Let AI prepare the client review; make the adviser own every judgment and action

A product launch is a catalyst, not a control framework

Anthropic launched Claude for Financial Advisors on September 14, 2026, followed by a public webinar and a GitHub reference implementation. The examples span onboarding, pre-meeting preparation, post-meeting follow-up, compliance review, prospect intake, portfolio rebalance review, and estate or tax briefs. The durable question is not whether one product can summarize a meeting. It is whether a firm can keep client identity, source cut-offs, calculations, judgment, communication, external writes, supervision, and records intact across any approved assistant.

The public repository is unusually helpful about its boundary. It says the plugin itself holds no client data, pulls from connected third parties, requires explicit approval for external writes, and assigns recommendations and advice to the adviser. It is also labeled a reference implementation that is not actively maintained or monitored and is provided as-is. Some security controls are structural; others depend on the model following instructions. That is useful implementation evidence, not regulatory approval or a measured productivity result.

Current community evidence is too thin to claim broad adoption or outcomes. One official repository, launch discussion, and general interest in Claude do not establish that advisory firms have safely deployed this workflow at scale. This guide therefore makes no adoption, time-saved, or error-reduction claim. It converts first-party product patterns and public regulatory guidance into a vendor-neutral operating procedure that a firm can test against its own policies.

Separate preparation from recommendation and action

A client review contains different kinds of work that should not share one permission. Retrieval finds source records. Reconciliation tests whether the records describe the same household and period. Deterministic computation produces performance, allocation, fees, and cash-flow measures. AI can organize the evidence and draft questions. A qualified professional interprets the client's circumstances, considers risks, costs, conflicts, and alternatives, and decides what—if anything—to recommend. A separate release step approves the exact language and system updates.

FINRA's 2026 GenAI report frames the relevant obligations through existing supervision, communications, books and records, and fair-dealing requirements. It highlights approval and governance, testing, monitoring, prompt and output logs, model versions, and human-in-the-loop review as useful considerations. SEC staff guidance on care obligations likewise emphasizes understanding the retail investor's profile and considering risks, rewards, costs, and reasonably available alternatives. Staff guidance is not itself a rule, but it is a useful design prompt: a fluent summary is not a substitute for missing client facts or professional analysis.

This guide is educational workflow guidance, not investment, legal, tax, or compliance advice. A firm must map it to its registrations, products, client agreements, policies, jurisdictions, supervisory system, privacy controls, and records schedule with qualified reviewers.

Freeze a household data contract before asking for prose

The most dangerous error can occur before the model writes a sentence: two people with similar names are merged, an old trust is treated as active, a spouse is assumed authorized, or portfolio data and planning data use different cut-offs. Start with a machine-readable run manifest. Make contradictions visible; never let the assistant choose the convenient value.

review_id: CR-2026-1042
meeting_at: 2026-10-03T14:00:00-04:00
adviser_owner: ADV-017
household:
  household_id: HH-8821
  people: [P-104, P-105]
  entities: [TRUST-14]
  authorized_participants: [P-104, P-105]
cutoffs:
  portfolio: 2026-09-28T20:00:00Z
  market_data: 2026-09-28T20:00:00Z
  crm: 2026-09-29T12:00:00Z
sources:
  - {id: SRC-01, system: portfolio, version: sha256:..., owner: ops}
  - {id: SRC-02, system: planning, version: plan-v17, owner: adviser}
  - {id: SRC-03, system: crm, version: export-1042, owner: service}
write_policy: draft_only
unresolved_conflicts: []

The source register should distinguish system facts, client statements, adviser notes, calculations, market data, approved research, and policies. A CRM note may be relevant, but it is not automatically current or authoritative. A model-generated meeting summary is a derivative artifact, not a primary source. Record the last-confirmed date and owner for facts that age: employment, tax status, liquidity needs, dependants, residence, objectives, risk profile, and time horizon.

Use deterministic tools for material calculations

Language models can help draft a formula, but material values should run through approved portfolio systems, spreadsheets, SQL, or reviewed code. Reconcile account counts, market values, cash, contributions, withdrawals, fees, returns, benchmarks, allocation weights, tax lots, and plan assumptions to control totals. Preserve the code or formula and the exact inputs. The site's spreadsheet review workflow provides a deeper formula and diff gate.

Performance deserves special care. Record methodology, gross or net basis, benchmark, fee treatment, cash-flow convention, period, and source. Do not let an assistant convert a portfolio-system number into promotional language. SEC adviser marketing rules impose conditions on advertisements, performance presentation, testimonials, and endorsements; firm compliance must decide how they apply to each communication.

Run the review as ten controlled handoffs

1. Open scope and identity

Name the review purpose, service tier, meeting participants, household ID, legal entities, accounts, adviser, and authorized representatives. Check for recent authority changes, death, divorce, trust amendments, powers of attorney, or account restrictions. An unresolved identity or authority conflict is a stop, not a drafting inconvenience.

2. Freeze sources and cut-offs

Export approved records at named times. Give each source an owner, version, confidentiality class, approval status, and stable reference. State whether data may enter the chosen tool and connector. If live connectors continue to change during preparation, either snapshot their output or record a new run version.

3. Reconcile the household population

Compare accounts, positions, cash, transactions, fees, tax lots, beneficiaries, plan facts, and CRM relationships. Differences enter an exception register with materiality, owner, required evidence, due date, and blocking status. Never smooth away a mismatch by choosing the newest-looking record.

