This question is often answered with one word: encrypted. That's true, and it says almost nothing. Encryption protects data that's at rest or in transit; the question you're actually asking is about data that's being used.

Below is the answer without reassurance up front. What stays yours, which parties get to see your information, what's shielded and how, where we've deliberately kept something out of the cloud, and which part of this question we simply can't answer.

01 / the entry pointWhose data is the system working with?

Yours. The system is built in accounts that are in your own name: your CRM, your mailboxes, your calendar, your telephony. There's no environment of ours that your customer data moves into and to which we hold the key.

That's not a detail, it's the core of the answer. The toughest security question isn't which lock is fitted where, but who's standing at the door when the collaboration ends. If everything stays in your own name, switching supplier is an administrative act rather than a hostage situation.

How that separation works exactly, and what stops if you cancel, is covered in who owns it once it works.

02 / the routeWhere does my customer data go?

Through a chain of cloud services, and you simply need to know that. An AI system isn't a box in your attic: it contains a CRM, a layer that links tasks together, and a language model that produces the actual text. Every link in the chain sees something of what passes through.

In our own setup, the three main ones are: a CRM for contacts and conversations, an orchestration layer that links the steps together, and Anthropic's AI for the thinking work. Around that sit smaller services for things like invoicing, telephony and document generation.

Every link is a party processing your customer information. A provider who can't give you a list of them doesn't have that list.

What you can demand about this, from us and from anyone: a list of the parties processing your data, where those parties do their processing, and a data processing agreement that sets it out in writing. Not as reassurance, but as a document.

03 / the safeguardsSo what is secured, and how?

Access, and that's arranged more tightly than most people expect. The underlying principle isn't "lock everything down" but: every key is allowed to do exactly one thing, and the more damage a key can cause, the fewer places it's allowed to sit.

Three things we've done in our own environment, and that you can ask any provider about:

That last point is underrated. Most of the damage AI causes isn't a data breach but a wrong message sent to the right person. A system that doesn't send by itself can't do that. Why we're so strict about this is covered in why we don't publish your posts automatically.

When choosing our own AI supplier, one party was ruled out because it trains on conversation data. That's exactly the kind of question you should be asking yourself: is a model being trained on my material, or not.

04 / the choiceWhat if something shouldn't go to the cloud?

Then you don't put it in, even if that's less convenient. For us, that's happened in practice with recorded phone calls: those are processed locally and deliberately don't go through the chain described above. A conversation with a lead is more sensitive material than a calendar appointment.

What's interesting is what that costs. There's a proposal to merge that processing with the rest, because it's technically neater and saves work. That proposal is still sitting there, because merging it would mean phone calls leaving the laptop. That trade-off hasn't been resolved by waving it away but by writing it down and leaving it open.

That's the most honest example we can give of how a choice like this works. There's no architecture where everything is secure and everywhere is convenient. There are trade-offs, and the question is whether a supplier names them or hides them.

05 / the limitWhat can't we answer about this?

The legal part. Whether your specific situation requires a data processing agreement, who is the processor and who the controller in it, and exactly what falls under your sector: that's work for a lawyer, and we're not one. We can describe how it's built; we can't determine whether it complies with the law in your case.

There's a second limit. There are things we haven't finished sorting out ourselves, and it would be odd to stay silent about that in a piece about honesty. Retention periods per type of data, for example: how long something stays and how deletion works, hasn't been written out on our side yet. That should exist, and it's coming.

So the answer to "is it safe" is neither yes nor no. It's: we can tell you exactly where your data sits, who can access it and what happens if you stop. Unbreakable doesn't exist, and anyone who promises that hasn't understood the question.

This text was produced with AI assistance and checked and approved for publication by a human.