Twenty Minutes a Week

01Do the division first

Carson Group built an internal AI assistant and reported that it saves five minutes per search and roughly five thousand hours a year across the firm. Five thousand hours sounds like a new department. Divided across the advisor base and the working year, it is about twenty minutes per advisor per week.

Twenty minutes a week does not become a new client. It becomes a slightly less rushed Tuesday.

This is not a criticism of Carson, whose disclosure was more specific and more honest than most. It is a criticism of the unit. Efficiency claims in this category are almost always expressed as hours per firm per year, which is the presentation that makes them sound largest. The only unit that supports a decision is minutes per advisor per week — because that is the unit in which an advisor either does or does not have room for another conversation. Before evaluating any tool in this piece, convert its claim.

Self-reported time savings are the weakest evidence class there is

The most-cited industry figure for AI efficiency in wealth management comes from F2 Strategy's 2026 survey: 68% of firms reported gaining 25% more efficiency in targeted workflows. That number deserves three qualifications, all of which are in the survey itself and none of which survive into the way it gets quoted.

It is 68% of the subset of firms that measure their AI investments at all — a denominator selected on the dependent variable, since firms that bothered to measure are disproportionately firms that got a result. The sample is forty firms. And it is "targeted workflows," not firm-wide. F2's own co-founder is blunter than the headline: "We're seeing a very loose correlation in 2026 between firms' spend on both AI technology and its tokens and a meaningful measurable value in a classic sense to the business."

Then there is the deeper problem. A randomized controlled trial run in early 2025 gave sixteen experienced open-source developers AI tools on real issues in their own repositories — projects averaging over a million lines of code that they had contributed to for years. They were 19% slower. Asked afterward, they estimated the tools had made them about 20% faster. A thirty-nine point gap between measured and perceived productivity, under randomization, in a population unusually numerate about its own workflow.

That study has since been qualified by its own authors, and the qualification is arguably more useful than the headline. In February 2026 METR reported newer data: a −18% estimate for the original cohort with a confidence interval running from −38% to +9%, and −4% for newly recruited developers with an interval from −15% to +9%. Both cross zero. More importantly, they are redesigning the experiment because AI adoption has made clean measurement hard — developers now decline to participate if they might be assigned to the no-AI arm, and self-report steering away from the tasks where AI helps most. The people most helped are selecting themselves out of the study.

So the durable finding is not the point estimate. It is the gap between what was measured and what was believed, plus the fact that the best-designed experiment in the field has been destabilized by the very adoption it was trying to measure. Every efficiency statistic in wealth management — including F2's — is self-reported, and none of them has anything like this level of methodological scrutiny behind it.

The practitioners closest to the problem already say so

The RIA Edge 100 — the largest and most sophisticated independent firms in the country — were asked in 2026 whether AI would change the number of clients an advisor can serve. They said no. Form ADV data puts the cohort at roughly seventy client accounts per advisor, and the ratio is roughly flat against prior years. WealthManagement.com's framing for the phenomenon was "shadow efficiencies" — real improvements that make an advisor more effective without changing coverage.

"I don't think you're going to find 25% efficiency, at least in our model."Carrie Delgott, president, COO and CCO, Wescott Financial Advisory Group — responding to precisely the figure quoted above. Her framing for the alternative: "deeper, not wider."

This is the finding that reorganizes the whole analysis. It is not that efficiency fails to materialize. It is that the firms achieving it have decided not to convert it into revenue — they are spending it on service depth, proactive planning, and advisor quality of life. Absorbed slack by choice, not by failure. One executive at a firm serving ultra-high-net-worth families keeps deliberately to ten to fifteen relationships per advisor so the team stays "one call away." No amount of meeting-note automation will change that number, because the number is a strategy, not a constraint.

F2 found the same thing from the buy side: firms are prioritizing "operational efficiency, advisor productivity, and headcount-neutral scale." Headcount-neutral scale means buying AI in order not to hire. That is a real and defensible return. It is simply not the return in the pitch deck, and it is calculated completely differently.

The one thing we know converts

There is a natural experiment in the advisor-productivity literature, and it predates AI. Advisors who add a paraplanner cut their own hours per client from roughly fourteen to ten, while total team hours per client rise from fourteen to twenty-two. Client counts go from about seventy-three to a hundred and twenty. But the striking number is revenue per client: it nearly doubles, from roughly $1,531 to $2,850.

