The L3 Ch1 L1 workflow audit produced the practice's motion inventory and the 90-day intervention list. This lesson takes each prioritized motion and answers the question every L3 workflow must answer before deployment: where does AI draft, where does AI extract, where does AI propose, and — most importantly — where does the registered, licensed human re-enter the loop to sign under Reg BI? The diagram that answers this question is the human-AI handoff diagram. It is the operational artifact that converts "we use AI for X" into a defensible Reg BI file, a passable Rule 2210 principal review queue, a clean Rule 4511 retention package, and an answer to the 2026 SEC examiner's question of "where exactly did a human exercise judgment on this recommendation?" Every multi-tool, multi-step workflow in L3 chapters 2 through 10 assumes the practice has built the handoff diagram for that workflow. This lesson installs the diagramming discipline.

The Four Handoff Modes — Draft, Extract, Propose, Decide

An AI tool integrated into an advisor workflow operates in one of four modes, each with a different handoff implication. The L3 advisor recognizes the mode at the moment the tool produces output and applies the matched verification and signoff discipline. Confusing the modes is the L2-to-L3 step-up that most advisors trip on.

Draft Mode

AI produces prose intended for client-facing or regulator-facing delivery. Examples: the Zocks-transcript-to-follow-up-email (L2 Ch2 L3), the quarterly commentary (L2 Ch6 L1), the difficult-conversation message (L2 Ch6 L2), the IPS narrative (L2 Ch5 L1), the Reg BI rollover memo (L2 Ch7 L2), the Roth conversion recommendation memo (L3 Ch2 L3), the estate-gap-and-beneficiary-audit attorney handoff (L3 Ch5 L2). The handoff implication: the draft is a marketing communication under Rule 206(4)-1 and a recordable communication under FINRA Rule 2210, and the registered, licensed human owns the substantive content. The handoff diagram for a draft-mode artifact places explicit human review nodes after AI generation, before any client or regulator sees the output. The retained record under Rule 4511 includes the system prompt version (L2 Ch8 L1), the user prompt, the AI's first-pass output, the human's edits (the redline), and the signoff identity plus timestamp.

Extract Mode

AI pulls structured data from unstructured source documents. Examples: Holistiplan reading a 1040 to a tax-planning matrix, FP Alpha extracting agents/trustees/distribution-mechanics from a revocable trust, Wealth.com reading a will and pulling beneficiaries, custodian statement OCR to a positions-and-basis table, Form 4 insider filings to an equity-comp picture, 401(k) plan document to a mega-backdoor Roth eligibility decision (L3 Ch3 L3), K-1 distributions to a pass-through-entity tax projection. The handoff implication: the extracted fields are facts about the source document, not recommendations; the verification discipline is field-by-field source-system reconciliation (Cardinal Rule from L1 Ch2 L3 applied to extraction). The handoff diagram for an extract-mode artifact places the verification node between the AI's structured output and the downstream consumer (the planning software, the tax memo, the eligibility decision). The retained record under Rule 4511 includes the source document, the AI's extraction output (JSON or table), the field-by-field verification log, and the verifying-person's signoff.

Propose Mode

AI generates a recommendation-shaped paragraph or list — a proposed allocation, a proposed Roth conversion size, a proposed rollover direction, a proposed estate vehicle decision. The proposal is not yet a recommendation under Reg BI §240.15l-1 or under the Investment Advisers Act fiduciary duty. The proposal is an input the registered, licensed advisor evaluates, modifies, or rejects. Examples: the L3 Ch2 L1 50-household Roth conversion screen produces 50 per-household proposed conversion sizes; the L3 Ch2 L2 per-client sizing memo produces a proposed dollar amount with bracket-fill, NIIT, and IRMAA rationale; the L3 Ch5 L3 advanced-vehicle decision tree produces a proposed SLAT/ILIT/CRT/CLAT/DAF allocation. The handoff implication: the proposal is a draft until the registered person reads it, evaluates it against the client's IPS and total household context, applies professional judgment, modifies as needed, and converts it into a recommendation by signing. The retained record under Rule 4511 includes the proposal, the human evaluation notes (the documented application of judgment), the modification (if any), and the signed recommendation. The conversion of proposal-to-recommendation is the discrete moment Reg BI's four obligations attach.

