Building products with AI
Own your code and data when you build AI with a vendor
What to keep in-house and what to negotiate when you hire someone to build your AI product.
· 5 min read · Product strategy · Vendor relationships · AI ownership
When you hire someone to build an AI tool for your business, you're making a bet that the relationship will work out. But good contracts assume it might not. You want to own three things: the code, the data, and the model. Here's why each one matters and what to ask for.
The code is your insurance policy
Your source code—the actual instructions that make the tool work—should live in a version control system you control. GitHub, GitLab, or Bitbucket all work. You own the account. The builder has access, but you keep the keys.
Why. If the builder goes out of business, you have the code. If you want to hire someone else to fix a bug or add a feature, you have it. If the vendor wants to raise their hosting fees by 300 percent in year two, you can move to someone cheaper without starting over.
A builder who says "we keep the code secure on our servers" is not saying they're being cautious. They're saying they own the relationship, not you. That's fine if you trust them completely and you never want to change anything. Most founders don't feel that way.
What to ask: "Can we store the source code in a GitHub account we own, with your team as contributors." A good builder says yes on the first conversation.
Your data stays yours
Training data—the examples you feed the AI to teach it how to behave—and customer data—the real invoices, emails, or records your tool processes—should be stored somewhere you control or at least somewhere you can get it back.
Why. Data is how the tool learns. If your builder stores it on their servers and you want to leave, they can say "sure, here's a zip file" or they can say "we delete it after 90 days, sorry." You lose the history that made the tool smarter. You also lose the ability to audit what happened to sensitive customer information.
A property manager in Surrey discovered this the hard way. They hired a team to build a tool that read lease agreements and extracted key dates. After six months, the builder wanted to renegotiate the contract. The property manager asked for their training data to get a second opinion on the tool's accuracy. The builder said it was stored on their cloud account and they'd need to ask their legal team. The conversation took three weeks. The data was still there, but the property manager learned they had no real control.
What to ask: "Where is the training data stored, and can we download it anytime. Can we also see a list of what customer data the tool processes, and where it's kept." If the answer is vague, keep asking until it's not.
The model is the thing that actually works
The trained model—the mathematical object that makes predictions or decisions—should be portable. You should be able to export it and run it somewhere else, or at least have a clear plan for what happens if you need to move.
Why. Models get better with time and real data. After a year, your model knows your business better than a generic one ever could. If you're locked into one vendor's platform and they sunset it or raise prices, you want to take that knowledge with you.
There are different types of models. Some are easy to move (you can download them as a file). Some are harder (they live inside a vendor's platform and you can only access them through an API, an interface that lets you ask the model questions). Ask which one you're getting.
A dental clinic in Kelowna built a tool that scheduled patient appointments based on treatment type and dentist availability. The builder trained it using a popular AI platform that made it easy to get started. Two years later, the clinic wanted to add more features, but the platform's pricing had doubled. Moving the model meant retraining it from scratch. The clinic lost six months.
What to ask: "Can we export the trained model, or at least get the weights and architecture so we can rebuild it elsewhere if we need to." If the answer is "it lives on our platform and that's how you use it," understand that you're renting it, not owning it.
What a good builder looks like
A builder who owns the code, data, and model usually says something like: "You own everything. We're building this for you. If you want to move it tomorrow, you can." They might add, "We hope you won't, because we're good at what we do, but you should be able to."
That's the voice of someone who's confident in their work and not afraid of competition. It's also the voice of someone who knows that trust is a better lock-in than a contract.
The conversation about ownership should happen before you sign anything. It's not confrontational. It's just clarity. A builder who gets defensive is telling you something. Listen.
Where to start
- Write down what you want the tool to do, then list what data it will touch (customer records, files, emails). That's what you need to own or control.
- Ask any builder you're considering: "Can we store the code in a GitHub account we own, download our training data anytime, and export the trained model if we move." See how they answer.
- If they say yes, ask how that works technically. Do they charge extra for it. Is there a time delay. Get specifics before you hire.