A while back, a professional services client asked us to build a voice clone for a member of their leadership team.

The idea was to clone their voice once, then use it across marketing campaigns, internal comms and whatever else came up. It was a flagship project, a chance to show what this technology could do inside a big organisation.

The creative side was the easy bit. Working out where that voice would actually go once we'd uploaded it was much harder.

In the end, the project didn't go ahead.

The problem nobody plans for

Enterprise businesses need to take data protection seriously, and for good reason. On top of UK GDPR there are internal policies, InfoSec reviews and procurement checks. Every new supplier has to get through all of them.

Most creative AI tools aren't built with that in mind. They're built for individuals and small teams who want results fast. Data gets processed on servers that could be anywhere in the world, often passed along a chain of sub-processors, with retention and training terms that can be hard to pin down.

In this instance, we looked at a tool called ElevenLabs, the obvious choice for voice cloning. The challenges with this tool are the same challenges we'd face for any of these well known creative AI tools. Its data residency option, which limits where your data is stored, is only available on Enterprise plans. Even then, its own documentation says processing can still happen outside the chosen region through its affiliates and sub-processors. We couldn't confirm that every link in that chain met the standard our client's data protection team needed, so that's where it stopped.

It's a pattern we keep seeing. Enterprise versions of these tools solve part of the problem, but they're priced for company-wide rollouts, not one campaign. So the idea either stalls in legal or someone uses a consumer tool and hopes for the best.

Why voice and likeness raise the stakes

Someone's voice and face are part of who they are. If a clone is misused, they can't reset it like a password.

Strictly speaking, GDPR only treats voice and facial data as special category biometric data when it's used to identify a person. A voice clone arguably isn't that. But it's still personal data, the potential for misuse is obvious and most legal teams treat it with the same caution.

And when we talk about creative applications, often any data captured for creative purposes (whether thats voice, likeness, etc) falls into high sensitivity data categories, increasing the importance of knowing how data is being handled and processed at every step.

What we're doing about it

Our answer is to keep the processing local.

By using open-source models running on our own hardware, the data never leaves our environment or jurisdiction. It isn't sent to a processing centre in another country and there's no chain of sub-processors to audit.

That is an important development because our clients already have framework agreements and data processing agreements in place with us. The work stays inside a relationship that has already been through their procurement and legal checks, rather than adding a new platform they've never assessed.

Around that we put the basics in place: written consent from the person being cloned, an agreed list of where their voice can be used and deletion of the source recordings and trained model when the work is done.

The honest bit

Local processing isn't a free pass. We're still handling personal data, so the paperwork, security and consent all still apply. Open-source models come with their own licences that need verifying for commercial use. And it takes more setup than signing up to a subscription. It also is a slower process: while we use professional-grade hardware in our array, it can't compete with huge data centres located around the world.

But it means a project like the one we lost can now go ahead, and we can give our clients total confidence in how data is being handled and processed.

What's next

Big organisations want the same creative ambition as consumer brands, and voice, likeness and generative content are a growing part of that. Data protection is a good reason to be careful, but shouldn't be the reason good ideas never get made - and we've now found an answer.