When back-office load was genuinely offloaded, the gain showed up as moving upmarket, not as serving more households at the same price. That is a specific, testable prediction for AI adoption, and it is the pathway the calculator below treats as most plausible.

02The model

Four numbers determine whether an advisor AI deployment returns anything: how much time it actually saves, how much of that survives your data environment, how much of what survives gets redeployed into revenue-generating work, and what an hour of that work is worth. Vendor ROI models set the middle two to 100% silently. This one exposes them.

Start with a firm size, then a scenario. The gap between "vendor claim" and "measured evidence" on identical firm inputs is the point of the exercise.

03The jagged frontier, category by category

The best-designed study of AI and professional knowledge work gave 758 BCG consultants eighteen realistic tasks. On tasks inside what the authors called the "jagged technological frontier," the AI-assisted group completed 12.2% more tasks, 25.1% faster, at higher quality. On tasks outside it — tasks that looked similar and felt similar — they were nineteen percentage points less likely to reach the correct answer.

The frontier is jagged because its boundary does not track apparent difficulty. That is the governing fact for an advisory practice: summarizing a client meeting and judging whether a recommendation is suitable feel like adjacent tasks and are on opposite sides of the line. Advisors adopting these tools cannot see the boundary from the inside.

The market has, so far, located it correctly by instinct. Advisors report using AI for meeting summaries at 31%, CRM updates 28%, meeting prep 26%, routine client communications 25% — and AI-generated financial recommendations at 3%. Ninety-three percent insist on retaining control of decisions and advice. Forty-six percent lack confidence in AI outputs. The category has found its edge without anyone drawing it.

04The data layer is a gate, not a nice-to-have

FinanceBench asked a frontier model real questions against real SEC filings under different retrieval conditions. Handed the correct evidence, GPT-4-Turbo answered 85% correctly. Made to find the answer itself in one company's filings, 50%. Made to find it in a shared store containing every company's filings, 19%.

Retrieval scope alone — not model quality, not prompting, not reasoning — moved accuracy thirty-one points. And a wealth firm running fifteen to twenty-five disconnected systems is the shared-vector-store condition. It is the worst case in that experiment, by construction.

One nuance worth carrying, because it changes what the risk looks like in practice: in the shared-store condition the failures were dominated by refusal — 68% of questions declined, against 13% answered incorrectly. A model that cannot find the evidence mostly declines rather than fabricating. That is the better failure mode, and it means the operational symptom of a bad data layer is not a stream of confident errors but a tool that quietly stops being useful, which is much easier to tolerate and much harder to notice.

This is the mechanism underneath every adoption-versus-outcome gap in the survey data. EY: 95% of wealth and asset managers have scaled AI to multiple use cases, 27% report substantial impact. Schwab, from an independent 533-advisor study fielded by Logica: 63% using AI, roughly 10% integrating it into business strategy. Two different researchers, different samples, a year apart, converging on the same shape. F2 supplies the reason: 64% of wealth firms and 83% of bank and trust respondents lack a unified data layer.

Advisor360's advisors have named "bad data and lack of integrated tools" their top technology complaint for three consecutive years — before, during, and after the generative AI wave. The constraint is not new and AI did not create it. AI simply made it expensive.

What "unified data layer" actually means

Eight component systems, at minimum: CRM, portfolio accounting and performance, custodian feeds, financial planning, document management, billing, held-away aggregation, and market data. Connecting them is not primarily an API problem. It is an identity resolution problem — matching the same household across systems that disagree about what a household is, and mapping accounts to households to contacts to service tiers. That is the piece that gets under-budgeted, and it is the piece that determines whether retrieval lands in the 85% condition or the 19% one.

Published implementation timelines, all vendor-sourced and converging across two independent providers: a native CRM accelerator with two or three custodians, four to eight weeks; middleware, eight to sixteen weeks; custom API builds, three to six months; full custom engineering, six to eighteen months with a permanent team. Cost figures are thin and entirely vendor-published — one puts custom integration at $50,000 to $500,000 per system pair, which for a six-system stack is fifteen pairs if built point-to-point. No independent benchmark for total data-layer cost in wealth management appears to exist.

