Most financial advisors meet generative AI the way they once met option-adjusted spreads or Monte Carlo simulation: through a vendor demo at a custodian conference, a Kitces webinar at 1x speed, or a junior associate's enthusiastic Slack message. The result is a foggy mental model that conflates four very different technologies — language models, classifiers, extractors, and so-called "agentic" actors — into one buzzword the practice owner can neither buy intelligently nor supervise defensibly. This lesson sets the working definition of AI you will use for the rest of your career as a fiduciary, anchored to instruments you already know — and draws the line, hard, between what AI can be trusted to do for a 200-household practice and what remains the exclusive territory of a registered, licensed, accountable human under the Investment Advisers Act of 1940 and Reg BI.

The Tuesday Morning Audit That Forces the Definition

Open Wealthbox on any Tuesday in May 2026. The 9:30 review meeting is the Hendersons: married, 64 and 62, $2.4M between a traditional IRA, a Roth, a joint brokerage, and a 529 for two grandchildren. Last year's 1040 sits in Holistiplan. Their RightCapital plan from December says 91% Monte Carlo. The advisor has roughly twenty minutes to prep, which is the entire premise of the productivity crisis the industry has been writing about for three years: there are too many households, too many planning surfaces, too many regulators, and a single human can no longer mentally hold the working state of more than a handful of relationships at once. Kitces' canonical time-and-task research documents the bucket split — roughly 20% of advisor working hours in client meetings, 36% in meeting prep and planning and servicing, 15% in business development, and 20%+ in admin and management — and the implication every advisor lives with: only a sliver of the week is the real-time advice judgment the client is actually paying for.

So the advisor opens an AI assistant and types: "Given the Hendersons' 1040 last year and their current plan, what are the top three planning conversations for the 9:30?" A useful answer comes back in eleven seconds. The answer mentions Roth conversion windows, a Social Security delay claim for the higher earner, and an unfunded revocable trust line item that was flagged "open" in 2022 and never closed. The advisor scans it, copies the framing, and walks into the meeting two minutes ahead of schedule.

That eleven-second turnaround is the entire point of AI for an advisor — and also the entire risk. To use it without fooling yourself, you need a precise definition of what just happened. The assistant did not plan the meeting. The assistant did not understand the Hendersons. The assistant did not assume any duty to them, did not have access to their account-level positions in real time, did not know whether the IPS allocation was breached by the December rebalance, and did not — in any sense recognized by the SEC, FINRA, or the Investment Advisers Act of 1940 — make a "recommendation." The assistant performed a single, narrow, statistically-driven task: it predicted the most likely next words in a response, given the prompt, the training data, and a temperature setting it inherited from the vendor. Everything you do downstream of that prediction is yours.

The discipline this lesson installs is the habit of holding both facts in your head at once: AI just compressed an hour of prep into eleven seconds, and AI did not do a single thing your Form ADV, your engagement letter, or your fiduciary duty actually obligates you to do. Confuse the two and you become the next AI-washing enforcement headline. Hold them distinct and you have an enormous, defensible productivity lever.

What "AI" Actually Means When You're Selling Advice

The acronym does too much work. When a vendor at a custodian conference uses the phrase "AI," they are almost always describing one of four very different technologies, and the advisor's job is to know which one is sitting behind the marketing slide.

Large Language Models and Generative AI

This is the technology most advisors actually mean by "AI" in 2026 — the GPT-class models from OpenAI, the Claude family from Anthropic, the Gemini family from Google, Mistral, Llama, the smaller open-weight models, and the wealth-vertical wrappers that sit on top of them (Jump, Zocks, FinMate AI, Sybill, Zeplyn, FP Alpha, Holistiplan, Wealth.com, Practifi's agent layer, Salesforce Einstein for FSC). A Large Language Model is, mechanically, a statistical predictor: given a sequence of tokens (roughly, word fragments), it predicts the next token most likely to follow. Run that prediction loop a few thousand times and you get a response that looks like writing. The "generative" part is the loop. The "AI" branding is marketing convenience — the underlying technology has more in common with statistical regression than with the science-fiction conception of a thinking machine.

