What two data protection specialists told us about AI and client data

Every AI solution in a firm touches data, much of it client data, so most of the risk in bringing AI in sits in that data. A firm that knows what it holds, secures it and gives every risk an owner can bring AI in with confidence. Neil Aremband of Pathway 2 Privacy and Ethan Oak of Oak Advisory Solutions made the same points in two long conversations with us over the last two weeks.

Neil and Ethan are data protection specialists who each run their own firm and often work on projects together, helping organisations in very different industries work out what data they hold, where it goes and what could go wrong with it, and then put the gaps right. Neil works across data protection as a whole, and Ethan has focused in particular on data subject access requests.

Data has to be secured before AI goes near it

"The way me and Neil work is data first," Ethan told us. "There's no point adding AI in if you don't have your data secured first." One of the first things they do in a new organisation is a record of processing activities, setting down for each thing the organisation does what data it uses and its legal basis, so that people understand "what are you doing and why."

Every AI solution uses some of that data. When we used to go straight to implementation, about seven times in ten we had to go back and fix the tech stack and the data first. Now we start with the firm's tasks and the layers under them, and for the data that means where it is held, how complete, accurate and secure it is, who can get to it and which compliance rules apply. Where a layer isn't ready, fixing it goes first on the roadmap.

Every use of AI needs its risks assessed, including what is already running

The first data to look at is what people already put into AI. Neil's way in is one question in a data protection review: "Do you currently use AI, or do you intend to use AI?" If the answer is yes, in Neil's view the firm needs an AI impact assessment, which "should be done in advance of using the AI, but it could be done retrospectively as well."

Most firms need the retrospective version. We sometimes find people on free or personal accounts with client data in them, a compliance risk the firm carries today that has to be brought inside its AI policy before anything new goes on top.

For anything new that holds a lot of personal or client data, Ethan's advice is a data protection impact assessment first. "We're looking at worst cases," Ethan explained. "What are the four biggest risks that could happen in these processing activities? What would we do to mitigate that?"

That's close to how we implement solutions. Each one goes on the roadmap with what it returns and what it puts at risk side by side, because a solution that saves hours while sending client data somewhere it shouldn't go costs the firm far more than it saves, and its risks are designed out before it goes live.

Where the data is processed decides much of the risk

One of the largest risks in any assessment is where the data goes once AI uses it. When client data is involved, Neil and Ethan would have it run "ideally local. But if not, UK-based." If a provider's servers are in another country, Ethan would suggest a transfer impact assessment "to assess the risk of the transfer from one jurisdiction to another."

Where each solution runs depends on the data it touches. When nothing sensitive is involved, it can run on the AI provider's own cloud. Client work runs on the firm's own server, kept in the UK for a UK firm, where the provider never trains on what it is sent, or on a machine in the office. Open-source models can now run locally on a firm's own hardware, so the most sensitive work never crosses a border and there is no transfer to assess.

To know exactly which AI your team already uses on client data, where that data is processed and who checks the work, take the quiz below. It takes a few minutes.

Every risk needs an owner, a rank and an escalation point

Wherever a solution runs, we write its risks down with how each is mitigated and who owns it. On the second call Neil asked whether those risks are ranked. "Even if one owner has five risks," Neil said, "they need to know of their five which are the highest priority, so they can look at those first." Ethan added who to go to when something goes wrong: "ownership of each part of the process, ownership of the risks, and so there's sort of escalation points."

That gives a firm governance it can act on, because the owner of a solution that goes wrong knows which risk to look at first and who to escalate to. We are adding the ranking and the escalation point to how we write up every solution's risks.

Once solutions are live, the register stays current. As the models change, a solution that worked well can start to produce worse work, so we check and adjust each one and review its risks again before the quality slips unnoticed. When a rule changes, the register of every solution, the data it touches and who owns it shows which ones to review first.

What firm leaders should do now

Neil and Ethan's advice sets out the order a firm needs to follow, whoever does the work.

Start with a record of the firm's own work, broken down into the individual tasks and processes that create value for its clients: who does each one, how long it takes, what data it uses, its legal basis and who signs it off.

Before any solution, check the layers those tasks sit on: the business and each department, the tech stack, the quality, completeness and security of the data, compliance and governance, and the people who will work alongside the AI. Bring the AI already in use inside the firm's AI policy.

Then set out a roadmap of solutions, each fully costed with the hours it frees up, what it costs to implement, its first-year return and what it does to the recovery rate, with its risks ranked beside its return.

Implement the solutions one by one, in the roadmap's order, each assessed for its worst case, run where its data allows, covered by the AI policy and owned by a named person who checks what it produces before it reaches a client, so the work keeps its standard or gets better. Once they are live, keep managing them and their risks as the models and the rules change, with the firm owning its knowledge and renting the AI, so it is never tied to one vendor.

Done in that order, the firm gets more capacity from the same team without adding headcount, and every solution is compliant, keeps client data secure and has someone answering for its risks.

If you want us to do it for you, book a free call and we'll find you at least $50,000 a year in new business that AI can unlock. It's compliant, data-secure and has governance built in.

Know exactly where your firm's AI puts client data at risk

Our quiz asks which AI your team uses, where client data goes, what your clients have agreed to and who checks the work. Your report shows where your firm stands today and gives you a 30-day plan to make its AI safe with client data.

We'll find you $50,000 a year in new business AI can unlock in 72 hours, or you don't pay.