How much of your deal data does an AI sales tool actually need?

We've built Ollie under European data protection rules from the start. Over time, I've come to think those rules and good engineering point in the same direction.

When companies evaluate an AI tool, the security questions tend to be familiar ones. Where is our data stored? Who can access it? Is it used to train models? How long is it kept? Which subprocessors are involved?

They're all worth asking, and we answer them in every security review we do.

There's another question we've spent a lot of time on while building Optivian, though, and it comes up far less often:

How much of your data does the AI actually need in the first place?

European data protection rules have a fairly simple principle here, which GDPR calls data minimisation: process only as much personal data as you need. It turns out that's also good advice when you're building on language models. The less you need to send to a model, the faster and cheaper it is to run.

For an AI sales co-worker like Ollie, which works from your team's conversations with customers, that matters quite a bit.

A CRM contains the handful of things your team decided were worth recording about a deal. The conversations behind it contain everything else.

Pricing and contract negotiations. Budgets and business priorities. Technical requirements. Procurement processes. Internal decision-making. Competitive evaluations. The names and views of the people involved.

Much of that information belongs to your customer, not just to you. Some of it was shared in confidence, and some of it is personal data about the people involved.

Optivian needs access to this material to understand what's happening in a deal, so it reads every conversation connected to it. But having access doesn't mean every AI request needs all of it. What we keep small is what goes to the model each time someone asks a question.

Optivian reads the emails, meetings, and notes connected to a deal as they arrive, and keeps a short summary of each one. Over time, those summaries build into a current picture of the deal, which we call deal intelligence.

When someone asks Ollie a question later, he starts from that picture. If more detail is needed, Ollie can go back to the relevant original conversation rather than pulling everything into context by default.

A less selective approach is to pull large amounts of source material into a model whenever there's a question to answer. Roope wrote about the cost of this in June: when a customer connected their AI straight to their CRM, email, and call recorder, the same question about a deal used, on average, six times the tokens it did through Optivian.

Cost is only part of it, though. Every time you send context to a model, you're also deciding how much customer information goes with it.

We don't train AI models on your data, and neither do the model providers we use.

Where data is stored, our security controls, subprocessors and the rest of it are covered in our Trust Center and DPA.

Keeping each model call small isn't something we added on top. It's part of how Optivian was built.

If you're evaluating an AI tool that works with your customer conversations, I'd add one question to the usual security review:

How much of my data does the AI actually need in the first place?