Wealth Management — “Supervised Wealth Planner Copilot”

Positioning: planning + scenario analysis + execution assistance, not “autonomous trading.”

1) Executive snapshot

ICP: independent wealth planners, RIAs, family offices, HNW individuals (with advisor oversight)
Wedge workflow: “Monthly review pack + rebalancing proposal + documentation.”
Autonomy: Supervised autopilot (agent drafts proposals; explicit approval for execution and client-facing outputs).

KPIs

time-to-review pack
compliance completeness
error rate in calculations / allocations

2) Product experience & UX

Advisor cockpit: intake → scenarios → proposal → approval → publish.
Trust UI: assumptions panel + citations + compliance checklist.

3) Agent design map

Skills: planner, risk manager, tax-aware reviewer (non-tax advice), compliance reviewer
Subagents: market data retriever, scenario engine, report generator, drift detector

4) Tool & data plane (MCP-centric)

Market data, holdings (read-only), CRM notes/KYC, disclosures/policies repo.

5) Context engineering plan

Pinned: IPS/risk constraints/prohibited assets/disclosures
JIT: latest holdings + market indicators
Immutable decision log: inputs + assumptions + versions per recommendation

6) Evals & observability

Offline: math correctness + constraint adherence + policy tests
Online: advisor edit distance, blocked approvals, p95 cost/pack, safety flags

7) Failure modes & mitigations

What breaks	Detect	Constrain	Prevent regression
Ungrounded recommendations	citation/assumption checks	proposal-only mode until approval	golden set + stress scenarios
Unauthorized execution	tool audits	execution tools off by default; approvals	CI policy tests
Prompt injection via docs	OWASP detectors	sanitize + isolate; least privilege	adversarial injection eval pack
Compliance gaps	checklist validator	block publish unless complete	compliance regression tests

8) Governance posture & rollout

High-risk posture: strong HITL, logging, and oversight by design.
Rollout: shadow mode → advisor-only beta → limited client publish → broader rollout.
Kill switch per capability (execute trades, send client email).

9) Business case + distribution loops

ROI: advisor hours saved + reduced errors + faster turnaround
Pricing: per advisor seat + per client case
Loops: client invites, review-pack templates, CRM embedded loop

10) Where RFT helps

Rule-following under complex constraints (tax, risk tolerance, liquidity, drawdown limits).
Decision discipline: consistent behavior under volatility, fewer inconsistent recommendations across turns.
Verification-first behavior: prefer “check + cite + simulate” over guessing.

What you train on (signals):

Trajectories: goals → profile → constraints → portfolio proposal → scenario checks → rebalance plan.
Graders:
- Constraint satisfaction (hard checks)
- Policy compliance (no forbidden advice flows, proper disclaimers)
- Portfolio sanity checks (risk/return proxies, diversification heuristics)
- Cost/latency penalties for over-tooling

Business metrics it can lift:

Higher plan completion rate (users reaching an actionable plan)
Lower support escalations / compliance incidents
Improved conversion to premium planning / advisory workflows (trust + reliability)