Answer first
Use AI as a tightly bounded evidence clerk. It can organize approved generic material, list missing information, and draft neutral meeting questions. A qualified human and the firm’s actual policies retain every client-specific, advisory, supervisory, privacy, disclosure, communication, and transaction decision.
Mission outcome
Make a review packet—not a recommendation.
Output 01
Permitted-input record
Document system approval, data boundary, role/capacity question, and named reviewers before drafting begins.
Output 02
Neutral meeting board
Source-indexed agenda questions, conflicts, gaps, and escalation flags with no advice or client language.
Output 03
Supervision handoff
One review packet for compliance, privacy, and qualified professional review—still a draft.
Output 04
Stop-state receipt
“Do not use,” “insufficient evidence,” or “candidate for firm supervision review.” Nothing executes.
Field route
Six moves before any client-facing work
- Confirm permission and capacity. Record the approved workspace, data class, firm path, role, jurisdiction, and reviewers. Do not infer any of them.
- Index sources without importing sensitive detail. Preserve firm-approved versions and owners. Client-specific data stays in the approved firm system.
- Ask neutral questions. Turn gaps, contradictions, or “best option” requests into named-reviewer questions—not advice.
- Separate preparation from communication. Agendas and evidence queues are drafts. No client email, summary, disclosure, public claim, or recommendation leaves this desk.
- Route regulated decisions. Capacity, disclosure, privacy, supervision, product, performance, account, and transaction questions go to their authorized owners.
- Stop at review readiness. The desk creates a record for supervision; it never creates a plan, recommendation, trade, or approval.
Safety lock
What the AI desk may and may not do
Safe preparation
- index approved generic source material
- list missing facts and contradictions
- draft neutral agenda questions
- prepare review queues and evidence states
- run a disconnected synthetic pilot
Human and firm authority
- all individualized advice or recommendations
- client data, records, disclosures, and communications
- supervision, compliance, privacy/security, capacity
- accounts, products, allocations, trades, and transactions
- public AI claims, marketing, and approval
Failure modes
Do not let a polished draft become false authority.
- Putting NPI into a consumer tool. The desk starts with input approval; unknown means do not use.
- Answering “what should I do?” Turn it into a source and reviewer question; never synthesize advice.
- Calling the agenda client-ready. Review, firm policy, and communication controls remain separate.
- Claiming AI makes the practice compliant. It does not; claims about AI use also require substantiation and review.
- Confusing a checklist with a record of supervision. The actual firm record and reviewer process live outside this draft.
Runnable artifact
Run the 30-minute synthetic meeting-prep drill.
Use the desk to prove it blocks a fictional account pattern, “best option” request, and client-email prompt while producing a neutral review handoff.