A feature that works in the demo still has to work in the margin.
Build with AI knowing what each feature costs, which model runs it and what it does with your customers' data.
Product teams choose models quickly, price features before the costs are clear and answer customer questions about AI they did not fully design. GenWA Analyse shows usage and spend by feature and team, and GenWA Guard keeps model changes tested and controlled.
of revenue goes on inference (the cost of running AI models) at scaling-stage AI business-to-business software companies, and the share does not fall much as they grow.
ICONIQ, State of AI report, January 2026 (investor survey of AI software companies)of software companies absorb AI cost entirely into existing prices. Only 13% pass it fully through to customers.
Nalpeiron, State of B2B Software Monetization, 2026 (vendor survey, 243 companies)of AI spend goes to premium models, which handle 46% of the tokens (the units AI is billed in).
Ramp AI Index, April 2026 (vendor data)of companies planned to change their AI pricing within the next twelve months.
ICONIQ, State of AI report, January 2026Every AI feature adds a cost line you did not have before, and a model you do not control.
of revenue goes on inference at scaling AI software companies, the cost of running the models.
ICONIQ, January 2026Spend arrives as one bill, not by feature, team or customer. What it costs you: margin leaks you only see in the quarterly numbers.
Premium models take 56% of AI spend for 46% of tokens (Ramp, 2026). What it costs you: routine tasks run at premium prices when a smaller model would do.
In 2025 a widely used coding tool changed from a request allowance to a credit pool and customers ran out in days, followed by an apology and refunds (TechCrunch). Model suppliers have also retired models on short notice. What it costs you: a roadmap and a margin that depend on someone else's pricing page.
Security questionnaires now ask which models are used and where data goes. What it costs you: slower deals and answers written from memory.
Without a shared place for what works, effort is repeated and quality varies. What it costs you: engineering time and inconsistent behaviour in the product.
Usage is easy to count; value is rarely recorded. What it costs you: features that stay in the product because nobody can show they should go.
Treat the model as a supplier you can swap, not a foundation you build on.
The price and capability of models change every few months, so a product wired tightly to one of them takes on that supplier's decisions. The teams that stay in control separate three things: the rule about which tasks need which quality, the route that picks a model to meet the rule, and the record of what each feature costs. A model change then becomes a tested switch, and the margin on each feature becomes a number the product owner can read.
Control applies to AI used through GenWA, meaning connected and routed integrations.
Six ways to see feature cost, choose models deliberately and change them safely.
Analyse shows the cost. Guard controls the change. Advisers, human and agentic, help you set the rules.
Usage Insight shows model mix and cost by team, tool and task across AI connected to GenWA.
About Usage InsightBudgets, alerts and forecastsCost Meter sets budgets and alerts per team and task and forecasts spend, so a new feature has a cost line before launch.
About Cost MeterThe right model for each taskModel Choice sets which model handles which task, data type and risk, so routine work does not run on the dearest model.
About Model ChoiceTested before it reaches usersChange Control and the Sandbox let you test a model or release before it reaches your people, and keep a record of the change.
About Change ControlSavings found in your own dataSavings Finder works out where spend can come down from your usage data. We do not promise a fixed saving.
About Savings FinderAnswers for customer questionnairesEvidence Packs gather the register, data-flow map and control matrix, so you answer from the record rather than memory.
About Evidence PacksMost routine work does not need the most expensive model.
Sort tasks by how much quality they need and how costly a mistake is. Send routine work to a smaller model, escalate difficult work to a stronger one, and keep a person in the loop where the consequence is high.
Animated chart matching tasks to models by difficulty and cost
- The input is varied language that rules struggle to read.
- A person can check the result quickly or the user can correct it.
- The feature earns more than it costs to run.
- The answer must be identical every time.
- A lookup, form or workflow does the job.
- The decision concerns credit, hiring or medical matters, where people decide.
Every AI feature carries a running cost and a model dependency. Some jobs do not need either.
Three sectors where product teams meet the sharpest AI questions.
Pick the sector closest to yours.