Decide Mode

AI takes an action — places a trade, sends an email, files a form, processes an RMD, initiates an ACATs. This is "agentic AI" in the FINRA 2026 Annual Regulatory Oversight Report's framing under Rule 3110 reasonable-design supervisory obligations. The handoff implication is the most stringent: the supervisory architecture must permit the action only within a pre-defined, written, narrowly-scoped envelope (allowed actions, prohibited actions, value thresholds, supervisory-review windows, kill-switch design), and Rule 4511 retention captures the agent's decision log, the action taken, the supervisory review, and the post-action audit. The handoff diagram for a decide-mode artifact requires (a) a pre-action human authorization at the policy level, (b) the agent's action log, (c) a post-action human review window before the action becomes irreversible (where possible), and (d) the audit trail. Decide-mode deployments are the least-common L3 pattern in May 2026 — most "agentic" wealth deployments are actually propose-mode with a streamlined approval step rather than true autonomous action.

What the Handoff Diagram Looks Like in Practice

The handoff diagram is a discipline, not a tool. The 2026 advisor practice convention uses one of three representations: a Visio/Lucid swim-lane (AI lane, advisor lane, paraplanner lane, CCO lane), a Mermaid or PlantUML text diagram checked into the firm's WSP repository, or a tabular workflow specification in the firm's compliance manual. Whichever format, the diagram contains the same six elements.

Trigger and Input

The event that initiates the workflow (calendar-driven: October 1 RMD calendar generation; transcript-driven: Zocks meeting end; document-driven: client uploads will to Wealth.com; population-driven: quarterly batch run of L3 Ch2 L1 Roth conversion screen). The inputs the workflow consumes (custodian feeds, planning software state, prior CRM activity, prior IPS edition, prior Reg BI evidence log entries, prior Smarsh archive context).

AI Mode and Output

Which of the four modes (draft, extract, propose, decide) the AI operates in for this workflow. The structured output schema (L2 Ch8 L2) the AI must conform to. The named AI tool (Jump, Zocks, Holistiplan, FP Alpha, Wealth.com, RightCapital, eMoney, MoneyGuidePro, Salesforce FSC + Einstein, etc.) and its model version. The system prompt version and user prompt template (L2 Ch8 L1).

Verification Checkpoint

The three-tier verification (L1 Ch2 L3, formalized in L3 Ch1 L3 next): source-system check (does the AI's number match the custodian feed, the Holistiplan extraction, the RightCapital plan?), regulatory check (is the rule cited correctly? Is the proposed action permissible under Reg BI, Marketing Rule, FINRA Rules 2210/3110/4511, Reg S-P, state DOI, NAIC Model #275?), client-fit check (does the recommendation fit this household's IPS, risk tolerance, prior decisions, life situation, and current planning state?). The verification node is human-occupied by default; in narrow extract-mode cases the source-system check can be partially automated against canonical data sources but the regulatory and client-fit checks remain human.

Human Judgment and Signoff Node

The discrete moment the registered, licensed advisor applies professional judgment, accepts or modifies the AI's output, and signs. For draft-mode artifacts, this is the redline-and-approve step. For extract-mode artifacts, this is the "I confirm the extraction matches the source" step. For propose-mode artifacts, this is the "I read the proposal, evaluated against client X's IPS and household context, modified Y, and recommend Z" step that converts the proposal into a Reg BI recommendation. For decide-mode artifacts, this is the policy-level authorization plus post-action review. The signoff identifies the person (registered representative number for BD-side, CRD for both BD and RIA advisors), the timestamp, the supervisory chain (Rule 3110), and the artifact ID.

Downstream Routing

Where the signed output goes — CRM record update (Wealthbox, Redtail, Salesforce FSC), planning software annotation (RightCapital, eMoney, MoneyGuidePro), trading platform execution (Orion Eclipse, 55ip), client delivery (DocuSign for trade authorization, email for client memo, mailed paper for elderly clients), regulator-facing system (IARD for ADV, FinCEN for SARs, custodian for ACATs/conversion forms), Smarsh / Global Relay archive routing with the tagged-Markdown metadata (L2 Ch8 L2).