The analogy advisors already know: think of an LLM the way you think about a Monte Carlo simulation. Both run a stochastic process to produce an output. Both are useful precisely because they sample from a distribution rather than committing to a single deterministic answer. Both fool naive users who treat the output as ground truth. And both are dangerous in the hands of an advisor who doesn't understand the assumptions baked into the model — for Monte Carlo, that's the return distribution, the sequence assumption, and the spending policy; for an LLM, that's the training corpus, the cutoff date, the system prompt the vendor wrote, the temperature, and the explicit prohibition against showing its sources unless you ask in a specific way.

Classifiers

A classifier is a model that sorts inputs into categories. "Is this email a follow-up request or a meeting cancellation?" "Is this transaction likely a duplicate ACH?" "Is this email outbound communication that needs Smarsh archiving under FINRA Rule 4511?" Classifiers have been embedded in advisor tools for a decade — your CRM's lead-scoring engine, your custodian's NIGO detection, your archive's policy engine — and they are usually invisible. They show up in vendor decks under the "AI" label in 2026 because the marketing tailwind made the relabeling free. The technology underneath is often a logistic regression or a gradient-boosted decision tree, neither of which would have raised an advisor's eyebrow in 2018.

Extractors

An extractor is a model that pulls structured fields out of unstructured documents. Holistiplan reading a 1040 line-by-line. FP Alpha pulling agents, trustees, and distribution mechanics out of a revocable trust. Wealth.com extracting beneficiaries, gift provisions, and incapacity language from a will. The category is sometimes called "intelligent document processing" or "OCR-plus." Holistiplan's 10,000-firm install base and FP Alpha's Estate Insights 2.0 are the category benchmarks; both have moved beyond pure OCR into LLM-assisted structured extraction. The output is a JSON object or a populated form — not a paragraph. Extractors are the most boring AI category and arguably the most valuable to a producing advisor, because they collapse three hours of associate work into ninety seconds with relatively low hallucination risk if the source document is clean.

Agentic AI

Agentic AI is an LLM that has been given tools — the ability to read a calendar, send an email, call an API, place a trade, file a form. The model decides, given a goal, which tool to invoke and in what order. The FINRA 2026 Annual Regulatory Oversight Report devoted a section to this category, framing it under Rule 3110's reasonable-design supervisory obligation and reasserting Rule 4511 retention for the artifacts an agent produces. The supervisory implications change the moment an LLM takes an action rather than producing a draft. Agentic tools are the most-hyped and least-deployed category in wealth in May 2026; most advisors who think they're using agentic AI are actually using a generative model that drafts something a human then approves and sends, which is not the same thing under FINRA's framing.

The Four Things AI Does Well, Mapped to Workflows You Already Run

An advisor's working model of AI capability should not be "it's smart" or "it's a copilot." That framing leaks. The accurate working model is four narrow capabilities — summarize, draft, extract, classify — each of which corresponds to a workflow advisors have been doing manually since before the iPhone.

Summarization

Take a 60-page eMoney plan, a 12-page Holistiplan tax scan, a 45-minute Zocks discovery meeting transcript, or a 30-email back-and-forth with a CPA, and produce a one-page brief. This is the workflow that delivered the headline "10+ hours per week saved" finding Zocks publishes, the Schwab 2026 RIA Benchmarking Study data point that AI adoption more than doubled vs. 2023, and the Jump/Zocks category dominance in the AdvisorTech map Kitces publishes every March. The analog advisors already know: this is the associate writing the briefing memo before the senior advisor walks in the room — except the associate works in eleven seconds, never takes a vacation, and never asks for a Slack DM about whether it's appropriate to use ChatGPT for that Tuesday's pre-meeting prep.

Drafting

Produce a first-pass version of a client follow-up email, a quarterly commentary, an IPS section, a Reg BI rollover memo, an ADV Part 2A disclosure paragraph, a client-facing concept memo on QCD, a difficult-conversation message about underperformance. AI drafting is the second-biggest source of the productivity lift, and also the biggest source of regulatory exposure, because every drafted artifact is a marketing communication under Rule 206(4)-1 and/or a recordable communication under FINRA Rule 2210, FINRA Rule 4511, and SEC Rule 204-2. The analog advisors already know: this is the junior associate writing a recommendation memo on the senior's letterhead — except the junior associate has never met the client, has no fiduciary duty, and produces output that confidently invents account numbers, cost bases, and SECURE 2.0 RMD ages when the temperature is set too high.