The most revealing datapoint is what Orion actually built when it built "AI." In its own description: an automated data pipeline drawing from CRMs, planning systems and client data lakes into Orion's lake. The AI is the interface. The pipeline is the product. A platform vendor spending at that scale concluded the same thing this section argues — and then sold it back as an AI story.

05What attaches the moment you switch it on

Five obligations activate when an advisor AI tool goes live. Two are well-defined, two are genuinely unanswered by any regulator, and one — the one nobody in the securities world is watching — is currently being litigated as a class action.

Supervision: defined, and broader than firms assume

FINRA's position is that its rules are technology-neutral and apply unchanged. Under Rule 3110, a supervisory system's policies "should address technology governance, including model risk management, data privacy and integrity, reliability and accuracy of the AI model." The sleeper clause is scope: FINRA rules apply "whether member firms are directly developing Gen AI tools for their proprietary use or when leveraging the technology of a third party, including through embedded features in existing third-party products."

Embedded features means Copilot in Outlook, the AI summarizer your CRM shipped last quarter, Zoom's meeting assistant — capability nobody procured as "AI" and nobody put through a governance review. For most firms that is the largest uninventoried AI surface they have. FINRA's 2026 effective-practices list is concrete about the remedy: robust pre-deployment testing, ongoing monitoring of prompts and outputs, "storing prompt and output logs for accountability and troubleshooting," "tracking which model version was used and when," and human-in-the-loop validation.

For registered advisers, Rule 206(4)-7 requires written policies reasonably designed to prevent violations, reviewed annually. There is no SEC rule, release, or staff guidance prescribing AI-specific content for those policies. The only signal is enforcement: the March 2024 AI-washing orders charged the compliance rule alongside the Marketing Rule, so the Commission's theory is that making AI claims without policies to test them is itself a compliance failure. AI has now been an examination priority three years running.

Communications: the rule is being rewritten because AI broke it

Rule 2210 applies to AI output — FINRA's advertising FAQ, updated December 2025, says firms "are responsible for their communications, regardless of whether they are generated by a human or AI technology." A bespoke meeting summary sent to one client is correspondence, subject to risk-based review rather than principal pre-approval. A templated AI recap sent to more than twenty-five retail investors in thirty days is a retail communication requiring principal approval before use. The template is the communication, not the individual send — which is exactly the distinction an automated workflow is designed to blur.

FINRA has conceded the problem. Regulatory Notice 26-14, published 9 July 2026, proposes to dismantle blanket principal pre-approval, stating that applying it to AI-generated retail communications "can be challenging" and that the difficulty is "compounded by the potential speed and volume of AI-generated communications." The replacement would require firms to write their own procedures determining which categories need pre-approval. It also offers the sector's clearest statement of the standard: "Gen AI communication tools can be part of a reasonably designed system, provided they are vetted, tested and monitored."

Books and records: genuinely unanswered

Is the prompt a record? Is the output? There is no SEC or FINRA guidance. That silence is the finding, and the practitioner analysis has moved well ahead of the regulators.

The conventional test is transmission: AI-generated content that is never sent is probably not a written communication requiring retention; content transmitted by email or chat triggers retention if the subject matter fits. The better analysis notes that transmission answers the wrong question — "Transmission tells you whether you have a communications problem; it does not tell you whether you have a records problem elsewhere." An untransmitted transcript can still be a required record under provisions that have nothing to do with communications: order memoranda, code-of-ethics violations and responsive action, performance-calculation support, annual compliance review documentation. A notetaker running in an investment committee meeting or a compliance interview can manufacture a required record that would not otherwise exist.

Two practical consequences follow, and the second is counterintuitive. Decide at the category level which meeting types may be recorded, before deployment rather than case by case — and prohibit notetakers outright on certain call types, such as calls with counsel and compliance interviews. And recognize that retaining everything is not the conservative choice: surplus transcripts create examination and discovery exposure that would not otherwise exist.

The market is pricing prompts as records ahead of the regulators. Commercial capture products now exist that ingest prompts and responses from enterprise AI platforms into the same archive as email. The off-channel communications sweep — over 100 firms and more than $2 billion in penalties — is the cautionary backdrop, and "shadow AI" on personal devices is reported by one examiner-facing consultant to occur at virtually every RIA examined.

