By the CleverAI team · Updated
It is written for product, operations and IT leaders about to spend real budget on an AI project: a first production use case, a pilot that stalled, or a board asking for an AI plan that someone has to be accountable for.
Decide what you are buying before you compare firms
AI consulting covers very different work, and firms are built for different parts of it. Comparing a strategy consultancy with a software studio on the same scorecard produces a confusing answer. Write down which of these you need before the first call:
- A strategy: which AI use cases to pursue, in what order, and which to skip
- A feasibility check: whether your data and systems can support one specific use case
- A build: software in production, with evaluation, access control and monitoring
- Ongoing operation: someone to run, measure and improve the system after launch
Know the kinds of firm you will meet
Large strategy consultancies are good at aligning executives and mapping opportunities across a business; they rarely write and run the software themselves. Systems integrators bring scale and existing vendor relationships, and tend to recommend the platforms they partner with. Specialist AI studios and agencies design and build the software, with smaller teams and narrower reach. Independent consultants can be excellent for advice and reviews but have limited capacity for a build. Platform vendors offer professional services that know their own product well and will, naturally, propose it.
None of these is the right answer by default. The right one matches what you wrote down in the step above, and is honest about the parts it does not do.
Questions to ask in the first call
A first call should tell you how the firm thinks, not only what it sells. These questions separate firms that have shipped AI software from firms that have presented it:
- Which AI systems have you built that are running in production today, and can we see one?
- Who exactly would work on our project, and are they the people in this call?
- How would you measure whether the system works, and when would we agree on that measure?
- What would make you advise us not to build this?
- Which parts would you build, and which existing tools or models would you use?
- Where would our data be processed and stored, and by which providers?
- What will it cost to run each month after launch, not only to build?
- What do we own at the end: code, prompts, evaluation sets, documentation?
Proof to request, and how to read it
Ask for evidence you can check yourself. Working software you can use is the strongest proof; a reference you can call is next; a polished case study page is the weakest, because it is written by the firm.
Read case studies carefully. Ask whether each project shown was an AI project, whether this firm built it or a partner did, and whether it is still running. Firms that build general software often show web and mobile projects alongside their AI work. That is a fair way to show how they deliver, as long as it is labelled as such and not presented as AI experience.
Ask to see an evaluation report from a past project, with client details removed: how answers were tested, what failed and what changed. A firm that has run AI in production will have one. If a firm claims a security certification, ask for the certificate or the audit report itself and check which company and scope it covers.
Red flags in a proposal
Some patterns in proposals predict trouble later. None is a reason to reject a firm on its own, but each deserves a direct question:
- Accuracy or return-on-investment figures promised before anyone has looked at your data
- No evaluation plan, or success criteria to be defined after the build
- A fixed price for a scope that is still undefined
- Running costs left out: model usage, hosting, monitoring and maintenance
- Code, prompts or evaluation data that stay the property of the firm
- A proprietary platform you cannot leave without rebuilding
- Senior people in the sales process and an unnamed team on the work
- No mention of security review, data protection or human review steps
How to compare quotes fairly
Quotes for AI work are hard to compare because each firm draws the scope differently. Ask every firm to price the same written brief, and to split the quote into discovery, build and operation. Ask each one for its list of assumptions: the data it expects to receive, the systems it expects to access, and the decisions it expects you to make.
Then compare what each price includes rather than the total. A lower quote that leaves out evaluation, integration with your identity provider, or handover documentation often costs more by the end. Put the expected monthly running cost next to the build cost, because a system that is cheap to build can be expensive to operate.
Working with a firm in another country
Many AI firms work remotely across borders, and it can work well. Check the practical points before you sign: which company you contract with and under which law, how meeting times are set across time zones, how written updates are shared between calls, and which data protection terms apply to your data.
Applying this guide to CleverAI
We wrote this guide, so here are our own answers. CleverAI is a trade name of Azinove SAS, a software company in France, and we work with US teams remotely; there is no US office. Our AI evidence is the software we build and run ourselves: Converso, a chat workspace for GPT, Claude and Gemini, and GitDoc, hosted documentation that AI agents read over MCP, are both live and you can use them. The client case studies on our site are web and mobile platforms the Azinove team built or co-built; they show how we work. They are not AI-led projects: two include an AI feature, but we do not present them as AI case studies. We do not claim security certifications. Ask us every question above.