Extraction

Pull structured fields out of unstructured documents. The wealth use cases are deep and high-value: 1040 line items to a tax-planning matrix, K-1 distributions and pass-through entity facts, trust and will provisions to an estate-gap audit, Form 4 insider filings to an equity-comp picture, 401(k) plan documents to a mega-backdoor Roth eligibility check, brokerage statements to a cost-basis reconciliation, and proxy statement DEF 14A disclosures to an executive-compensation context for a discovery meeting. The analog: an associate spending three hours reading a trust binder. Extractors collapse the time and — crucially — produce auditable, structured output that integrates into RightCapital, eMoney, MoneyGuidePro, Wealthbox, Redtail, Salesforce FSC, and Holistiplan downstream.

Classification

Sort inputs by category. Pattern detection in transaction histories ("is this client suddenly spending materially differently?"), NIGO detection at the custodian, off-channel communication detection in Smarsh and Global Relay, lead scoring at Catchlight or SmartAsset, sentiment analysis on client portal usage, fraud detection on inbound wire requests. Classification is the workhorse the back office has used quietly for years; it shows up under "AI" branding now and is, by far, the lowest-risk category for an advisor to deploy because the output is a discrete label rather than a prose claim a client or regulator can sue you over.

The Five Things AI Cannot Do, Period

This list is the regulatory and fiduciary spine of the lesson. Every item below is a thing the practice will be tempted, eventually, to outsource to an AI tool. Every item is a thing the SEC, FINRA, the state DOIs, the CFP Board, the NAIC, and your E&O carrier will hold you accountable for the day the tool gets one wrong.

Hold a Fiduciary Duty

The Investment Advisers Act of 1940 imposes a fiduciary duty on a registered investment adviser. That duty has two components — a duty of care and a duty of loyalty — articulated in the SEC's 2019 Interpretation Regarding Standard of Conduct for Investment Advisers. Neither component can be discharged by a machine. The "duty of care" requires the adviser to provide advice in the best interest of the client based on the client's objectives and to seek best execution and to provide advice and monitoring over the course of the relationship. The "duty of loyalty" requires the adviser to eliminate or make full and fair disclosure of conflicts. A language model has no client, no objectives, no relationship, no awareness of conflicts, and no capacity to disclose. The fiduciary remains the registered human signing the ADV. Any vendor pitch that hints at AI "taking on the fiduciary burden" is, by definition, marketing language that does not survive the Advisers Act.

Exercise Professional Judgment

Professional judgment is the act of weighing facts a client has shared, facts the model cannot know, and facts no one has yet articulated — and deciding what to recommend, what to defer, and what to push back on. It is the part of the workflow CFP Board's Code and Standards and FINRA's Reg BI Care Obligation under §240.15l-1(a)(2)(ii) most directly govern. An LLM can produce a recommendation-shaped paragraph; it cannot exercise judgment. The output looks like judgment to the untrained reader and is, in fact, a statistical interpolation from training data combined with the advisor's prompt. The advisor's job is to read the output, apply judgment, and either accept, modify, or reject — never to deploy the output verbatim. The Cardinal Rule (L1 Ch2.3) installs the verification protocol that makes this discipline routine; this lesson establishes why it is non-negotiable.

Take Responsibility for a Reg BI Recommendation

Reg BI's four obligations — Disclosure, Care, Conflict, and Compliance, codified at 17 CFR §240.15l-1 — apply to "recommendations" made by a broker-dealer or its associated persons to a retail customer regarding securities transactions or investment strategies involving securities. The 2025-2026 FINRA AWC pattern of inadequate rollover Reg BI documentation is the canonical enforcement example. The four documented alternatives in a rollover (leave in plan, roll to new employer plan, roll to IRA, take cash) must each be considered and documented by the registered person, not by the tool that drafted the memo. An AI-generated Reg BI memo is a draft; the registered person's review, edit, and signoff is the recommendation. Confuse the two and the next FINRA AWC will reference your firm.

Source or Cite Reliably Without Human Verification

LLMs hallucinate. The polite framing in the industry is "confabulation" — the model produces a confident, well-formed statement that has no factual basis in its training data. The wealth-specific hallucination gallery is well-known by May 2026 and includes: an invented account number on a trade authorization, a fabricated cost basis on a tax-loss harvest, the wrong RMD starting age (the SECURE 2.0 confusion between 70½, 72, 73, and the 75 step-up in 2033 is the most-cited example), a made-up FINRA rule citation, a plausible-sounding but inaccurate Social Security Primary Insurance Amount calculation, hallucinated trust language that doesn't actually appear in the document the model was given, and confident citations to "FINRA Rule 2210(d)(4)(B)" or "IRC §408(d)(6)" used wrongly (the latter governs IRA transfers incident to divorce, not the backdoor-Roth pro-rata rule, which lives at IRC §408(d)(2) read with §72(e)(8) and is reported on Form 8606). Every cite in an AI-produced artifact requires source verification before the artifact leaves the advisor's screen.

