Runaway AI bill? We find the loop, prove the fix against your invoice, and get paid only from what it saves.
For companies where the model bill is a line item with its own meeting: support agents at scale, voice agents, agentic pipelines. Read-only, inside your walls, with a receipt.
The problem, specifically
At scale the bill is the sum of agent loops. One support conversation is eight model calls: retrieval, classification, reasoning, tool calls, a summary, an evaluation. The third call re-sends the whole ticket history. The classifier runs on the most expensive model. Caching is off. Nobody can see this from the invoice, and the observability tools show traces, not money-per-step. So the bill grows with volume and nobody can say which step to change.
What we do
1. Read the traces you already have. Langfuse, LangSmith, OpenTelemetry GenAI, Datadog LLM observability. No SDK, no proxy in the request path. The output is cost per conversation, per workflow, per step, ranked — with the dominant step and the shape of the waste named: context growth, heavy-model routing, cold cache, over-retrieval.
2. Propose the change, with assumptions stated. Route the read/extract/classify steps to a lighter model. Replace re-sent history with a running digest. Turn on cache-stable prefixes. Cut retrieval to what is used. Each proposal says what it assumes and what would prove it.
3. Prove it on real traffic. The harness pre-registers the arms, shadows a sample of live requests through the change, and checks parity on your own fitness criteria — accuracy, resolution rate, customer-visible quality — with your judge, not ours. Savings that came from doing less work are not counted.
4. Issue a receipt against the invoice. The baseline is the provider invoice, not a dashboard. The receipt is an open specification; anyone can validate it. What is on the receipt is what we bill against.
Two doors in
Engineering: the trace store, above. Finance: connect Ramp or Brex and see what the company actually pays every AI and compute vendor — last month, growth, and who holds the card. The two views meet on the same page: the invoice line and the loop that produced it.
How it runs
The engine is open source and runs as a CLI in your VPC; nothing leaves. The hosted site is optional. Credentials are never stored by us. Everything is read-only until you apply a change, and every change ships with its undo.
How we charge
A share of verified savings, measured on the receipt, for a fixed term. If the receipt says $0, the invoice is $0. No seat pricing, no platform fee, no charge for the diagnosis.
Typical shape of a first engagement
Week 1: connect the trace store, read two weeks of traces, deliver the ranked loop report with three proposals. Week 2–3: shadow trial on the top proposal, parity judged by your criteria. Week 4: receipt, and a decision. One engineer on your side, part time.
Start by connecting a trace store yourself on the loops page — the read is free and takes a minute — or write to the address on the site footer with the provider and the monthly figure, and we'll reply with what we'd expect to find.