Listen on
Transcript
Not every task needs the strongest model. A simple question can use a smaller, faster route. Good AI workflow means matching the tool to the size of the job.
Full transcript of Briefing 03, 0:15, published 3 August 2026.
Key points
- Capability and suitability are different questions
- Smaller models answer routine questions faster and more cheaply
- Routing is a design decision, not a downgrade
- Reserve the largest model for work that genuinely needs it
The default that costs the most
The simplest architecture is to send everything to the most capable model available. It removes a decision, and for a prototype that is the right trade. At production volume it becomes the single largest avoidable cost in most deployments, and it makes every routine request slower than it needs to be.
The alternative is not to accept worse answers. It is to notice that a large share of requests are classification, extraction, formatting or lookup — work where a smaller model is not merely adequate but often preferable, because latency matters more than depth.
Routing as a first-class design choice
Treating model selection as part of the workflow design means deciding, deliberately, which classes of request go where, and what happens when a small model is unsure. A well-built escalation path — cheap first, capable on demand — usually outperforms both extremes.
It also has a governance benefit. Once routing is explicit, an organisation can say which work is being handled by which system, which is difficult to reconstruct after the fact when everything goes to one endpoint.
Frequently asked questions
Why not use the most capable model for everything?
Because much routine work does not benefit from the extra capability, and paying for it adds cost and latency to every request. Capability is only useful where the task requires it.
What kind of work suits a smaller model?
Classification, extraction, routing, formatting and short factual lookups — high-volume tasks with narrow scope, where speed is usually worth more than depth.
How do you avoid quality problems when routing?
By designing an escalation path: handle the request cheaply first, and pass it to a more capable model when confidence is low or the task turns out to be harder than expected.