missing evidence, and failure modes."A worked example: customer churn analysis
If you ask for "ideas to reduce churn," a fast model may be enough. If you ask whether the business should change pricing, that deserves a higher rung.
Task:
Analyze whether our recent churn spike is caused by onboarding, pricing, product gaps, or support delays.
Routing:
Level 3: Thinking
Why:
- Multiple possible causes
- Needs evidence from several sources
- Could affect customer communication and pricing decisions
Instructions:
- Compare at least three hypotheses
- State what evidence would support or weaken each one
- Separate facts from guesses
- End with confidence level and follow-up data needed
Do not:
- Recommend a pricing change without stronger evidence
- Write customer-facing messaging yetThe prompt
Use this before starting a task when you are unsure how much reasoning it needs:
Before answering, route this task.
Task:
{what I need}
Choose one:
1. Fast
2. Focused
3. Thinking
4. Expert review
Explain the choice in one sentence based on:
- Risk
- Ambiguity
- Consequence
- Need for evidence
Then answer using the chosen level.
If the task could affect customers, money, legal/compliance, public content, private data, or irreversible actions, choose Thinking or Expert review.The review checklist
- Risk: What happens if the answer is wrong?
- Ambiguity: Is there one clear answer, or several plausible paths?
- Evidence: Does the task need sources, files, data, or citations?
- Consequence: Will the answer influence money, customers, policy, or public messaging?
- Review: Does the output need a second pass for gaps and failure modes?
Why it works
Using a powerful model for every tiny task wastes time. Using a fast answer for high-consequence work creates quiet risk. The routing ladder gives you a lightweight judgment step before the model starts working.
It also improves prompts. When you say "think carefully" only for tasks that deserve it, your AI workflow becomes more deliberate: fast where speed matters, careful where correctness matters.