Keep Client NPI Secret by Default

Nonpublic personal information of an advisory client is protected under Regulation S-P (17 CFR Part 248), the GLBA Safeguards Rule, and an increasing patchwork of state laws (NY DFS 23 NYCRR 500, California CPRA, Texas DIR, and the 2024-2026 wave of state insurance-commissioner AI bulletins anchored to NAIC Model #275). The May 2024 Reg S-P amendments tightened the obligations: covered entities must adopt a written incident response program, notify affected individuals within 30 days of a breach involving sensitive customer information, and conduct oversight of service providers. Pasting a client's name, SSN, account number, or DOB into a free public LLM is, in the practical reading of the rule, an information-sharing event with a third-party service provider that does not have an executed agreement with your firm. The model will not protect you. Reg S-P, GLBA, NY DFS Part 500, and your own ADV Part 2A disclosure obligations will. Lesson L1 Ch5.2 and L4 Ch4.1 develop the practical NPI handling rules; this lesson establishes the principle.

Why the Difference Matters Under Reg BI, the Marketing Rule, and Fiduciary Duty

The five "cannot do" items are not abstract. Each maps to a specific regulatory regime and a specific enforcement vector that has produced 2024-2026 actions.

The 2024-2025 SEC AI-washing settlements — Delphia, Global Predictions, and the broader 2025 enforcement cluster — turned on advisers' marketing language overstating AI capability or misrepresenting how the firm used AI. The SEC Division of Examinations Risk Alerts on Marketing Rule compliance, the January 2026 staff FAQs on third-party ratings, hypothetical performance, and testimonial mechanics, and the SEC's repeated re-publications of "Additional Observations" all point to one operating reality: every claim about AI capability is a marketing communication, and the "clear and prominent" disclosure standard of Rule 206(4)-1 governs it. An advisor who says "our AI manages your portfolio" when the AI in fact drafts an internal memo a human reviews is misrepresenting capability and walking directly into the Marketing Rule.

The 2025-2026 FINRA AWC pattern on inadequate Reg BI rollover documentation turned on broker-dealers failing to document consideration of reasonably available alternatives. AI tools that generate rollover memos collapse the time cost of producing those memos by an order of magnitude; the regulators are reading them as carefully as ever. A registered person who signs an AI-drafted rollover memo without independently verifying that all four alternatives were considered — leave in plan, roll to new employer plan, roll to IRA, take cash — is exactly the named-respondent profile in the AWCs.

The FINRA 2026 Regulatory Oversight Report's framing of agentic AI under Rule 3110 reasonable-design supervisory obligations made one principle explicit: when AI takes action, the supervisory architecture must change. The CCO's WSPs need a discrete section on AI use, the principal review queue under Rule 2210 needs to handle AI-drafted content at volume, and the Rule 4511 retention obligation extends to the prompts, the outputs, the human edits, and the signoff trail. None of this is solved by the vendor's product datasheet. All of it is solved by the practice's policy, written by a human, supervised by a human, and capable of being explained to a 2026 SEC examiner in plain English.

The CFP Board's Code and Standards, the Investment Advisers Act fiduciary duty, the NAIC AI Model Bulletin and Model #275 for the annuity-licensed advisor, and the state-by-state cyber and AI rules (NY DFS 23 NYCRR 500, California CPRA, Colorado SB 21-169 on insurance models, and the broader state DOI patchwork) all converge on the same point: the registered, licensed human owns the recommendation. The tool produces drafts. The drafts are not the recommendation. The drafts are records. The records get retained. The recommendation is yours.

The Working Mental Model You Will Use for the Rest of This Program

Here is the mental model to install before Lesson 2 begins:

AI in a wealth practice is four narrow capabilities — summarize, draft, extract, classify — wrapped in marketing that overstates them. The four capabilities collapse the time cost of meeting prep, follow-up, planning extraction, and pattern detection by roughly an order of magnitude. The collapse is real; the productivity studies (Schwab 2026, Zocks 10+ hours/week, the Kitces AdvisorTech map's category dominance for Jump and Zocks) are not vendor marketing — they reflect what actually happens when an advisor with a strong workflow plugs the tools in.

