Token Cost Margin Audit (Delivery)
A checklist with 7 steps: Pull 30 days of per-client token spend before quoting any new AI retainer.
By InnovaAI ResearchPublished
What are the steps?
Token Cost Margin Audit (Delivery)
- 01
Pull 30 days of per-client token spend before quoting any new AI retainer
Export usage by model and endpoint from your gateway logs, then split input versus output tokens. Output tokens typically cost several times more than input, so a blended per-call average hides where the margin actually leaks.
- 02
Rank the top five highest-volume client workflows by cost per completed task
Measure cost per finished deliverable, not cost per API call. A workflow that runs 40 cheap calls to produce one report can outspend a single expensive call that ships the same output.
- 03
Strip full-document attachments and long conversation history from prompts that tolerate compression
Replace pasted source documents with a short summary plus retrieval of the relevant passages. Token trimming of this kind can cut per-task cost by half or more without changing output quality.
- 04
Set a hard monthly spend ceiling per client account at the gateway layer
Route traffic through a proxy that enforces budget caps and rate limits, so a runaway agent loop cannot burn a retainer's margin overnight. Helicone and Portkey both expose per-key spend controls for this purpose.
- 05
Model the same workload across at least two providers before locking pricing
Price the identical prompt set against a second vendor and a smaller model tier. A cheaper model that passes your quality bar on classification or extraction tasks can carry most of the volume while a frontier model handles only the hard cases.
- 06
Write the token budget into the client statement of work as a named line item
State the included volume, the overage rate, and what triggers a re-quote. Agencies that bill AI work as an opaque flat fee absorb every provider price change themselves.
- 07
Re-run the audit quarterly and whenever a provider changes pricing
Provider pricing and model lineups shift faster than annual contract cycles. Schedule the review so a cost change surfaces before the client invoice does.
More for AI Infrastructure
- Operating ProceduresAlgolia Client Workspace Setup (Onboarding)
- Operating ProceduresDigitalOcean GPU Droplet Deployment for Client AI Workloads (Delivery)
- Operating ProceduresIQ Routing Client Cost Optimization Setup (Onboarding)
- Operating ProceduresCoral Bricks Client Agent Endpoint Handoff (Onboarding)