Case study · Regulated financial services · Wealth & pensions · Switzerland
A real engagement, client anonymised. Figures measured with the client and rounded; not audited by a third party.
A 120-person financial firm brings all 68 of its AI uses under clear rules in four and a half months
An inventory of AI use, rules matched to risk, approved tools with controlled access, and approval within 48 hours for simple uses.
Scope and limitations
Rounded figures. A real engagement: the client’s name and any detail that would identify it have been removed.
- Company
- Swiss regulated financial services firm (wealth management and pensions), around 120 people
- Scope
- Generative AI use across the whole company, and the data passing through those tools
- Software in place
- Microsoft 365 and business applications; no AI tool chosen by the company
- Starting point
- Consumer AI tools used everywhere, with no inventory, no rules and no approval
Starting situation
Consumer AI tools had spread through the teams without a single decision being made: drafting letters, summarising client files, preparing analyses. Nobody knew who was using what, or with which data. Then an employee pasted a client’s data into one of these tools. The incident was caught just in time, but it showed that the risky path had become the default.
Management’s first instinct was a total ban. That would have solved nothing: use would have moved to personal phones, out of sight, and the firm would have lost what these tools were genuinely contributing. The board was not asking for a charter: it wanted proof that use was under control, in a business where pension data leaves no room for approximation.
The question asked
Ban everything and lose the useful work, or set rules and make the safe path faster than the workaround?
Constraints
Financial regulation and data protection
A regulated business handling personal data under the Swiss data protection act: every usage rule had to be written, traceable and defensible, not a mere recommendation.
Pension data
Clients’ pension data is among the most sensitive the firm holds. Exposing it to an unapproved tool was the first scenario to rule out.
A board that wants proof
A signed policy was not enough: the board asked for an inventory, controls, logs and indicators reviewed on fixed dates.
No habit of IT risk
The firm managed its financial risks rigorously but had no written practice for the risks of digital tools. The rules had to be ones the teams could own, not ones imposed on them.
What we put in place
Two deliverables: usage rules and the tools that make them real. One principle from the first day to the last: every use gets a risk level, every level gets controls to match, and the approved path must be faster than the workaround.
The building blocks
Stages of the engagement
Inventory and risk analysis 4 weeks
Team-by-team interviews and observation of real practice: 68 generative AI uses found, the data involved identified, leak and error scenarios described.
Rules and controls 4 weeks
Uses classified into four risk levels, controls per level, usage rules drafted with external counsel, sign-off by management.
Approved tools go live 6 weeks
Approved tools available with company accounts, role-based rights, usage logs, two-speed approval in operation.
Training and handover 4 weeks
Training by role (users, approvers, management), internal publication of approval decisions, first full dashboard, follow-up taken over by the client.
- Permitted uses defined by risk level
- No client data in an unapproved tool
- Approval within 48 hours for a low-risk use
- Usage logs reviewed every month
- In an incident, who alerts whom and how fast is written down and named
- Quarterly review of the rules with management
Measured results
At the start there was nothing: no inventory, no controls, no approval route. The figures below measure what the rules made visible and then brought under control, including six months of use after go-live.
Rounded figures. A real engagement: the client’s name and any detail that would identify it have been removed.
What we would do differently
Every study also says what could have been done better.
- 01
Train before delivering the tools, not after. Teams knew approved tools were coming and the wait caused frustration; training during go-live would have turned that delay into preparation.
- 02
Bring in counsel at the inventory stage, not only when drafting the rules. Two rounds of rework would have been avoided had the legal constraints guided the classification from the start.
- 03
Publish approval decisions from day one. We started mid-engagement, and that openness is what got the approval route adopted: everyone could see it decided quickly, and why.