AI's Wealth Management Land Grab Begins

Two frontier AI laboratories shipped finance products within five days of each other this month. Neither led with a benchmark.

On 10 September, OpenAI launched ChatGPT for Financial Services, built with Morgan Stanley and Evercore as design partners and pointed at bankers and equity researchers. On 14 September, Anthropic went after the other end of the advice chain, releasing Claude for Financial Advisors and plugging it into Charles Schwab's custody platform, and with it, the 16,000-plus registered investment advisers Schwab serves.

Both launches are, in substance, integration announcements. After three years of argument about whether language models can reason about markets, the commercial question has quietly become far more prosaic: who controls the pipes into the systems where client money actually sits?

What happened#

Anthropic's release bundles two things. First, connectors giving Claude authenticated access to the platforms advisers already run on: Schwab Advisor Center for custodial data, BlackRock's Advisor Center for model portfolios, Vanguard's advisor solutions, Envestnet's Tamarac and MoneyGuide, Orion and its Redtail CRM, SS&C Black Diamond, Addepar, iCapital for alternatives, Wealthbox, Wealth.com for estate and tax records, and the meeting-capture tool Zocks. These sit alongside existing links to Salesforce, Microsoft 365, FactSet, S&P Global, Morningstar and Box.

Second, a library of eight task-specific "skills" mapped to an adviser's day, meeting prep, post-meeting notes, rebalance review, estate and tax briefs, alternatives summaries, prospect intake, adviser onboarding, and a compliance workflow that screens client-facing language against the SEC Marketing Rule.

Schwab's role is the commercially significant part. Schwab Advisor Services says it is the only RIA custodian currently integrated with Claude, exposing balances, positions, transactions, cost basis and money-movement status. The firm reported $13.41tn in total client assets at end-August, roughly $5.9tn of it in Advisor Services.

OpenAI's product attacks the institutional end, shipping with embedded data from Daloopa, PitchBook, LSEG News, Crunchbase and Fiscal.ai, pass-throughs to S&P Capital IQ, MSCI, Moody's and Factiva, and information barriers enforced through separate workspaces, the compliance scaffolding a sell-side firm demands before a model nears a deal team.

Background: why the pipes matter more than the model#

A registered investment adviser is a fiduciary, and in US practice that duty is discharged through documentation: what was recommended, on what basis, reviewed by whom. The Investment Adviser Association counts 16,544 SEC-registered advisers managing $176.8tn for 73.7 million clients as of 2025, a vast, fragmented industry whose real constraint is time and evidence rather than analytical horsepower.

That shapes what a useful product looks like. A model reasoning brilliantly about a portfolio it cannot see is worthless. A merely competent model with permissioned, logged access to the custodian, the CRM, the performance system and the planning software changes the economics of a practice, because the bottleneck was never the analysis. It was reconciling five systems that do not talk to each other.

Market implications#

The equity read-through is thin, and honestly so. Schwab closed at $107.79 on 15 September, up 0.45%, noise against a $186bn market capitalisation, on a day it also posted record August net new assets of $64.8bn. Investors are pricing a retention story with no near-term revenue line. That looks about right.

The structural consequences deserve more attention than the tape does.

Custody has been a commoditised spread-and-scale business for a decade; an exclusive AI integration is a real, if temporary, differentiator in the fight for adviser flows, and it puts Fidelity, Pershing, Altruist and LPL on a clock. Vendor concentration is the flip side. Route a meaningful share of an industry's workflow through two model providers and you have created a third-party dependency of supervisory interest, on 11 September the Federal Reserve, FDIC, OCC and NCUA jointly proposed replacing existing third-party risk management guidance, and IOSCO's 2025 consultation report names over-reliance on third-party vendors as a distinct AI risk category.

For wealthtech vendors the question is sharper: is the conversational layer a feature you own, or a commodity that turns your platform into a data source? The connector list reads like a group of firms that decided being reachable beats being bypassed.

None of this moved markets this week; the FOMC's 15–16 September meeting dominated, with a quarter-point rise widely priced and Treasury yields at multi-year highs. These launches play out over quarters, not sessions.

Technical deep dive#

The engineering that matters is the connector layer, and it rests on the Model Context Protocol. MCP is an open standard, built originally by Anthropic and then donated in December 2025 to the Agentic AI Foundation, a Linux Foundation body whose platinum members include AWS, Anthropic, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. A provider such as Schwab writes one server exposing its data and actions; any compliant AI client consumes it under the end user's own credentials.

That cuts against the exclusivity narrative. The protocol is neutral and what Schwab built is portable in principle, so the barrier facing a rival lab is contractual rather than technical. Read "exclusive" as "first".

The harder engineering problem is the division of labour between the model and everything else. Rafael Loureiro of Wealth.com put it well in trade coverage of the launch: complex planning needs three layers, extraction, computation and judgement, and language models are strong at the first and third, while tax arithmetic is deterministic and must not be probabilistic. A well-built system routes calculation to a calculation engine and uses the model to orchestrate and explain. A badly built one lets the model do the sums. The two are indistinguishable from the outside until the first error reaches a client's return.