Recording consent: the exposure nobody in securities is watching

No SEC or FINRA guidance addresses meeting-recording consent. Neither the AI notice, nor the 2026 oversight report, nor the examination priorities mentions it. The exposure is private civil litigation and state law, which is precisely why compliance departments oriented toward securities regulators are missing it.

In re Otter.AI Privacy Litigation is consolidated in the Northern District of California, pleading the federal Wiretap Act, California's Invasion of Privacy Act, Illinois's Biometric Information Privacy Act and more. The core allegation is architectural rather than incidental: the notetaker "seeks permission only from meeting hosts... but not from all participants." That host-only consent model is the industry standard design. Speaker diarization — how a transcript labels who said what — arguably extracts a voiceprint, which creates biometric exposure independent of the wiretap claim.

Two points practitioners routinely get wrong. First, any list of "two-party consent states" you have seen is probably wrong. Roughly a dozen states impose an all-party or anti-secret-recording rule, but at least six are medium-dependent or judicially narrowed: Connecticut is all-party for telephone and one-party in person; Oregon is the inverse; Michigan reads all-party on its face but courts have recognized a participant exception since 1982; Nevada and Delaware are contested. Illinois — the state everyone cites — requires the recording to be surreptitious and the conversation to carry a reasonable expectation of privacy, so a disclosed, visible notetaker is likely fine there. Massachusetts is a secrecy statute rather than a consent statute, and is the harshest in practice.

Second, on a multi-state video call, the location of the advisor, the firm and the server is irrelevant. The controlling authority is a brokerage case: California's Supreme Court held that California law governed recordings made by a firm's Atlanta branch of calls with California-resident clients, because California's interest in resident privacy would otherwise be severely impaired. If any participant is physically in an all-party state, assume that state's law applies. California statutory damages run $5,000 per violation; Illinois biometric damages are $1,000 to $5,000 per violation and accrue per scan.

Vendor practice converges on disclosure in the calendar invite, a verbal announcement, capturing the affirmative response inside the transcript, and switching off on objection. Some wealth-specific vendors have moved to no-recording, text-only architectures explicitly citing the Otter litigation. None implements a true per-participant consent gate. That is an unremediated industry-wide exposure, and the firm — not the vendor — is the one with the client relationship.

Privacy: the clause your AI vendor will not sign

The amended Regulation S-P requires written policies obliging service providers to notify the firm "as soon as possible, but no later than 72 hours after becoming aware that a breach in security has occurred." Compliance dates were December 2025 for larger entities and June 2026 for smaller ones — both now passed.

This is the practical blocker in AI procurement, and it is asymmetric: large model providers' standard terms do not generally commit to 72-hour breach notice, and a mid-size RIA has no leverage to negotiate one. The firm must paper it or document an exception. Separately, disclosing client information to an AI vendor without an opt-out depends on the service-provider exception, which requires contractual confidentiality and purpose limitation. If the vendor trains on your data, the purpose-limitation prong is arguably broken. No regulator has confirmed or rejected that reading. Vendors that state plainly they do not train on client data — several wealth-specific ones do — are materially easier to diligence, and that single contractual term is worth more than any feature comparison.

06Build versus buy

Purpose-built advisor AI runs roughly $800 to $2,400 per seat per year at list price. A hundred-advisor firm buys the category for something between $80,000 and $240,000 annually. Savant Wealth Management, at 727 employees, has disclosed spending $15–20 million a year — roughly $50 million over three years — building its own data warehouse and internal AI. Its stated goal is to quadruple revenue while only doubling headcount.

Those two numbers are not close enough for cost to be the deciding factor. The build case cannot be about cost. It has to be about differentiation or data control, and a firm should be made to say which — out loud, in writing, before the first hire.

Three things the build case consistently gets wrong

The model is the cheap part. In published enterprise retrieval-application costings, engineering and data preparation account for 50–60% of budget and model API calls for 5–15%. Most firms budget the inverse. The expensive work is connectors, identity resolution, document cleanup and evaluation — precisely the work in §04, which is why a build that skips the data layer produces the 19% condition at ten times the price.

