Services
Four engagements. One artifact each.
Each engagement ends with one thing: instrumentation running in your stack, or a model finance can open and audit. No decks. Pick by where you are: before the build, after the first bill, or once the numbers need watching.
01 · Audit
Spend attribution audit
Start here if you have a token bill and no workflow names on it.
We instrument up to five workflows in your stack, tagging every model call with workflow, task, and outcome. Then we attribute one month of token spend to those tasks, retries and regenerations included, and capture the minutes people spend reviewing and reworking what the model produced. The result is a cost per task and a cost per accepted output for each workflow, computed from your own logs.
- Duration
- 3–4 weeksIndicative. Depends on how many workflows are in scope and how far the logs sit from the task.
- Artifact
- Attribution report per workflowCost per task, cost per accepted output, and retry share for each workflow. The instrumentation stays running in your stack.
- You need to provide
- Log and review-tool accessRead access to the LLM gateway or client-library logs, read access to the review tool, and one engineer who knows where the calls are made.
- Not included
- Prompt changes, model changes, or vendor negotiation.
02 · Model
Unit-economics model
For a workflow whose spend is attributed and whose payback is still an argument.
We set a human baseline for each workflow from sampled time studies, historical throughput, and the fully loaded hourly cost your finance team already uses. Against that baseline we produce the four numbers per workflow: cost per accepted output, payback, breakeven acceptance rate, and breakeven token price. Each comes with its sensitivity to price, volume, and quality, and a written recommendation to keep, fix, or stop.
Some workflows get “stop”. We write that down too.
- Duration
- 2–3 weeksIndicative. Runs after the audit, or alongside it when the attribution already exists.
- Artifact
- The model, every formula visibleA spreadsheet-portable model finance can open and audit. The formulas are the same ones as the public Token ROI model, so anyone can check the arithmetic.
- You need to provide
- Attribution data and an hourly rateThe attributed spend from the audit or your own equivalent, access to the people doing the task today for a sampled time study, and the fully loaded hourly cost from finance.
- Not included
- Building or changing the workflow itself.
03 · Monitoring
Ongoing monitoring
For workflows that have a verdict and a price list that keeps changing.
Prices, models, and volumes change every month, and a verdict from March is not a verdict in September. We re-attribute spend monthly for every instrumented workflow, alert you when one crosses its breakeven acceptance rate or its breakeven token price, and reissue the keep, fix, or stop verdict each quarter. When a workflow stops paying, you find out from the ledger, not from the invoice.
- Duration
- Monthly · 3-month minimumIndicative. A retainer; the first re-attribution runs the month after the audit or the model completes.
- Artifact
- Monthly one-page ledger per workflowPlus a standing dashboard in your stack, built on the instrumentation from the audit. The alert thresholds are the breakevens from the model.
- You need to provide
- Standing access and rate changesContinued read access to the gateway logs and the review tool, and someone in finance who tells us when a contracted rate changes.
- Not included
- Incident response or on-call for the workflows themselves.
04 · Pilot design
Pilot design
For a workflow that does not exist yet and a budget that has to prove it.
Before you build, we pick the workflow with the best odds, define what counts as an accepted output, set the human baseline, and size the pilot so it can actually prove or disprove payback. The success threshold is written down before the first call is made, so the result cannot be reinterpreted afterwards. Most pilots fail on review time, not on token price. This is where that gets designed out, or found early.
- Duration
- 1–2 weeksIndicative. Runs before the build. Can be folded into an audit when one is already planned.
- Artifact
- Pilot spec with a pre-registered success thresholdThe workflow, the acceptance definition, the baseline, the sample size, and the number the pilot has to beat. Run the candidate through the calculator first to see whether it is worth the two weeks.
- You need to provide
- A shortlist and the people doing the taskCandidate workflows, access to the people doing the task today, and the rates you expect to pay for tokens.
- Not included
- Running the pilot or building the workflow.
Terms
The same terms, every engagement.
These do not change with the size of the engagement. They are what makes the number reproducible.
Included
What every engagement includes
- No proxy in your request path. We tag calls inside your code and read your logs. Nothing sits between you and the model.
- Your data stays in your stack. Logs, prompts, and outputs stay in your warehouse under your access controls. We leave with a report, not a dataset.
- Formulas you can audit. Every figure in the report traces to a formula on the sheet and a query against your logs. Finance can rerun both.
- A written verdict, including “stop”. Each workflow gets keep, fix, or stop, in writing, with the numbers that led there.
Excluded
What we do not do
- Resell tokens, models, or seats. We have no margin on your spend and no reason to want it higher.
- Take vendor referral fees. If a cheaper model would change the verdict, we say so, and nobody pays us for the recommendation.
- Produce a number we cannot reproduce. If your logs cannot support a figure, the report says the figure is missing. It does not guess.
Pricing
Fixed price per engagement
We quote a fixed price per engagement after a 30-minute scoping call. The quote names the workflows in scope, the artifact, and the delivery date, and it does not change once the work starts.
If the scoping call shows that an engagement is not warranted yet, the same note says so. Book the scoping call.
Not sure which engagement you need?
Send us one workflow and last month's token bill. The scoping note names the engagement that fits, or says that none of them is warranted yet.