Verity for AI companies
AI businesses have a variable cost per request and a fixed price per seat. Verity attributes cost to accounts and tracks what is actually running in production.
Verity runs the company. Training infrastructure and model tooling stay where they are.
Verity / Company
LiveProduction accounts
64
Gross margin
48%
Accounts below cost
3
Pilots past their window
5
Needs attention
Illustrative figures. Verity shows your own company in this shape.
How the business runs
An AI company has a cost structure most software businesses do not: every request costs money. When pricing is per seat or per month and cost is per token, per image or per hour of compute, three accounts costing more to serve than they pay is not unusual — it is the default outcome unless cost is attributed per account.
The second characteristic is that what is running in production drifts. Two model versions live that nobody deliberately chose, deployed for a test and never reverted, is a common and consequential state, because behaviour customers depend on has changed without a decision.
The third is evaluation. Model changes need evidence before deployment, and evaluation depends on annotated data, which depends on labelling capacity. An annotation backlog is a deployment blocker one step removed.
The fourth is the pilot-to-production gap. Five pilots past their evaluation window with no recorded decision is the most expensive state in the business — engineering attention and serving cost with no commercial outcome either way.
The fifth is that customers ask what changed. A behaviour difference needs an answer that connects the account, the version it was served and when the change happened.
Verity attributes serving cost to accounts, records what is deployed and why, and tracks evaluations, annotation work and pilot decisions.
What Verity calls these things
What gets in the way
AI company difficulties come from variable delivery cost and changing model behaviour.
Serving cost is not attributed to accounts
Compute and inference spend is a single line and unit economics are unknown.
In Verity Cost is attributed per account and per feature against the price they pay.
Production model versions drift
Deployments made for tests remain live and nobody chose them.
In Verity Deployed versions carry the decision, evaluation and date behind them.
Evaluations wait on annotation
A model change cannot be evidenced because labelled data is not ready.
In Verity Annotation work is tracked as capacity against the evaluations depending on it.
Pilots have no decision point
An evaluation window passes and the pilot continues without a commercial outcome.
In Verity Pilots carry a decision date, criteria and an owner.
Behaviour change questions cannot be answered
A customer reports different results and nobody can say what changed.
In Verity Accounts, versions served and change dates are linked.
Heavy users are not identified early
Usage grows past plan assumptions and margin erodes silently.
In Verity Usage against plan assumptions is monitored per account.
The complete system
One system across cost, deployments, evaluations and accounts.
Control
One permission model and one audit trail, with serving cost attributed per account, feature and model version against the revenue each produces.
Why it matters here Variable cost against fixed price is the defining economic risk of the business.
In practice Three accounts costing more to serve than they pay.
Work
Each deployment carries its model version, the accounts it serves, the evaluation behind it, the person who approved it and the date.
Why it matters here What is running should be a decision with evidence attached.
In practice Two live versions nobody deliberately chose.
Records
Evaluation runs carry datasets, metrics, comparison against the incumbent and the decision they supported.
Why it matters here A deployment without an evaluation is a change without evidence.
In practice Evaluation results linked to the deployment they justified.
Workflows
Annotation work carries datasets, annotators, throughput, quality checks and the evaluations waiting on it.
Why it matters here Labelling capacity is an upstream constraint on shipping model changes.
In practice Annotation backlog blocking three evaluations.
Relationships
Accounts carry their plan, usage, serving cost, model versions, pilot state and decision dates.
Why it matters here A pilot without a decision date does not end.
In practice Five pilots past their evaluation window.
Orders
Plans carry their usage assumptions, actual consumption and margin per account.
Why it matters here Pricing assumptions have to be compared with real usage.
In practice Usage against plan assumptions by account.
People
Staff carry assignments, evaluation ownership, deployment approvals and annotation throughput.
Why it matters here Deployment approval is a named responsibility.
In practice Deployments by approver with evaluation attached.
Reports and analytics
Cost per account and feature, gross margin, deployment history, evaluation outcomes and pilot conversion come from the records.
Why it matters here Unit economics and model governance are both measurable.
In practice Gross margin by account and by feature.
Verity AI
Verity AI answers from your own account, cost, deployment and evaluation records, respects permissions, and can create assigned follow-ups.
Why it matters here The useful questions are about which accounts lose money and what is deployed.
In practice "Which accounts cost more than they pay?" returns three with usage patterns.
Communication
Pilot reviews, change notifications and usage conversations attach to the account.