Maintenance is what kills mid-size builds, not construction. The canonical case is Commonwealth Financial Network, which built its advisor platform, then spun it off as a separate company in 2019 because ongoing maintenance costs could not be carried without more scale. They built it, it worked, and they still could not keep it. Generic benchmarks put annual maintenance at 15–20% of initial build cost, permanently.

The famous builds are mostly not builds. Morgan Stanley did not build a model — it built retrieval, evaluations and workflow on a vendor model, growing an internal corpus from 7,000 to 100,000 documents and reaching 98% adoption among advisor teams. Bank of America's Meeting Journey ran on infrastructure it had already built for Erica since 2018; the marginal build cost was low because the platform existed. Both are correctly described as assembly, not construction. Neither has ever disclosed team size, budget, or timeline — so any mid-size firm modelling itself on them is modelling against numbers that do not exist publicly.

What is genuinely transferable from Morgan Stanley is not the architecture but the discipline: purpose-built evaluation sets graded by advisors and prompt engineers, and daily regression testing against a question suite. That is an operating cost, forever, and it is the line most build models omit entirely. It is also, as Part II found, the only published evaluation methodology anywhere in the sector — and it belongs to a build, not a purchase.

The decision, by firm size

07Vendor risk — and a correction to the consensus

The standard advice on this category has been that standalone AI notetakers are doomed, because CRMs and custodians will absorb the feature and give it away. The absorption is real: Nitrogen bundled statement parsing and proposal drafting into existing subscriptions at no additional cost in February 2026, Morningstar embedded the same functions into its advisory suite in March, and Robinhood's Cortex for Advisors arrived free to advisors custodying on TradePMR in June.

But the prediction failed. CRM response was slow — one major CRM did not ship a notetaker until autumn 2025 and another still lacks one. The specialists won on speed, and the assessment by mid-2026 had inverted: it is now the CRMs that look threatened by the standalone tools rather than the other way around. Jump reached roughly 27,000 advisors and raised $80 million; Zocks raised $45 million; both are expanding into onboarding, account opening, analytics and client data — becoming systems of record.

So the real vendor risk is not the one everyone is watching. It is not your AI vendor gets absorbed. It is your AI vendor becomes your system of record and you never decided that it should. A tool bought for meeting notes at $100 a seat, two module upgrades later, is holding your client data model. That is a governance decision being made by procurement.

Where absorption has bitten is the proposal and statement-parsing category, where three major platforms now give away what standalone vendors charge for. Thinly capitalized point solutions in that lane are the ones to diligence hardest.

What to actually ask a vendor

Two useful screens do not yet work. SOC 2 Type II is now table stakes across the notetaker category and therefore differentiates nothing. And no advisor-AI vendor we could find publishes ISO 42001 certification — the AI management-system standard has not penetrated this market, so buyers cannot use it as a filter yet.

08Sequencing

Ordered so that each step produces the artifact the next step needs, and gated so that nothing is deployed into a data environment that cannot support it.

1. 1Inventory what you already haveEvery embedded AI feature in your CRM, productivity suite, meeting platform and custodian portal — the ones nobody procured. FINRA's rules reach embedded third-party features explicitly, and shadow AI is reported at virtually every RIA examined.Artifact: an AI use inventory. Without it you cannot write a defensible policy, and it is the first thing an examiner asks for.
2. 2Write the consent and recording policy before the first tool goes onDecide at the category level which meeting types may be recorded and which may never be — counsel calls and compliance interviews in the second bucket. Draft the verbal script; capture affirmation inside the transcript.Gate: this is a legal exposure with active class litigation and no securities-regulator guidance. It is cheaper to write than to unwind.
3. 3Deploy one inside-the-frontier tool and measure in minutes per advisor per weekMeeting capture is the clearest win. Run it for two quarters. Measure against a pre-recorded baseline, not against recollection — §01 is what happens otherwise.Gate: do not add a second tool until the first has a measured number.
4. 4Decide, explicitly, what the freed time is forDeeper service, more households, upmarket movement, or not hiring. Each has a different measurement, and the last one is not a revenue calculation at all. Unassigned freed time is absorbed by default — that is the finding, not a risk.Artifact: a written conversion mechanism. This is the input the calculator in §02 cannot supply for you.
5. 5Build the data layer before buying anything that depends on itNext-best-action, cross-system analytics and anything agentic fail without it — and fail expensively, because the failure looks like plausible output. Start with identity resolution across CRM, portfolio accounting and custodian feeds.Gate: this is the step that separates firms that get value from firms that get invoices. It is also the asset that survives whichever vendors do not.
6. 6Stand up supervision proportionate to what you deployedPrompt and output logging, model-version tracking, human review before client contact, and a review procedure for AI-assisted communications. Then re-read Rule 2210 against your templated outputs, since the template is the communication.Timing: comment on FINRA Regulatory Notice 26-14 by 11 September 2026 if you will be operating at scale.
7. 7Only now consider buildingAnd only if you can name the differentiation or data-control reason, and can fund evaluations and maintenance permanently at 15–20% of build cost per year.Test: if the answer to "why build" is "it will be cheaper," the answer is no.