Cost per feature and customer security reviews
Margin depends on model choice, and customers send questionnaires about AI. Usage Insight shows cost by feature, and Evidence Packs answer questionnaires from the record.
See technology and softwareSearch, recommendations and assistants at peak
Peak trading multiplies AI cost and puts brand and consent rules on every message. Cost Meter sets caps, and the Veto checks customer-facing content.
See retail and consumerAI features inside regulated journeys
Features that touch customers carry conduct rules, and people decide on credit. Rules set what AI may do in the journey, Guard checks it, and the Log supports your own review.
See financial servicesThree routes for a product team, from a cost review to a controlled build.
Advisers are human and agentic. Founder-led businesses can also apply to the Accelerator, a fourth route paid by equity, project, retainer or licence.
Our advisers, human and agentic, review your AI features, cost per task and model choices, and design the rules that decide which model does which job.
See advisoryProductsAnalyse and Guard run in your AWS or Azure account, or hosted by GenWA on AWS as agreed. They show cost by team and tool and keep model changes tested and recorded.
See AnalyseIncubatorWe build and run an AI workflow with you in an agreed engagement in a controlled environment, with cost caps, tests and the record designed in. You own the prompts, workflows and record.
See the IncubatorGet a first estimate of what you could save on AI cost.
Enter your monthly AI spend and a few assumptions to see an indicative range for what better model choice, shorter context and tidier instruction files could save. It is an estimate from your figures, not a quote or a promise of saving.
AI cost calculator
Indicative, from the numbers you enter
- Only the work a smaller model could do is moved. Premium models stay for the rest.
- This covers model choice only. Your health check looks at more, using your own usage.
Items marked assumption can be edited. No figures are stored or sent.
How this is worked out
Monthly saving = monthly spend x share on premium models x share a smaller model could do x price difference. The low and high figures move the middle share down and up by the range you set. Annual figure = middle monthly figure x 12.
Three colleagues who share the decision, and three functions to bring alongside.
The people who own technology, margin and the board's questions, and the teams beside product.
Owns the route AI takes through your systems.
For the chief technology or information officerChief financial officerOwns the number for AI spend and the case for it.
For the chief financial officerChief executiveWants AI features that grow the business without surprises.
For the chief executiveEngineering teamsBuild and run the features.
AI for engineering teamsCompliance teamsAnswer customer and regulator questions about AI.
AI for compliance teamsCustomer service teamsHear first when a feature goes wrong.
AI for customer serviceWant to see feature cost by team and task?
Book a demo and we will walk through usage, model mix and a forecast, using example data.
Straight answers.
How do product teams control the cost of AI features?
Measure cost by feature, team and task, set budgets and alerts, and match the model to the difficulty of the job. GenWA Analyse shows usage and spend for AI connected to GenWA, and Model Choice sets which model handles which task.
Should every AI feature use the best model?
No. Routine tasks often run well on a smaller, cheaper model, while difficult or high-consequence work may justify a stronger one. Test on your own tasks, not on public benchmarks, and decide per task.
What happens when a model supplier changes a model or price?
A change becomes a tested switch rather than a rewrite when your rules are written about tasks, data and risk, not about one supplier. Change Control and the Sandbox are there to test and record changes before they reach your people.
How do we answer customer questions about AI and data?
Answer from a record rather than memory. Evidence Packs gather the register of AI in use, the data-flow map and the control matrix, mapped to the frameworks you select.
Does GenWA build our AI features?
We advise and, through the Incubator, can build and run an AI workflow with you. We deploy controls, including harnesses, around your features. We do not take over your product roadmap.
Does GenWA see all AI used in our product?
It sees AI used through GenWA, meaning connected and routed integrations. An advisory review helps you find other AI in use and decide what to bring under control.
Is this useful for a team of ten people?
Yes. A product team in a business of ten people or more can start with a usage view and a simple rule for model choice, and add controls as it grows.
Know what each AI feature costs, and who can change it.
Cost views cover AI connected to GenWA. We do not promise a fixed saving.