Cost Ledger

Servicing agent · one mid-size insurer

420,000 tasks/month · 6,200 in / 540 out · mid-tier general

• Prompt caching on stable prefixes−34%
  measuredMeasured 30–45% depending on prefix stability. The 60% figure in vendor material did not reproduce on any client pattern we ran.
• Summarised recall instead of raw history−31%
  measuredThe largest single lever measured, and the one nobody markets. Comes free if you have already adopted the memory layer.

  Requires · An episodic memory layer — see the Recall Probe.

• Batch the asynchronous half−12%
  measuredOnly applies to the ~40% of this workload that is not user-facing. Latency-tolerant work only.
• Context pruning before the call−9%
  measuredSmall and fiddly. Included because teams spend disproportionate effort here for a single-digit return.
• Distil the classifier step to a small model−7%
  measuredReal but narrow — only the classification step, which is a small share of this workload's tokens.
• Vendor 'automatic optimisation' tier−5%
  not reproducedClaimed at 25%. We measured 5% on our patterns and could not reproduce the rest. Shown at the measured number.