What AI does not do — at all, ever, in 2026 or 2030 under any plausible regulatory regime — is hold the fiduciary duty, exercise professional judgment, take Reg BI responsibility, cite reliably without verification, or keep NPI secret by default. These five "cannots" map to the SEC Marketing Rule, Reg BI, FINRA Rules 2210/3110/4511, Reg S-P, and the state cyber/AI patchwork. The registered, licensed, accountable human owns each one.

The discipline that follows from the model is straightforward: every AI-touched artifact gets a verification pass, a documented human review, an explicit recommendation owner, and a retained record. The Cardinal Rule (L1 Ch2.3) is the operational form of that discipline. The chapters that follow this one will build the regulatory layer (Ch4), the tool-by-tool map (Ch3), and the personal-accountability layer (Ch5) on top of it.

You will not, after this lesson, be able to deploy a Roth conversion screen across fifty households (that is L3 Ch2). You will not be able to draft a Reg BI-compliant rollover memo from a Zocks transcript (L2 Ch7.2). You will not be able to write WSPs for agentic AI under Rule 3110 (L4 Ch3.3). What you will be able to do is hold a clean, accurate, vendor-pitch-resistant working definition of AI in your head — the working definition that lets you walk into any custodian conference, any compliance training, any prospect's question, or any examiner's interview, and answer the question "what does AI do at your firm?" without flinching, without overclaiming, and without setting the firm up for the next AI-washing enforcement headline.

How to Explain This to a 70-Year-Old Client Tomorrow

The L1 exit state is that you can explain GenAI accurately to a client, a CCO, and a junior associate. Here is the 90-second script for the client, refined over hundreds of advisor-client conversations and now sitting in firm playbooks across the wirehouse and RIA channels:

"You'll see me using AI tools in two places this year. The first is for prep. When you and I get on a review call, I'm using an AI tool to pull a summary of your accounts, your last meeting, and your tax return so I walk in already up to speed. It saves me an hour and gives you a more focused meeting. The second is for follow-up. When we end a meeting, the same AI drafts the follow-up email and the trade authorization in a few seconds, and then I review and edit before anything goes out. The AI never sends anything to you, never makes a recommendation for you, and never replaces my judgment about what's right for your household. Every recommendation in your file has my name on it, and the same fiduciary standard applies whether I typed the first draft or an AI did. Your information is protected — we use vendor-hosted AI under our security policies and our SEC privacy obligations, not free public chatbots. If anything ever changes about that, I'll update our agreement and tell you directly."

That paragraph is the deliverable for the L1 capstone (a one-page personal AI Use Policy) restated for a client audience. It is short on purpose. It is accurate. It is consistent with what the SEC, FINRA, and your state regulator are asking you to disclose. It is what survives.

Key Takeaways

• AI in 2026 is four narrow capabilities — summarize, draft, extract, classify — wrapped in marketing that overstates them. Hold the four-category model in your head and you will be able to read a vendor deck in 30 seconds.
• The productivity collapse is real. Schwab's 2026 RIA study and Zocks' published data on 10+ hours/week saved aren't vendor fluff — they reflect what happens when meeting AI, planning AI, and extraction AI plug into a 200-household practice.
• AI cannot hold the fiduciary duty under the Investment Advisers Act of 1940, cannot exercise professional judgment under the CFP Code or Reg BI Care Obligation, cannot take responsibility for a Reg BI recommendation under §240.15l-1, cannot cite reliably without human verification, and cannot keep client NPI secret by default under Reg S-P.
• Every AI-touched artifact is a regulated artifact: a marketing communication under Rule 206(4)-1, a recordable communication under FINRA Rule 2210, and a retained record under FINRA Rule 4511 and SEC Rule 204-2. Treat every prompt, output, edit, and signoff as part of the record.
• The registered, licensed, accountable human owns the recommendation. The tool drafts. The human signs. The 90-second client explanation in the prior section is the deliverable that captures this for the L1 capstone.
• The Cardinal Rule (next chapter) is the operational form of this principle: source-system verification, regulatory verification, client-fit verification — applied to every AI-touched artifact before it leaves your screen.