4. Recalculate and validate

Run approved calculation tools and compare control totals. Test sign, units, currency, period, benchmark, fees, and cash-flow handling. Record calculation failures separately from data-quality failures so the right owner can fix them.

5. Build a dated client-fact timeline

Show what changed since the last review and when each fact was confirmed. Separate the client's statement, the system record, the adviser's interpretation, and an AI-generated hypothesis. Turn stale facts into questions for the meeting instead of silently treating them as current.

6. Assemble a topic-to-evidence matrix

For every possible discussion topic, list the client objective or constraint, supporting facts, calculation, costs, risks, conflicts, material alternatives to consider, missing evidence, and qualified owner. The matrix prepares judgment; it does not perform it.

7. Draft a question-led meeting brief

Lead with decisions the client needs to make, commitments from the prior meeting, material changes, exceptions, and questions. Avoid turning a model's guessed cause into a fact. If concentration rose because one holding appreciated, cite the calculation; if a life goal may have changed, ask.

8. Review communication risk

Flag promissory, absolute, unbalanced, misleading, performance, testimonial, endorsement, complaint, privacy, tax, legal, and unapproved-product language. Compliance or legal owners decide applicability. AI may detect patterns, but it cannot waive review or certify a communication.

9. Draft post-meeting records from confirmed notes

Create separate drafts for the client recap, CRM note, tasks, and proposed data changes. Attribute client statements and adviser explanations. Distinguish decisions made from issues deferred. Never backfill what “must have been discussed” from the pre-meeting pack.

10. Approve exact versions and writes

Bind approval to the precise communication, CRM note, field-level changes, destination, and audience. For external writes, preserve the prior value, proposed value, approver, time, tool identity when relevant, result, and rollback route. Books-and-records retention should cover the records required by the firm's rules and policies, not every token indiscriminately.

Worked example: a review that should stop before advice

Consider a household review with two spouses and a revocable trust. The portfolio system shows $2.4 million across seven accounts as of September 28. The planning system shows $2.31 million from September 12. CRM says one spouse expects to retire in 2028; the last signed plan says 2030. A concentrated employer-stock position is 23 percent of investable assets. The client also emailed about a home purchase, but no amount or date is confirmed.

The assistant may reconcile that the valuation difference is mostly explained by date and market movement, calculate concentration and cash-flow scenarios with approved tools, and draft questions about retirement and the purchase. It must not infer a new time horizon, recommend selling employer stock, claim a tax outcome, or treat the email as a finalized liquidity need.

A good system returns “not ready” here. The pre-meeting brief can still be valuable: it tells the adviser which conflicts to resolve and which calculations can wait. Productivity comes from focusing professional time, not from manufacturing a recommendation before the facts exist.

Failure modes the workflow must block

The Anthropic reference implementation explicitly notes that some controls depend on the model following instructions. That is why the most consequential boundaries should also exist outside the prompt: connector permissions, read-only defaults, data-loss prevention, approved destinations, write confirmation, logging, and supervised exception routes.

Pilot for 30 days without touching live client records

Week 1: define one review class

Select a recurring review with known sources and moderate complexity. Map current preparation time, corrections, missed commitments, late data, review steps, and record destinations. Approve the exact tool, connectors, data classes, and test accounts. Use closed or synthetic households first.

Week 2: build the source and calculation gates

Create the run manifest, source register, reconciliation rules, calculation workbook, exception log, topic matrix, and approval manifest. Seed identity collisions, stale objectives, mixed dates, missing fees, wrong benchmarks, and unapproved marketing language. The process should stop on each one.

Week 3: run shadow reviews

Prepare packets in parallel with the existing process. Do not send AI drafts or write to CRM. Reviewers compare completeness, traceability, calculation accuracy, question quality, false positives, and minutes spent resolving exceptions. Record edits rather than scoring “looks good.”

Week 4: authorize a narrow lane

If the evidence supports it, allow one bounded output such as a pre-meeting internal brief. Keep client communication and CRM write-back behind separate gates. Document model or tool version, monitoring, incident response, change approval, retention, and a rollback to the prior process.

Final human review gate

FAQ

Can AI recommend investments for a client review?

This workflow does not delegate recommendations. AI may organize supplied evidence and draft alternatives, but a qualified adviser must evaluate the client's profile, risks, costs, alternatives, conflicts, and firm obligations before making or communicating a recommendation.

Can the AI write directly to the CRM?

Use draft-only output by default. A named adviser or authorized operations reviewer should approve the exact note and fields before any write. Preserve the prior value and a write receipt where the system supports them.

Does this workflow satisfy FINRA or SEC requirements?

No generic workflow guarantees compliance. Firms must map it to applicable legal, regulatory, supervisory, communications, privacy, and records obligations with qualified compliance and legal reviewers.

Do we need to retain every prompt?

Follow the firm's records analysis and applicable requirements. Retain enough to reproduce material transformations, supervise communications and actions, investigate errors, and preserve required business records. More indiscriminate logging is not automatically better if it creates privacy or security risk.

What if the tool cannot cite a source field?

Treat the output as an unverified draft and do not use it for a material claim. Prefer connectors and exports that preserve stable references, or manually attach the evidence before review.