09What we could not verify

Advisor time-allocation figures carry an additional caveat worth stating: the canonical study's activity detail sums to roughly 53 hours per week while the same advisors self-report 43. Every derived per-hour figure moves about 19% depending on which base you use.

10Next

Part IV takes up the question this piece keeps deferring: what happens when the model stops drafting and starts acting. The engineering problem is the one every firm meets at the edge of §03's frontier — how an investment policy statement becomes a machine-readable constraint set the agent cannot argue with: tracking-error budgets, VaR caps, concentration limits, drawdown triggers, and the position limits, rate limits and circuit breakers that Part II found missing from every published agentic retail product.

Part V is the validation playbook — evaluation design, backtest hygiene, champion–challenger promotion, shadow deployment, drift monitoring, and treating a foundation-model version bump as the model change FINRA already says you should be tracking.

11Sources

Advisor economics & productivity

1. Kitces Research on Advisor Productivity — "How Do Financial Advisors Actually Spend Their Time?" (n>1,000 lead advisors); 3rd edition, Sept 2022; 4th edition, Sept 2025 (2024 data)
2. Kitces Advisor Marketing Survey, July 2024 (n≈1,000 firms) — client acquisition cost, and the finding that 71% of CAC is advisor and employee time
3. Kitces AdvisorTech research, Future Proof, Sept 2025 — firms focusing technology on cost efficiency are 40% less productive
4. Kitces, "Best AI Notetakers For Financial Advisor Meetings" (fieldwork Fall 2024) — adoption and satisfaction by tool
5. Charles Schwab RIA Benchmarking Study, 2024, 2025 and 2026 editions
6. Charles Schwab 2025 RIA Compensation Report (n=1,046 firms, 16,000+ employees)
7. Schwab / Logica Research, "2026 RIA & AI Research Study: Advisor AI in Action" (n=533, fielded Oct 2025)
8. The Ensemble Practice, 2025 Careers & Compensation Survey (n=230 RIAs)
9. Cerulli Associates — advisor time allocation and client coverage ratios (2022 and 2025 editions; exhibits paywalled)
10. Morningstar behavioral research on meeting frequency (n=399 advisory clients)
11. WealthManagement.com, "AI Won't Alter Client-Advisor Ratios but Create 'Shadow Efficiencies,' Say Edge 100 RIAs," 2026
12. InvestmentNews, "RIAs are building and investing in data warehouses and in-house AI tools," 2026 — Savant, Carson, Commonwealth
13. InvestmentNews, "AI in wealth management: budgets surge but ROI remains elusive," 24 July 2026 — F2 Strategy Q2 2026 Trend Report (n=40, $8.6T)
14. WealthManagement.com, "Plancorp Saves 700 Hours a Month With AI," 3 April 2026 (28-day study, 60+ staff, vendor-partnered)
15. EY, GenAI in Wealth & Asset Management Survey 2025 (n=100 firms)
16. Advisor360°, Connected Wealth Report 2025 (Coleman Parkes, n=300) and 2026 (FUSE Research, n=301) — note the research vendors differ; the two years are not comparable

