Anyone buying AI help runs into the same problem: the people selling it are also the people explaining it. This page is a checklist that holds up despite that conflict of interest, and the conflict applies here too. I sell exactly this service, so the fair thing is to hand you the scoring sheet and suggest you use it on me first. No competitors are named below. The categories and tests are enough.
Four services sold under one label
“AI consultancy” covers four quite different services:
- Strategy advice. Reports and roadmaps about what you could do. Needed occasionally, sold constantly.
- Enablement. Setting up the tools you already own and training your people to use them.
- Building. Custom automations, agents and software delivered into your business.
- Running. Keeping what was built monitored, secure and improving after go-live.
I think most disappointment in this market comes from buying the first when you needed the other three. A strategy deck doesn’t chase an invoice. Ask every candidate which of the four they do themselves, and who exactly will be doing the building.
Pricing models, and what each one rewards
Each pricing model nudges the supplier towards a different behaviour:
| Model | What it rewards | Verdict |
|---|---|---|
| Day rates | Time spent, which quietly rewards slowness | Only against a tightly defined scope |
| Fixed price per deliverable | The supplier scoping honestly | The fair shape for audits and first builds |
| Monthly retainer | Depends: stated monthly outputs, or just availability | Good for ongoing work with stated outputs |
| Percentage of savings | Sounds aligned, then stalls on arguments about measurement | Rarely survives contact with the accounts |
If a supplier can’t put a price on a defined first engagement, they’re still selling the strategy phase.
Red flags to watch for
These are the signs, seen from inside the trade, that a supplier will waste your budget:
- Slides without systems: nothing live you can try today
- Invisible builders: delivery handed to unnamed subcontractors while the brand does the talking
- Security as an appendix: no written answer to where data goes or what happens when the AI is wrong
- Technology-first proposals: a solution looking for a problem in your budget
- Lock-in by design: systems only the supplier can operate, and exports that never arrive
- Manufactured urgency: in a market moving this fast, pressure to sign is sales technique, not scarcity
What a good AI consultancy looks like
A good supplier can be checked rather than taken on charm. There are live systems you can try and a public record of what was built. There’s a written delivery method, starting from a specification, with security checks before anything goes live and monitoring afterwards. Scoping is honest, including telling you when a cheaper tool, or no build at all, is the right answer. Results are measured (hours saved, error rates, running costs) and reported without you having to chase. And leaving is easy: documentation, handover and your ownership of what you paid for, agreed up front. The questions in the FAQ fit this into five minutes of any sales call, and the vendor due diligence sheet gives you somewhere to write the answers.
My own answers to the checklist
Since I’ve asked you to use this on me, here are my answers. The systems are live on the Labs page. The method is written down on the Software Factory page, along with what I’d refuse to build. First engagements are fixed-price Audits and Kick-starters, and nothing gets priced until there’s a written spec and a clearly defined outcome.