Skip to main content

Boardroom Answers · AI & Data · AI, Data & Analytics

LLM bills are where automation programmes go to die. Walk me through your cost controls — caching, model tiering, budgets — with specifics, not adjectives.?

The question a Chief Automation Officer (CAO) asks.

The short answer

Three-tier model routing, two layers of caching with the savings metered, spend allowances and rate limits that fail closed, and a ledger that prices every attempt including the discarded ones — LLM cost as engineering, not weather.

The full executive answer

Specifics, in the order they save money. Model tiering: every module is routed by declared complexity — the deepest synthesis work runs on Claude Sonnet with an Opus option on extended budgets, standard analysis on Sonnet, and short phrasing and classification on Haiku, which is a fraction of the cost — so quality is spent where it counts and nowhere else, all reconfigurable without code changes. Response caching: identical inputs return the stored validated answer at zero model cost, and cache hits are recorded as zero-cost events so the savings are visible in dashboards rather than invisible. Prompt caching underneath: the large fixed portion of each prompt — identity, task template, schema, examples — is marked for Anthropic’s cache, so repeated generations re-bill that prefix at the reduced cached rate; that is the single biggest lever on input cost, and we track cache-read and cache-write tokens separately in the ledger to prove it is working.

Then the guardrails on runaway spend, which fail closed: per-tier monthly AI allowances — on the order of a few dollars on trial up to around a hundred and fifty on the top self-serve tier, custom for enterprise — enforced before each paid call; a per-organisation rate limit of thirty paid generations per five minutes that denies on infrastructure error rather than waving traffic through; and a monthly budget ceiling that refuses generation when the ledger is unreadable, because an unmeterable call is an unaffordable call. Every attempt is accounted — including the embarrassing ones: discarded corrective retries, consensus second opinions and adjudicator calls each land in the cost ledger, so our margin math has no hidden column. The honest trade-off on the table: consensus and adjudication multiply cost on flagship modules by design, and they are configuration-disablable under cost pressure — a documented dial, not an accident.

Grounded in: FinOps Foundation practice applied to LLM spend (unit economics, allocation, visibility); ISO/IEC 42001 resource-management controls; provider prompt-caching economics (Anthropic cached-token pricing).

Want this answered live, on your data?