Pricing in this market is opaque, partly because engagements genuinely vary and partly because vagueness benefits sellers. Here are real ranges and, more usefully, what moves them.
Market rates by engagement model
These are broad US market ranges as of 2026, not Guardian's rates — we scope per engagement because honest pricing requires knowing what your data looks like.
| Model | Typical range | Best for |
|---|---|---|
| Individual specialist, hourly | $150–$400/hr | Narrow expert problems, review, architecture input |
| Boutique firm, hourly | $200–$500/hr | Full delivery of a scoped system |
| Large consultancy, hourly | $300–$800/hr | Programmes needing org-wide change management |
| Fixed-price project | $75k–$300k | One well-defined process, clear success bar |
| Retainer / fractional | $8k–$40k/mo | Ongoing capability, multiple small initiatives |
| Discovery / assessment | $15k–$50k | Deciding what to build before committing |
A discovery engagement is frequently the correct first purchase. Paying $25,000 to find out that your intended use case is blocked by a data problem is dramatically cheaper than finding out four months into a $250,000 build.
What actually drives the price
Data accessibility — the single biggest factor. Clean API access to well-structured data is a different project from extracting semantics out of 40,000 scanned PDFs. This can swing effort by 3–5x on otherwise identical use cases.
Integration depth. A pipeline that outputs a report is cheap. One that writes back into a heavily customized ERP with a change-control process is not.
Regulatory burden. Healthcare, financial services, insurance, and government work all add documentation, review cycles, and audit requirements. Budget 20–40% on top for genuinely regulated deployments.
Accuracy target. Going from 90% to 95% is meaningful work. Going from 95% to 99% can cost as much as everything before it, because you are now fighting the long tail. Be sure you need it.
Evaluation rigour. Building a proper eval set costs real money and is the thing most worth paying for. Vendors who skip it are cheaper and riskier.
Scope discipline. The most expensive engagements are the ones where "while you're in there" accumulates for six months.
The cheapest quote you receive is frequently the most expensive outcome, because it usually omits evaluation, exception handling, and integration — the three things that determine whether the system ships. Compare what is included, not the headline number.
Costs buyers forget
- Ongoing inference. A recurring operating cost, not a capital one. See LLM API pricing.
- Human review. If the system routes 15% of cases to a person, someone must be staffed for that. This is often the largest ongoing line item and the most commonly omitted from business cases.
- Maintenance. Prompts drift, models are deprecated, upstream schemas change. Budget 15–25% of build cost annually.
- Internal time. Your engineers, data owners, and subject-matter experts will spend significant hours. Labelling an eval set alone is often 40–80 hours of expert time.
- Change management. Training, documentation, and process redesign. Skipping it is the most common reason a working system goes unused.
How to buy well
- Buy discovery separately from build, and keep the right to walk away after it.
- Insist on a numeric success bar in the statement of work, measured on data the vendor has not seen.
- Require the eval set as a deliverable. It is the most durable asset you get.
- Put your engineers in the codebase from week one, or you will be buying a system you cannot maintain.
- Ask what happens to the 15% the system cannot handle. A vendor without a crisp answer has not built this before.
Frequently asked questions
Why is AI consulting more expensive than regular software consulting?
Partly scarcity of experienced practitioners, partly genuine uncertainty. Conventional software has predictable effort for a known feature. AI work has an irreducible research component — you often cannot know the achievable accuracy until you try on real data. Competent firms price that uncertainty in, either as a discovery phase or as a wider range.
Should we hire in-house instead?
For sustained, multi-year AI work, in-house is usually cheaper and strategically better. Consulting makes sense for the first project, for specialist gaps, or when speed matters more than cost. The best outcome is a consultancy that builds your first system and leaves your team able to run it.
What is a reasonable first-project budget?
$75,000–$150,000 for one well-defined process end to end, including evaluation and integration. Materially below that usually means scope has been cut somewhere that matters. Materially above, for a first project, usually means scope is too broad.
Do fixed-price AI projects work?
They work when the scope is genuinely well-defined and a discovery phase has already de-risked the data. Fixed price on an unexplored problem either costs you a large risk premium or produces a vendor incentive to cut corners on the parts you cannot see, which is typically evaluation.
How do we compare two proposals with very different prices?
Normalize on inclusions: is a formal eval set included, what accuracy is committed and measured how, is production integration in scope, what is the exception-handling design, who owns the code, and what does year-two support cost. Price differences usually collapse once you make those explicit.