Generative AI in a Regulated Advisory Firm
A 400-person advisory firm has run a six-month generative AI pilot with 80 consultants using a public LLM through personal accounts. Internal survey reports 6.5 hours saved per week; partners want firm-wide rollout. The DPO has found client contracts, including two with audit rights, pasted into prompts; no data processing agreement is in place; the vendor's consumer terms allow model training on inputs. Junior staff use AI to draft advice that is sent with light review. One client has asked whether AI was used on its account. The firm's professional indemnity insurer has requested the AI governance policy, which does not exist. The managing partner wants a decision before the next partner meeting; the CIO favours a private enterprise tenant; the CFO questions the cost. The pilot budget is €60k; enterprise tenant quote is €180k annually.
1. What immediate containment actions should the firm take, and what must be disclosed to clients?
Read the worked answer
Contain first, then assess, then disclose facts-not reassurance.
Immediate actions:
- Suspend use of personal accounts and consumer tools for client data. Preserve relevant logs and prompt records for the incident review.
- Classify what was pasted: client identifiers, contract terms, personal data, or privileged advice. The response differs by category.
- Stop any training or retention by the vendor; confirm in writing what was retained and whether deletion is possible.
- Convene the DPO, legal, risk and the managing partner. Open a formal incident record and inform the insurer.
Disclosure:
- Notify clients whose data was pasted where contracts, audit rights or law require it. Do not over-notify before facts are established, but do not wait for a perfect assessment either.
- Answer the client who asked directly and truthfully: what tools were used, on what data, and what review occurred.
- Tell clients what has changed, including the suspension and the remediation plan.
Reasoning: the firm's core asset is trusted advice. A cover-up is a bigger risk than a contained disclosure. The professional indemnity insurer will ask whether the firm identified and managed the exposure promptly.
2. Enterprise tenant or consumer tools? What assessment criteria and contract terms should drive the decision?
Read the worked answer
Recommend an enterprise tenant for any client-confidential work; allow consumer tools only for non-sensitive drafting under policy.
The €180k annual cost must be compared with the risk exposure, not just the €60k pilot budget. One client audit or contract breach can exceed the difference.
Assessment criteria:
- Data use: no training on inputs, no human review of prompts, documented retention and deletion.
- Tenant isolation: single-tenant or logically isolated environment, SSO, role-based access, audit logs.
- Compliance: data processing agreement, subprocessor list, breach notification, residency options, current security reports.
- Professional controls: eDiscovery, legal hold, model versioning, output logging, and an exit plan for data export.
- Usability: if the enterprise tool is materially slower or harder, consultants will revert to consumer tools. Adoption is part of security.
Non-negotiable contract terms: no training on firm data, deletion on exit, audit rights, liability for data breach, and notice of material model changes. The consumer tool can remain for brainstorming with no client data, but that boundary must be technically enforced and audited.
3. How should the firm measure whether the AI pilot created value, given the 6.5 hours-per-week claim?
Read the worked answer
Replace self-reported time savings with task-level evidence and quality-adjusted output.
The 6.5 hours claim is a perception metric, not a value metric. It may include time shifted to review, correction or re-prompting.
Measure:
- Task-level before/after: proposal drafting, research summarisation, meeting notes. Time each stage, including review and rework.
- Quality: reviewer scores, error rates, client revisions, and compliance exceptions. Faster bad advice is a cost, not a benefit.
- Commercial outcomes: proposal cycle time, win rate, utilization, realization and client satisfaction. Do not attribute all changes to AI.
- Control group: compare a pilot cohort with a similar non-user group, or stagger rollout.
- Adoption and risk: active use, policy exceptions, data-handling incidents, and hours spent on governance.
Set a decision rule in advance: scale only if quality does not decline, net time savings survive review, and no material data incidents occur. This protects the firm from a firm-wide rollout based on an optimistic survey and gives the CFO a defensible business case.