AI Policy

1. Principles

Four ideas guide every decision we make about AI:

• Usefulness before novelty. We recommend AI when it clearly beats the simpler alternative, and we say so when it does not.
• People stay responsible. A model can assist a decision. A person remains accountable for it.
• Everything can be explained. If we cannot reconstruct why a system produced a result, we do not put it in production.
• Your data stays yours. What you share with us serves your project and nothing else.

2. Your data in the AI systems we build

When we build an AI capability for you:

• Your data is used to operate your system and for no other purpose.
• We never use it to train shared or general-purpose models.
• We agree with you on which providers are used and where data is stored before anything is sent. We favor arrangements in which the provider is contractually barred from training on your data.
• We configure the provider to retain data for the shortest time the system needs.
• We log inputs and outputs so that any decision can be reviewed later, and those logs are protected by the same access controls as the remainder of the system.

3. Human oversight

Where the consequences justify it, we design for human review and not for full autonomy. In practice that includes:

• confidence thresholds that send uncertain cases to a person
• a visible way for users to question or override a result
• monitoring that detects changes in the system's behaviour after launch

When a decision affects a person's access to healthcare, credit, employment or something of similar weight, we treat human review as mandatory.

4. AI tools in our own work

Our engineers use AI-assisted development tools. These are our rules for them:

• Client code and data enter an AI tool only when the project agreement allows it, and never through a consumer-grade service.
• Protected Health Information and other regulated data are never submitted to a general-purpose AI tool.
• Code produced with AI assistance is reviewed by a person and must meet the same standard as code written by hand.
• If a client asks us not to use these tools on their project, we do not use them.

5. Limitations

AI systems can be wrong, including when they sound certain. They can reflect biases present in the data they learned from, and their behaviour can change when a provider updates a model.

We reduce these risks through testing, human review and monitoring. We do not claim to remove them, and we will tell you plainly where a system's limits are.

6. Questions

If you want to know how any of this applies to your project, write to hi@trivoandco.com. We are glad to go through it with your technical or legal team before you sign anything.