Productivity evidence & AI performance

17. Dell'Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon & Lakhani, "Navigating the Jagged Technological Frontier," Organization Science, 2025 (n=758 BCG consultants)
18. Brynjolfsson, Li & Raymond, "Generative AI at Work," Quarterly Journal of Economics 140(2), 2025 (n=5,179 support agents)
19. METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," arXiv:2507.09089 — the 19%-slower RCT
20. Islam, Kannappan, Kiela, Qian, Scherrer & Vidgen, "FinanceBench," arXiv:2311.11944 (Patronus AI, Contextual AI, Stanford) — 85% oracle / 50% single store / 19% shared store
21. MIT NANDA, "The GenAI Divide: State of AI in Business 2025" — the widely-misquoted 95% figure; not peer-reviewed, not a random sample
22. Milemarker and Vantagepoint — data-layer integration architectures, timelines and cost ranges (both vendor-published)

Regulatory & legal

23. FINRA Regulatory Notice 24-09, "Firms' Use of Generative Artificial Intelligence and Large Language Models," June 2024
24. FINRA 2026 Annual Regulatory Oversight Report, GenAI section, 9 December 2025 — effective practices including prompt/output logging and model-version tracking
25. FINRA Regulatory Notice 26-14, proposed Rule 2210 modernization, 9 July 2026 (comments due 11 September 2026)
26. FINRA Rule 3110 (Supervision), Rule 2210 (Communications with the Public), Advertising Regulation FAQ item D.8 (updated 22 Dec 2025), Regulatory Notice 21-29 (vendor management)
27. SEC Rule 206(4)-7; Rule 204-2 (Advisers Act books and records); Exchange Act Rule 17a-4 and the 2022 amendments permitting the audit-trail alternative (compliance May 2023)
28. SEC Division of Examinations FY2026 Examination Priorities, 17 November 2025
29. SEC, Regulation S-P amendments, adopted May 2024, 89 FR 47786 — incident response, service-provider oversight, 72-hour breach notification; compliance Dec 2025 / June 2026
30. Skadden, "How and When SEC Recordkeeping Rules May Apply to AI-Generated Content," 13 September 2024
31. Cooley / TheFundLawyer (Stacey Song), "AI Notetakers and the Books and Records Rule," 24 April 2026
32. ABA Business Law Today (Diamond & Sheridan), "AI Prompts and Responses: Records or Not, Here We Come," 9 June 2026
33. Troutman Pepper Locke, "AI Heightens Old Compliance Risks For Investment Advisers," 23 June 2026
34. In re Otter.AI Privacy Litigation, No. 5:25-cv-06911 (N.D. Cal.), consolidated 22 October 2025; motion to dismiss argued 20 May 2026
35. Kearney v. Salomon Smith Barney, Inc., 39 Cal. 4th 95 (2006) — choice of law for multi-state call recording
36. Matthiesen, Wickert & Lehrer and RecordingLaw state-consent charts — cited to show that the two lists disagree

Vendors & market structure

37. Vendor pricing pages retrieved 2 August 2026: Jump, Zocks, RightCapital, Wealthbox, Catchlight, Salesforce Financial Services Cloud, OpenAI, Microsoft 365, Google Workspace
38. InvestmentNews, "Is $50 the new $120? Price compression comes to AI notetakers," 9 January 2026
39. XY Planning Network, "Battle of the Bots: Comparing and Ranking Nine AI Notetaker Solutions," 23 July 2026 — hands-on transcription accuracy testing
40. WealthTech Today, "AI Notetakers & Agentic OS for Financial Advisors: The 2026 Strategic Buyer's Guide," 8 May 2026
41. Feature absorption: Nitrogen "Nucleus" (19 Feb 2026), Orion "Denali AI Enterprise" (26 Feb 2026), Morningstar AI Assistant (9 Mar 2026), Robinhood Cortex for Advisors on TradePMR (3 Jun 2026)
42. Funding: Jump $80M Series B (Feb 2026), Zocks $45M Series B (Jan 2026), Conquest $80M Series B (Jun 2025), Rogo $160M Series D (Apr 2026)
43. Morgan Stanley / OpenAI case study — AI @ Morgan Stanley Assistant, Debrief, evaluation and daily regression testing methodology
44. Bank of America, "Merrill and Bank of America Private Bank Launch AI-Powered Meeting Journey," 26 March 2026