AI unit economics · engineering, not slides
Know which AI workflows earn their spend.
Netesis instruments your AI workloads, attributes token spend to the task, sets a human baseline, and hands you a model your CFO will sign. Some workflows will not pay back. We tell you which.
Example scenario from the ROI model. Not a client result.
The problem
A line item nobody owns.
Two dashboards, no sentence
Finance and engineering are looking at different numbers.
Finance sees a bill for tokens that grows every month. It has a vendor name on it and no workflow name. Nobody can say which product feature, which team, or which task it paid for.
Engineering sees latency, error rates, and a quality score. Good numbers, but none of them is a cost, and none of them says whether the output was worth the review time a person spent on it.
Nobody owns the sentence that connects the two: this workflow costs $4.67 per accepted output, against $5.42 for a person, and pays back its build in 2.3 months.
Figures from the example scenario in the ROI model, not a client result.
That sentence is our product.
Method
Four steps, one verdict per workflow.
Everything runs inside your stack. Nothing leaves it. The output is a number per workflow and a written recommendation to keep it, fix it, or stop it.
-
01
Instrument
Tag every call with workflow, task, and outcome in your stack. No proxy, no data leaves.
-
02
Attribute
Roll token spend, retries, and review time up to the task, then to the workflow.
-
03
Baseline
Measure what the task costs when a person does it. Defensible, not flattering.
-
04
Verdict
Cost per accepted output, payback, breakeven acceptance, breakeven price. Keep, fix, or stop.
Services
One artifact per engagement.
Each one ends with something running in your stack or a document finance can audit. No decks.
-
01 · Audit
Spend attribution audit
Instrument up to five workflows, attribute one month of token spend to tasks, and capture review and rework time.
-
02 · Model
Unit-economics model
A human baseline study and the four-number verdict per workflow, with sensitivity to price, volume, and quality.
-
03 · Monitoring
Ongoing monitoring
Monthly re-attribution as prices, models, and volumes change, with an alert when a workflow crosses its breakeven.
-
04 · Pilot design
Pilot design
Before you build: pick the workflow, define acceptance, set the baseline, and size the pilot so it can prove or disprove payback.
Token ROI model
Run the numbers before you run the pilot.
Enter one workflow: volume, tokens per task, your contracted prices, the acceptance rate you actually get, and what a person costs. The model returns cost per accepted output, payback, and the two breakevens that decide whether it survives a price change or a quality dip.
- +$14,346 Net savings per month
- 2.3 months Payback on a $40,000 build
- 64% Breakeven acceptance rate
No sign-up. Nothing is sent anywhere. The formulas are on the page.
The part nobody sells
Some workflows should be shut off.
When the numbers say a workflow does not pay, we write that down and put it in the report. In the example above, doubling review time turns +$14,346 a month into −$6,454; doubling the token price only moves it to $13,725.
Nobody funds a pilot hoping to hear that. A verdict that can only say “keep” is not a verdict, and being told which workflow to stop is the deliverable.
Send us one workflow and last month's token bill.
We come back with a scoping note: whether an engagement makes sense, and what it would measure. Sometimes the answer is that it does not.