Why it matters here A behaviour change customers notice needs to be a notification, not a discovery.
In practice Change notice recorded against the accounts affected.
Schedule
Evaluations, annotation capacity and deployment windows are planned together.
Why it matters here A deployment date is only real if the evaluation and its data are ready.
In practice Deployment windows planned against annotation capacity.
Suppliers
Providers carry cost, commitments, capacity and reliability.
Why it matters here Compute commitments are a large fixed obligation against variable demand.
In practice Committed compute against actual consumption.
Work in motion
These already happen. Recorded, unit economics and model governance both become visible.
Attaching the evaluation to the deployment is what makes the change reversible with reason.
Attribution per account is the only way flat pricing survives variable cost.
Labelling is a capacity constraint on shipping, not a background task.
A recorded decision to stop is worth more than an indefinite pilot.
Knowing which accounts were served which version is what makes the answer possible.
Verity AI
Verity AI reads the same account, cost, deployment and evaluation records the company creates as it operates. It answers from your own company, respects permissions, and can turn an answer into a pricing conversation or a deployment review.
Verity / Ask
Verity AI only returns what the person asking has permission to see.
Without chasing
Each runs from the company’s own records at the point the condition is met.
When
An account’s serving cost approaches its revenue
When
A deployment is made without an attached evaluation
When
A pilot passes its decision date
When
An evaluation is blocked on annotation
When
Usage exceeds plan assumptions
What you can understand
Unit economics, deployments and evaluations from operating records.
Verity records the company’s operations. Training infrastructure and model tooling continue as they are.
One system, different ways of seeing it
Everyone works from the same records.
Founder
Do the unit economics work?
Cost per account, gross margin, accounts below cost, pilot conversion.
Engineering lead
What is deployed and why?
Live versions and approvals, evaluations attached, rollback position, annotation capacity.
Research lead
What can we evaluate?
Evaluation queue, dataset availability, annotation throughput, results against incumbent.
Account manager
Where does this customer stand?
Usage against plan, cost to serve, pilot decision date, version and change history.
Where it is used
Inference and compute cost held per account and feature against the revenue each produces, which is the only way flat pricing survives a variable cost to serve.
Live model versions carrying the evaluation, the approver and the date, so production reflects decisions rather than leftover tests.
Annotation tracked as capacity against the evaluations that depend on it, making labelling a visible constraint on shipping.
Pilots carrying criteria, a decision date and an owner, so an evaluation converts or closes instead of continuing indefinitely.
Accounts linked to the versions they were served and the dates they changed, so a customer reporting different behaviour gets a real answer.
Consumption compared with plan assumptions per account, surfacing margin erosion before it becomes a renewal problem.
Plain-language questions across cost, deployments, evaluations and accounts, with pricing and review actions raised in the same step.
Getting there
Training infrastructure and model tooling continue and are mapped during implementation. Accounts with plans and usage, cost attribution structures, deployment and evaluation history, annotation records and compute commitments are brought across.
Implementation runs about four weeks: discovery and mapping, configuration, migration, then an ongoing operations partnership.
Questions
Verity AI answers questions from your own account, cost, deployment and evaluation records: which accounts cost more to serve than they pay, what gross margin is by feature, which model versions are live and what justified them, which pilots are past their decision date. Each answer can become a pricing conversation or a deployment review.
Because the cost is variable per request and the price is usually fixed per seat or per month. Without attribution the company knows its total compute spend and not which customers are profitable, and heavy users erode margin invisibly.
Every deployment carries the model version, the evaluation that justified it, the person who approved it and the accounts it affects, so what is running in production is a recorded decision rather than an accumulated state.
Annotation carries datasets, annotators, throughput and quality, with the evaluations waiting on it linked, so labelling capacity is visible as the upstream constraint on shipping model changes.
Pilots carry criteria, a decision date and an owner, and a decision is recorded either way. An indefinite pilot consumes engineering attention and serving cost with no commercial outcome, which is the most expensive state in the business.
Accounts are linked to the model versions they were served and the dates those changed, so a customer reporting different results receives a specific answer rather than an investigation.
No. Training infrastructure, experiment tracking and model tooling continue as they are. Verity holds the company around them — accounts, cost, deployments, evaluations, annotation and pilots.
About four weeks: discovery and mapping of cost attribution, deployment process, evaluation practice and plan structures, configuration, migration of accounts and history, then an ongoing operations partnership.
A few accounts usually carry the margin problem. Tell us how serving cost is attributed today.