AI, explained
How to price an AI project
An AI project has two costs that behave completely differently, and pricing goes wrong when they are collapsed into a single number.
Why the two parts have to be separate
The build is a one-off engineering cost with a defined end. Operating is a permanent cost that starts the day it goes live and never stops, because retrieval quality degrades as the corpus grows, model versions change under you, input distributions shift, and somebody has to notice. Bundling both into one figure produces either a build price that looks expensive or an operating commitment nobody funded, and the second is how systems quietly rot.
What sits in the build price
- Discovery against real data, not a sanitised sample. The awkward formats are where the time goes and quoting before seeing them is guessing.
- The evaluation set. Someone has to write the labelled cases the system is measured against, and it is real work that buyers routinely forget to fund.
- Integration with source systems, which is usually the largest line and the least discussed.
- Guardrails, permissions and the audit trail.
- Handover documentation and the traces that prove it works.
What sits in the operating fee
- Rerunning the evaluation set on a schedule and after every model version change.
- Corpus maintenance as documents are added, superseded or withdrawn.
- Cost monitoring and ceiling enforcement.
- Exception review and the fixes that come out of it.
Why per-seat pricing dies
Under a per-seat model, revenue is flat per user while cost rises with how much each user actually uses. Your best customers, the heavy users who prove the product works, become your worst margins, and the rational response is to throttle exactly the people who love it. Charge on the unit that drives cost. That means per outcome where an outcome is countable, such as a document processed, a case resolved or a report produced, or a tiered allowance with defined overage.
Setting the number itself
Price against the value of the problem, not the cost of the build. If a process consumes a known number of hours a month at a known cost, or if a missed deadline carries a known penalty, that arithmetic sets the floor and the client can check it themselves. Cost-plus pricing on an AI build systematically underprices, because the engineering is often modest while the process cost removed is not.
Digiton scopes and prices this way across 8 countries. An AI audit produces the scope, the evaluation plan and the two numbers.
Frequently asked questions
How do I price an AI project?
Separate a fixed price build from a monthly operating fee. The build covers discovery on real data, the evaluation set, integration, guardrails and handover. The operating fee covers evaluation reruns, corpus maintenance, cost monitoring and exception review. Metered inference either passes through or sits in a capped allowance.
Why does per-seat pricing fail for AI products?
Revenue per seat is flat while cost scales with usage, so heavy users, the ones proving the product works, become the worst margins. The rational response becomes throttling your best customers. Charge on the unit that drives cost: a countable outcome such as a document processed, or a tiered allowance with defined overage.
What do buyers forget to budget for in an AI project?
Three things: writing the labelled evaluation set the system is measured against, integration with source systems which is usually the largest engineering line, and the ongoing operating cost after go-live. A build funded with no operating phase has nobody assigned to accuracy once the model version changes.
Related
Ready to put AI to work?
Book a discovery audit and we will map the highest-ROI AI agents and automations for your business.
Book a discovery audit →