When Should Your AI Use a Stronger Model?

A customer email asks whether you serve a particular neighbourhood. Another asks for a custom quote involving three locations, an unusual deadline, and terms that conflict with your standard agreement. Should both requests go through the same AI model?

That question matters more than the name of this month's top model. On November 9, Anthropic released Claude Fable 5.2 for demanding technical work. Its own guidance recommends starting most coding and knowledge tasks with a different everyday model, then moving to a more capable or more economical model as the work requires.

Give routine work a routine path. Define when the work should escalate.

Try the five-minute escalation exercise

Choose one task your team repeats every week: replying to enquiries, reviewing a proposal, preparing a report, or sorting support requests. Then answer four questions.

1. What does an ordinary request look like? Write down the information needed to complete it. For a service business, that might be the customer's location, requested service, and preferred date.
2. What makes the request difficult? Look for missing information, conflicting instructions, several documents that must be reconciled, an unfamiliar situation, or a decision with financial consequences.
3. What can AI do without approval? It might classify the request, find information in approved material, or draft a reply. Decide separately whether it can send the reply, promise a price, or change a customer record.
4. How will you know escalation helped? Measure accepted results, staff correction time, response time, and total cost. A stronger model earns its place when it improves an outcome enough to justify the extra effort.

Copy this decision card into a note for your team: Routine request → first response → escalation trigger → person who approves → result to measure. You can complete it in a few minutes without buying new software.

Download the one-page AI Escalation Decision Card ↓
A printable worksheet for choosing one task and defining its routine path, escalation trigger, and approval point. No email required.

OpenAI's model-selection guidance also recommends comparing models on representative work and keeping the lightest setting that meets your quality bar. Its deployment guidance recommends measuring cost per successful task, rather than model cost alone. For a broader comparison of models and constraints, see our business guide to choosing an AI model.

What an escalation rule could look like

Consider a small commercial cleaning company handling incoming enquiries:

• Standard service-area question: Check approved service-area information and draft a reply. Escalate if the location or service is unclear.
• Routine request for a quote: Collect square footage, schedule, and contact details. Escalate if requirements are missing or outside standard packages.
• Three sites, a next-day deadline, and attached terms: Organize the request for review. A person assesses pricing, capacity, and contractual commitments.

The rule does not have to start with an elaborate system. A team member can apply it manually while testing the workflow. Once the pattern is clear, the business can decide whether software should route requests automatically.

Test the rule on work you already know

Take ten recent, anonymized requests. Mark which were routine, which needed deeper analysis, and which needed a person's decision. Run the same requests through your proposed workflow and compare the results with what actually happened.

Pay attention to exceptions. If the AI handles nine easy requests well but confidently mishandles the tenth request containing a contractual change, the average score hides the problem. Likewise, if every request goes to the strongest model or to a person, the escalation rule may add little value.

The goal is a workflow your team can explain: “We use AI to prepare standard responses. We ask for deeper analysis when the information is incomplete or contradictory. A person approves prices, commitments, and unusual terms.”

That is a more useful operating decision than simply asking whether the newest model is better.