Decision-1, then Haiku 5.5, then a person: a three-tier gate

A three-tier gate works like this: Microsoft-Decision-1 answers first and is allowed to say it cannot tell; unsure cases go to Claude Haiku 5.5, which can explain itself in words; anything still unclear goes to a person. Only a clean pass from a tier leads to a paid Sume call. Microsoft's own best-practice list suggests including an abstention option such as cannot tell when the model should not be forced to choose.

Tiers and what each costs you

Haiku 5.5 is priced from $0.10 per million input tokens and $0.50 per million output tokens, per Anthropic's models overview; check its pricing page for long-prompt rates. Its default effort is medium. A decision model returns no text, so the second tier is where a written reason appears.

How to wire the abstention

Give the choice question an explicit fourth option and describe it, for example cannot_tell: the brief is missing a subject, length or ratio. Microsoft notes that scores can move with wording and option order, so test both before you set a threshold. Send tier 2 only the brief and the question, not the Sume key, and ask it for one word plus one sentence. Keep tier 2 optional: if you would rather have a person look at every unsure case, drop it and the cost is a short queue.

Where Sume's gates still apply

After the last tier says go, the call to Sume is the same as without any model. On hosted MCP that means an idempotency_key, optionally a dry_run first, and max_spend_credits if you want a ceiling. The gates are described on the tools and gates page. The tiers reduce bad submits; they do not replace those controls.