Decision briefs

Decision briefs that end in an action.

Only changes that can affect model choice, workflow cost, delivery risk, or product pricing belong here. Each brief connects to a scenario and a deterministic calculation.

Illustrative local preview — not current provider data
01PRICE CHANGEAI support

When input and output prices move in different directions

A blended headline can hide whether your retrieval-heavy or generation-heavy workflow actually benefits.

Decision impact
Recalculate by workload mix before switching routes.
Editor view
Price per million is context, not the product decision. Weight the rate by the workflow that produces a successful outcome.
Evidence state
Illustrative editorial pilot · production source review required
Calculate this scenario
02MODEL RELEASERAG search

Long context expands the workflow—and its failure surface

A larger context window may remove retrieval steps, but it can also increase submitted-token cost and retry exposure.

Decision impact
Compare retrieval, long-context, and hybrid paths.
Editor view
Do not value context size in isolation. Model the call count, context actually submitted, and what happens when the answer fails review.
Evidence state
Illustrative editorial pilot · production source review required
Calculate this scenario
03APPLICATION CASEAI support

A support automation case should expose human review cost

Vendor case studies often highlight resolution gains while leaving escalation and review effort outside the headline.

Decision impact
Add human-handoff and failed-task assumptions.
Editor view
Label vendor-provided outcomes and reproduce only the economic assumptions you can defend.
Evidence state
Illustrative case pattern · no vendor claim reproduced
Calculate this scenario
04BENCHMARKCoding agent

A benchmark refresh can invalidate an old recommendation

Task sets, model versions, reasoning budgets, and scaffolds change. Old ranks should not silently become current advice.

Decision impact
Pin the exact evidence version in the decision report.
Editor view
Use a benchmark to narrow questions, then validate your own success criterion and operating cost.
Evidence state
Method note · external values remain link-only
Calculate this scenario
05OPERATIONSCoding agent

Retries can erase a cheaper model's apparent advantage

A low per-call price can lose when failure recovery creates more attempts or additional review steps.

Decision impact
Compare cost per successful task, not per submitted call.
Editor view
Reliability belongs in the denominator. Make the success definition explicit before comparing candidates.
Evidence state
Synthetic scenario · deterministic formula 2.0.0
Calculate this scenario
06PRICINGRAG search

A generous free tier is a unit-economics decision

Free allowance depends on successful-task cost, paid conversion, payment fees, fixed cost, and the target margin.

Decision impact
Calculate a sustainable whole-task allowance.
Editor view
Do not convert a token budget directly into a product promise. Price the customer outcome and preserve a margin buffer.
Evidence state
Synthetic scenario · deterministic formula 2.0.0
Calculate this scenario