Retention and Supervisory Record

The Rule 4511 / SEC Rule 204-2 retention package: the trigger event log, the inputs snapshot, the system prompt version, the user prompt, the AI output, the verification log, the human signoff, the downstream routing confirmation, and the supervisory review entry. Tagged in Smarsh with document_type, household_id, advisor IDs, retention policy, AI tool/model/prompt versions, and related records. Surfaced during quarterly CCO sampling (L4 Ch3 L2), L4 Ch5 ROI dashboard metrics, M&A diligence queries (L4 Ch8), and SEC/FINRA exam responses.

A Concrete Walk-Through — The Hendersons' Roth Conversion Workflow

The Hendersons (the canonical L1 example — married couple, 64 and 62, $2.4M, MAGI ~$148K, $96K Roth conversion window before the 24% bracket cliff, unfunded revocable trust line item) are now in the L3 Ch2 Roth conversion workflow. The handoff diagram for their per-client sizing memo (L3 Ch2 L2 next chapter) reads as follows.

Trigger and input. Quarterly Q4 batch run; inputs are the Hendersons' Wealthbox household record (current age, account inventory, beneficiary status), Holistiplan-extracted prior-year 1040 (AGI $148,000, no IRA distributions), RightCapital plan (Monte Carlo, asset allocation, withdrawal projection), prior IPS edition (2022, flagged "trust funding line item open"), prior Reg BI evidence log (no prior Roth conversion recommendation on file).

AI mode and output. Propose mode. The L3 Ch2 L2 prompt produces a JSON object: { household_id, recommended_conversion_amount, bracket_fill_analysis, NIIT_check, IRMAA_two_year_lookback_projection, state_tax_impact, pro_rata_check_against_aggregated_pre_tax_basis (the IRC §408(d)(2) + §72(e)(8) Form 8606 calculation; not §408(d)(6) which governs divorce transfers), five_year_clock_implications, alternatives_considered_summary }. Tool: RightCapital + custom GPT-4-class prompt with firm system prompt v3.2. Schema: regbi-conversion-memo-schema v2.1.

Verification checkpoint. Source-system: does the recommended_conversion_amount match the gap between Holistiplan AGI ($148K) and the top of the 24% bracket? Does the IRMAA lookback projection use the correct two-year reference year? Does the pro-rata check sum across all the Hendersons' aggregated traditional/SEP/SIMPLE IRA balances? Regulatory: is the Care Obligation citation §240.15l-1(a)(2)(ii)? Are the four alternatives (no conversion, partial conversion, full bracket-fill, multi-year ladder) named and evaluated? Is the IRC citation §408(d)(2) read with §72(e)(8) for pro-rata (not §408(d)(6))? Client-fit: does the Hendersons' IPS support the conversion given their post-2022 stale state? Does the recommendation interact correctly with the unfunded-trust line item still open? Does the recommendation coordinate with the L3 Ch4 L1 Social Security claiming analysis (the higher earner's delayed-to-70 decision)?

Human judgment and signoff. Senior advisor reviews the AI proposal, walks through the three verification tiers, applies judgment on (a) whether to recommend the $96K bracket-fill or a more conservative $74K to also stay below the next IRMAA tier, (b) whether to coordinate the conversion year with delaying SS claim, (c) whether the trust-funding line item needs to be addressed first or in parallel. Senior advisor's documented judgment: "Recommend $74K conversion in 2026 to fill 24% bracket while preserving Tier 2 IRMAA position; coordinate with higher-earner SS delay to 70 (separate memo L3 Ch4 L1); flag trust-funding for IPS update workflow L2 Ch5 L2 with attorney handoff." Signoff: senior advisor name, CRD, timestamp, supervisory chain to CCO.

