AI implementation cost and engagement models
AI implementation is bought in four phases: assessment, pilot, production build and run. Each phase has its own pricing model, and most of the total cost sits in the production build and the years of running that follow. None of the firms in this guide publish price lists, so this page explains how quotes are built and what moves them.
The four phases and how each is priced
Partners rarely quote one number for the whole job. They price each phase separately, because the scope of the next phase is only known once the previous one ends.
| Phase | What you get | Typical length | Usual pricing model |
|---|---|---|---|
| Assessment | A ranked list of use cases, a data and systems check, and a scoped first project | 1 to 4 weeks | Fixed fee, sometimes credited against the build |
| Pilot | A working version for a small group of real users, with measured results | 4 to 12 weeks | Fixed price for a fixed scope, or a capped time budget |
| Production build | Integrations, access control, monitoring, testing, documentation and rollout | 2 to 6 months | Time and materials with milestones, or fixed price per milestone |
| Run | Monitoring, fixes, model updates, prompt and evaluation upkeep, small improvements | Ongoing | Monthly retainer or managed service fee |
What drives the price
Two projects with the same goal can differ several times over in cost. The reasons are usually visible before the first workshop.
- Number of integrations. Each system the AI must read from or write to (CRM, ERP, ticketing, document stores) adds design, build and test work.
- State of the data. Clean, permissioned, searchable documents are cheap to use. Scattered files with unclear access rights are not.
- Risk of a wrong answer. A drafting aid reviewed by a person needs less testing than an agent that sends messages or changes records by itself.
- Regulation. Healthcare, finance and public-sector work adds documentation, hosting constraints and review steps.
- Team shape. Day rates at a global integrator, an AI-native consultancy and a boutique studio differ, and so does the number of people placed on the project.
- Users and volume. More users mean more support, more edge cases and higher model usage bills.
Costs that are not in the partner's quote
The services quote is only one part of the budget. Model usage is billed by the model provider, normally per token, and grows with volume. Cloud hosting, vector search and monitoring tools carry their own fees. Your own staff will spend time in workshops, on test reviews and on answering questions about how the process works today. That internal time is the cost buyers most often leave out.
Ask every bidder for an estimate of monthly running costs at your expected volume, split into model usage, infrastructure and support. A partner who has run similar systems can give a range and say what would push it up.
Engagement models compared
| Model | Works well when | Watch for |
|---|---|---|
| Fixed price | Scope is narrow and well understood, as in an assessment or a tightly defined pilot | Change requests, and a partner padding the price to cover unknowns |
| Time and materials | Scope will change as you learn, as in most production builds | No natural cap, so agree a budget ceiling and review points |
| Retainer | The system is live and needs steady upkeep and small changes | Unused hours, and unclear response times |
| Forward-deployed team | You want engineers embedded with your staff for months | Dependence on people who leave when the contract ends |
| Outcome-based | The result is easy to measure and both sides trust the numbers | Disputes over attribution, and long contracts |
How to keep the first project affordable
Start with one process, one team and one or two systems. Choose a task where a person checks the output, so the testing burden stays small. Use the tools you already pay for where they are good enough. Put a ceiling on the pilot budget and write down the result that would justify spending on production.
Be wary of two extremes. A very cheap pilot often skips integration and security work, which then appears as a surprise in the production quote. A very large first contract commits you before anyone has seen the system work with your data.
Common questions
Do AI implementation partners publish prices?
Almost never. Pricing depends on scope, team mix and region, so firms quote after an initial conversation. Ask for the day rates behind a fixed price and for the assumptions it rests on.
Is the pilot or the production build more expensive?
The production build, in nearly every case. A pilot proves the idea with a small group. Production adds integrations, permissions, monitoring, testing and rollout, which is where most of the engineering time goes.
Who pays for model usage?
Usually you do, directly to the model or cloud provider, on your own account. That keeps the bill transparent and lets you change partner without moving infrastructure.