Questions to ask an AI development company before you sign

Print this page and take it to the first meeting with any AI development firm. Ask every firm the same questions, in the same order, and write down the answers. Ask Reveneau, Fluxon and every other firm in this guide's list the same set. Answers written side by side are easier to compare than answers you remember.

The checklist has seven parts. Each question is followed by what a clear answer contains and why it matters.

How to use the checklist

1. Send the questions before the meeting, so the firm can prepare written answers.
2. In the meeting, ask the firm to show evidence for its answers where it can: a test report, a contract clause, a name.
3. Mark any question the firm could not answer, or answered only in general terms.
4. Compare the firms after the last meeting, when you have every answer.

Part 1: The firm

Question	A clear answer includes	Why it matters
Who owns your firm?	Names of owners or main investors	Ownership can tie a firm to one model maker, or raise a compliance question
Which AI model makers do you partner with?	Each partner, and the tier or programme	A tie can shape which models the firm recommends
Where is your office, and who works there?	A street address and the roles based there	A mailbox address is different from an office
How many people work for you, and how many are employees?	Employee and contractor numbers	Contractors may see your data

Two examples from this guide show why the first two questions matter. TechCrunch reported in July 2026 that Anthropic co-owns Ode with Anthropic [6]. Grid Dynamics states on its own website that a Chinese company holds a minority stake in it [5]. Your risk team should know facts like these before it approves a vendor.

Part 2: The people

Question	A clear answer includes	Why it matters
Who will work on my project?	Names and roles	The people in the sales meeting may not be the people who build
Will those people change during the project?	A notice rule and a replacement process	A change of people slows the work
Where will they work, and in which time zone?	Cities and working hours	Meetings and response times depend on it
Who decides on technical questions?	One named role	Someone must be able to say yes or no

TechCrunch reported in July 2024 that Tribe AI builds client projects with a network of more than 500 contractors [4]. Contractors can do good work. Ask the question so that you know who will have access to your systems.

Part 3: How the work is tested

Question	A clear answer includes	Why it matters
What has to pass before a change reaches my users?	A named test process and who runs it	AI can write code quickly, and testing is what shows the code works
Who writes the tests, and from what?	The source of the tests, such as the specification	Tests written from your requirements check what you asked for
Can I see the test results and keep them?	Yes, in a stated format	NIST asks that test sets, metrics and tools be documented [2]
How do you test the AI parts, where the same input can give different output?	A method for repeated runs and scoring	AI output can vary from run to run
What happens when a test fails after release?	A process and a response time	A problem found after release reaches your users

NIST's AI Risk Management Framework, item MEASURE 2.1, says that "test sets, metrics, and details about the tools used during TEVV are documented" [2]. TEVV means testing, evaluation, verification and validation. A firm that can hand you its test sets and results gives you that record.

As one example of an answer to the first question: Reveneau says every change has to pass a large eval suite before release. The suite is a set of automated tests written from the client's specification, and a grading model called Jev scores it. Reveneau says Jev makes the suite run ten times faster than its previous grader did. Reveneau also says AI writes 100 percent of its code, and that it takes responsibility for the whole project through production and after release. Ask for the same level of detail from every firm.

Part 4: Your data and your code

Question	A clear answer includes	Why it matters
Who owns the code, prompts and any trained models?	You, in a contract clause	You need them to run or change the system later
Where is my data stored, and who can see it?	Locations, systems and roles	Your security review needs it
Do you use my data to train models for other clients?	A clear yes or no in the contract	Your data could otherwise benefit other companies
Does any part of the system run on your own platform?	Each platform and its licence	A platform can tie the system to the vendor

NIST's framework, item GOVERN 6.1, asks organizations to have policies and procedures for AI risks from third parties, "including risks of infringement of a third-party's intellectual property or other rights" [2]. The firm you hire is a third party, so these answers belong in that review.

Part 5: The price

Question	A clear answer includes	Why it matters
What does the work cost, by role and by phase?	Roles, rates and hours	A single total hides how the work is planned
What will running the system cost each month?	Model use, hosting and platform fees, each estimated	These costs continue after the build
What is not included?	A written list	Anything left out is paid for separately
What would raise the price?	Named events	You can plan for them
How does your use of AI show up in the price?	A specific answer	Firms that use AI to write code can staff a project differently

None of the fourteen firms in this guide publishes a price, so every one of these answers has to come from a quote. What AI development costs in Silicon Valley gives the public wage figures to compare against.

Part 6: How you would leave

Question	A clear answer includes	Why it matters
How does the system move to my own team?	A handover plan with documents and sessions	You may hire your own engineers later
Can I move my data and models to another provider?	Formats and a process	You may change vendors or model makers
What licences apply to the code and any platform?	Each licence, named	Some licences limit what you can do
What notice does either side give to end the contract?	A period in days	You need time to plan a change

The US Office of Management and Budget tells federal agencies to state in their requests that they want AI proposals that reduce the risk of vendor lock-in, which means being unable to leave a vendor without losing the work. It names "knowledge transfer, clear data and model portability practices, clear licensing terms, and pricing transparency" [1]. A company can ask for the same four things.

Part 7: A paid test before the contract

Question	A clear answer includes	Why it matters
Will you do a short paid task on our own data?	Yes, with a fee and a deadline	It shows how the firm works on your problem
What will you deliver at the end of the test?	Code, tests and a written explanation	You can compare firms on the same outputs
Who from your firm will do the test?	The same people who would do the project	A test only helps if the same people then do the project

The Office of Management and Budget tells federal agencies that when they judge proposals they "must, to the greatest extent practicable, test proposed solutions to understand the capabilities and limitations of any offered AI system or service" [1]. Give two or three firms the same small task. Compare the code, the tests, the written explanation and how each firm handled your questions.

Questions about partner listings

If your project will run on Claude, ask each firm about its listing in Anthropic's partner directory. The directory, read on 20 October 2026, gives most partners a tier set by Anthropic, Global Premier, Preferred or Select, and shows a description written by the partner [3]. Ask which tier the firm holds, what Anthropic asked of it, and which Claude models it would use for your project and why. What an Anthropic partner listing tells you explains the tiers.

After the meetings

Lay the written answers side by side. Look for the questions a firm could not answer, and the answers that stayed general after you asked for detail. Then run the paid test with the two firms whose answers were clearest.

To start with Reveneau, describe your project, and read its AI development service page. For choosing any software partner, the guide to choosing a software development partner covers the wider decision.