Downstream routing. Wealthbox household record updated with the recommended conversion (custom field per L2 Ch8 L2 schema); RightCapital plan annotated with the 2026 conversion scenario; Orion Eclipse queued with the December conversion trade-ticket draft; client DocuSign generated for the trade authorization; client memo emailed; Smarsh tagged-Markdown archive routed with document_type=regbi_roth_conversion_memo and full metadata; tax-prep handoff note to the Hendersons' CPA generated for the November tax-planning conversation.

Retention and supervisory record. Smarsh archive package: trigger event (Q4 batch run, date), inputs snapshot (Wealthbox + Holistiplan + RightCapital + IPS + prior Reg BI log), system prompt v3.2, user prompt, AI JSON output, three-tier verification log, senior advisor signoff with judgment narrative, downstream routing confirmation (Wealthbox + RightCapital + Orion + client + CPA), supervisory chain to CCO with quarterly sampling flag. Retention policy: Rule 4511 5-year practical.

The Defensible Reg BI Documentation Point in Every Workflow

Reg BI §240.15l-1 imposes four obligations on broker-dealers and their associated persons making recommendations to retail customers — Disclosure, Care, Conflict, and Compliance. The Investment Advisers Act fiduciary duty parallel — the duty of care and the duty of loyalty articulated in the SEC's 2019 Interpretation — applies to RIAs. The 2025-2026 FINRA AWC pattern on inadequate rollover documentation showed regulators are not relaxing the documentation expectation despite the order-of-magnitude productivity collapse AI enables.

The handoff diagram's signoff node is the singular Reg BI documentation point. Everything before it is preparation, drafting, extraction, or proposal; everything after it is execution, distribution, retention, or supervisory review. The signoff node is the moment Reg BI's four obligations attach. The signoff identifies the responsible person, captures the documented application of judgment, names the alternatives considered (the four-alternatives discipline from L2 Ch7 L2), records the client-specific rationale, and creates the artifact a 2026 FINRA exam will request.

Practices that AI-deploy without diagrammed handoffs typically fall into one of two failure modes. Mode A: the signoff is implicit — the senior advisor reads the AI output, presses "send," and considers the recommendation made; the Rule 4511 retention is incomplete because the documented application of judgment was never captured as a discrete artifact. Mode B: the signoff is performative — the senior advisor's "approval" is a fast click without independent verification; if the AI's underlying calculation was wrong (bracket math, pro-rata aggregation, IRMAA tier, Form 8606 line citation, conflation of §408(d)(2) with §408(d)(6)), the recommendation propagates the error and the Reg BI defense collapses. Both modes are the FINRA AWC pattern. The handoff diagram's explicit verification checkpoint and explicit signoff node defend against both.

The CCO and Supervisory Implication

The CCO's job under FINRA Rule 3110 reasonable-design supervisory obligations and SEC Compliance Rule 206(4)-7 written-policies obligations is to ensure the practice's AI-touched motions have diagrammed handoffs, that the handoffs implement the WSPs accurately, that the principal-review queue (Rule 2210) processes the AI-drafted client content at acceptable risk-based sampling, that the Rule 4511 retention package is complete per workflow, and that exception handling has documented escalation paths. The L4 Ch3 chapter develops the CCO's supervisory architecture in detail; for L3 Ch1 L2's purposes, the takeaway is that the handoff diagram is the artifact the CCO supervises against. Without it, the CCO is supervising a black box. With it, the CCO has discrete checkpoints to sample, discrete decisions to review, and discrete signoffs to audit.

The diagrammed handoff also clarifies the principal-review economics. A 200-household practice running L3 workflows at scale generates hundreds of AI-touched artifacts per month. Risk-based sampling (L4 Ch3 L2's design) prioritizes the artifacts where the diagram shows the highest judgment-bearing weight — propose-mode and decide-mode artifacts get higher sample rates than extract-mode; client-facing draft-mode gets higher sample rates than internal-only; new-tooled workflows get higher sample rates until the diagram-and-output baseline is established. The CCO who tries to sample uniformly across AI-touched content drowns; the CCO who samples against the handoff diagram's risk weights stays current.

