Insights · Governance
What to agree in writing before an AI project starts
Many disputes in technology projects come from things that were assumed rather than agreed. With AI, the list of things worth agreeing is a little longer.
A good engagement letter is not a formality. It is the place where both sides find out whether they mean the same thing by the same words. For AI and software work, we settle the following points in writing before any work starts, and we suggest that any organization buying this kind of work asks its supplier to do the same.
This is a practical checklist, not legal advice. Your own lawyer should review any agreement that matters to your business.
1. Scope, and what is outside it
What will be delivered, in terms a non-specialist can check. Just as important: what will not be delivered. "Integration with your systems" means very different things to different people. Name the systems.
2. How success will be measured
The measures agreed at the start, how they will be reported, and how often. If the work is a pilot, the threshold that decides whether it continues.
3. Who owns what is built
AI projects produce more kinds of work than conventional software: code, configurations, prompts and instructions, workflows, evaluation tests, documentation, and sometimes models adapted to your data. The agreement should say who owns each of these when the engagement ends, and what the supplier may reuse elsewhere.
4. What data is used, and where it goes
Which of your information the project will use; whether any of it is personal information; which outside services will process it and where they say they process it; whether those services may keep it or use it to train their own models; and who may see it. In Canada, federal and provincial privacy laws govern how personal information is collected, used and disclosed, including when it is sent to an outside service. The agreement should make clear how the project meets them.
5. Access, and its removal
Which systems the supplier will have access to, under whose accounts, and how that access is removed when the work is done.
6. Accountability for outputs
AI systems make mistakes. The agreement should say how outputs will be checked before anyone relies on them, who is responsible for those checks, and how errors will be reported and corrected.
7. Price, and how it can change
A fixed price for defined work, or a clear rate for work that cannot yet be defined. Usage-based charges from AI services should be estimated in advance and reported as they occur, so that nobody is surprised by a bill.
8. How you can leave
Complete documentation, a handover to your own people or another supplier, and the return or deletion of your data. An organization should never be dependent on one supplier because nobody else can understand what was built.
If a point on this list feels too detailed to raise at the start, that is usually the point that causes the dispute at the end.
None of this makes a project slower. It makes the first weeks more precise and the last weeks far less contentious. At NLM Cognitive Solutions, it is how every engagement begins.
NLM Cognitive Solutions
To discuss this article or your own situation, write to hello@nlmcognitive.com.