FDE Times: Explaining the LLM bill to a CFO — price each outcome, not each call
A 4 November FDE Times guide shows why CFOs need cost per completed outcome (total spend ÷ successful work), not cost per call: a worked claims example turns ~1.8¢/call into ~7.5¢/outcome once retries, cache writes and failed cases stay in the numerator, then asks for three assumption-backed scenarios reconciled monthly.
FDE Times (4 November 2026) argues that the finance conversation after a pilot is an FDE skill, not an account-manager afterthought. CFOs already understand unit economics; token counts and public price lists do not map to “a claim processed” or “a ticket closed.”
Drawing on FinOps framing from Finout and CloudNuro, the piece separates cost per inference (spend ÷ model calls) from cost per outcome (allocated spend ÷ work that actually finished). Failures and retries still cost money, so they stay in the numerator and simply do not count as outcomes. A worked insurance-claims example with Anthropic-style prompt-caching multipliers shows how ~USD 601 for 33,000 calls (~1.8¢/call) becomes ~7.5¢ per fully processed claim when only 8,000 of 10,000 claims clear without human intervention.
Cache-hit rate, retry rate and success rate dominate the margin of error—especially when night-time traffic lets a 5-minute cache cool, or when reasoning tokens hit high per-call ceilings. The practical ask: open with a base and bad-case cents-per-outcome figure, state which assumption drives the bad case, put tokens in an appendix, and agree a monthly forecast-vs-actual reconciliation.
Source: fdetimes.net