The Handoff Discipline and the L3 Chapters Ahead

Every workflow lesson in L3 chapters 2 through 10 implicitly references the handoff diagram for the workflow it describes. L3 Ch2's Roth conversion sequence assumes the per-client sizing memo's diagram has the verification checkpoints and signoff node specified above. L3 Ch3's RMD calendar assumes the population-level extract-and-propose handoff with batch verification and per-household signoff. L3 Ch5's estate-gap memo assumes the document-intake extract-mode handoff feeding the propose-mode gap-and-action-item memo with attorney-handoff signoff. L3 Ch6's equity-comp exercise modeling assumes the extract-from-grant-agreements feeding the propose-with-AMT-and-QSBS-rationale handoff. L3 Ch7's family-structure workflows assume the document-intake-plus-life-event-trigger handoffs with attorney/case-manager coordination signoffs.

The L3 Ch1 L3 lesson (next) takes the three-tier verification pattern (source-system, regulatory, client-fit) and formalizes it as discrete checkpoint specifications inside every multi-step workflow, with Rule 4511 retention baked into the checkpoint chain. The handoff diagram becomes the scaffold; the verification pattern becomes the checkpoint specification. Together they constitute the L3 Ch1 design package every chapter-2-through-10 workflow consumes.

L4 Ch3 then takes the handoff diagrams plus verification specifications and builds the CCO's supervisory architecture on top — the principal-review queue, the risk-based sampling, the exception escalation paths, the agentic-AI WSPs, the incident response. L4 Ch5 measures the workflow's actual delivery against the L3 Ch1 L1 audit's baseline. L4 Ch6 governs the diagram-and-WSP change-control process. L4 Ch8 uses the diagram-WSP-supervision package as the M&A diligence artifact. The handoff diagram is foundational to every later level.

Key Takeaways

• AI operates in one of four handoff modes — draft, extract, propose, decide — and each has a different verification and signoff discipline. Confusing the modes is the L2-to-L3 step-up most advisors trip on.
• The handoff diagram is the operational artifact that turns "we use AI for X" into a defensible Reg BI file, a clean Rule 4511 retention package, and a 2026 SEC exam answer. Six elements: trigger/input, AI mode/output, verification checkpoint, human judgment/signoff node, downstream routing, retention/supervisory record.
• The signoff node is the singular Reg BI documentation point. Everything before is preparation; everything after is execution. The signoff identifies the responsible person under §240.15l-1, captures the documented application of judgment, names alternatives considered, and creates the artifact the 2026 FINRA exam will request.
• Propose-mode is the dominant L3 pattern in May 2026 — recommendation-shaped proposals from AI that the registered person evaluates, modifies, and converts to a Reg BI recommendation by signing. Decide-mode (agentic) deployments are the least-common; most "agentic" wealth deployments are propose-mode with streamlined approval.
• The Hendersons walk-through shows the diagram in practice — propose-mode JSON output (with correct §408(d)(2) + §72(e)(8) pro-rata citation, not the §408(d)(6) divorce confusion), three-tier verification on bracket fill / IRMAA / pro-rata / alternatives / IPS-fit / SS coordination, senior-advisor judgment narrative, signoff, downstream routing to Wealthbox / RightCapital / Orion / client / CPA / Smarsh, Rule 4511 retention package.
• The two failure modes the diagram defends against: implicit signoff (no discrete judgment artifact) and performative signoff (fast click without verification). Both are the FINRA AWC pattern on inadequate rollover documentation; the diagram's explicit verification and signoff defend against both.
• The CCO supervises against the handoff diagram. Risk-based sampling under FINRA Rule 2210 / 3110 prioritizes propose-mode and decide-mode, client-facing draft-mode, and new-tooled workflows. Without diagrams, the CCO supervises a black box; with diagrams, discrete checkpoints to sample.
• Every L3 chapter-2-through-10 workflow assumes the handoff diagram exists for it. L3 Ch1 L3 next formalizes the three-tier verification as discrete checkpoint specifications. L4 Ch3 builds the CCO supervisory architecture on top. L4 Ch8 uses the diagram-WSP-supervision package as the M&A diligence artifact.