Note what nobody claimed. Anthropic states that investment recommendations, client communications and compliance determinations remain subject to human review and approval. OpenAI's headline benchmark is modest by design: GPT-6 Astra scores 69.9% on OfficeQA Pro against 60.2% for its predecessor. A meaningful gain, and a long way from unsupervised reliability.

Critical analysis#

Begin with the premise. Anthropic's post cites Kitces research showing a typical practice spends only a sixth of its time in client meetings. Michael Kitces rejects the conclusion usually drawn from his own data. In a July webinar reported by Financial Advisor magazine he called the goal of pushing advisers to 80% of their time in meetings a "dystopia", noting that reasonably successful advisers already sit at 20–25%. The time-liberation thesis borrows empirical authority from a researcher who disputes it.

The evidence on AI and knowledge work is real, bounded and strikingly uneven. The strongest peer-reviewed result, Brynjolfsson, Li and Raymond in the Quarterly Journal of Economics, covering 5,179 support agents, found a 14% average lift in issues resolved per hour, 34% for novices, and almost nothing for experienced staff. The most important caveat comes from Dell'Acqua and colleagues at Harvard Business School: across 758 consultants, GPT-4 delivered 12.2% more tasks and 25.1% faster work inside its competence, but users were 19 percentage points less likely to be correct on a task just outside it. The frontier is jagged, and practitioners cannot see the edge.

One cautionary tale belongs here. The most-quoted academic claim that GPT-4 beat analysts at directional earnings prediction, Kim, Muhn and Nikolaev's 2024 preprint, was withdrawn by its authors in February 2025 after a co-author found inconsistencies in the data and analyses. It was never peer reviewed. It still circulates in vendor decks.

Practitioners are split, sensibly. Alois Pirker of Pirker Partners warned that Claude will meet the same specificity wall Salesforce did, since one large firm contains many kinds of adviser wanting very different things. Will Trout of Datos Insights doubted the workflow promises survive contact with real integrations; Orion's Reed Colley was warmer on governed access to proprietary data. Neither vendor disclosed pricing beyond a launch usage credit, and adoption economics, not capability, will decide whether small practices ever see any of this.

Historical context#

This is a change in distribution rather than in what models can do. The closest analogue is the robo-advice wave of 2014–2016, which promised to disintermediate advisers and ended up a feature inside incumbent platforms. Anthropic launched a general Claude for Financial Services in July 2025; this release narrows it to a vertical and pushes it through a partner.

What is new is agentic depth. Earlier adviser software read data. These systems draft, stage actions and update systems of record, with a human gate before anything reaches a client, a materially different supervisory problem. The SEC's 2026 examination priorities commit examiners to reviewing "for accuracy registrant representations regarding their AI capabilities" and to testing whether firms supervise AI use adequately. Buy an AI compliance feature and you have also bought a new examination surface.

Key takeaways#

1. The contest has moved to plumbing. Both launches compete on authenticated access to adviser and banker systems rather than on reasoning benchmarks.
2. Distribution is the asset. Being the first integrated RIA custodian matters more to adoption than any capability gap between frontier models.
3. The moat is contractual. MCP is an open standard governed by a body containing both labs, so exclusivity here means first-mover advantage, not a defensible barrier.
4. Productivity gains are real but jagged. Peer-reviewed effects cluster at 12–14% and concentrate among less experienced workers; accuracy falls sharply outside the model's competence.
5. Governance paces adoption. With AI supervision an explicit SEC examination priority and bank regulators rewriting third-party risk guidance, compliance is the binding constraint.

Frequently asked questions#

Is AI now giving investment advice? No. Both vendors keep investment recommendations, client communications and compliance determinations subject to human review and approval. What is automated is the preparation, summarisation and drafting around the advice.

Why does the Model Context Protocol matter? It standardises how an AI system reaches external data and tools. A custodian writes one server; any compliant client uses it under the end user's permissions. That makes integrations portable and limits how long any exclusivity can last.

Is Schwab's exclusivity meaningful? Commercially, for now, it differentiates the platform in competing for adviser flows. Technically, no. Nothing in the protocol stops Schwab or a rival custodian building equivalent connectors for other AI providers.

What is the biggest implementation risk? Routing deterministic computation, tax calculations, performance attribution, required minimum distributions, through a probabilistic model instead of a calculation engine. Those errors look plausible and are hard to catch on review.

How should firms evaluate these tools? Test on tasks with verifiable ground truth, measure error rates rather than hours saved, and require every output to trace back to a source record. Capability is uneven in ways demonstrations rarely reveal.

Does this affect markets in the near term? Not materially, Schwab's shares moved less than half a per cent. Any revenue effect will surface over several quarters, in adviser retention and platform economics.

References#

• Anthropic, "Claude for Financial Advisors", 14 September 2026.
• Anthropic, "Claude for Financial Services", 15 July 2025.
• Charles Schwab, "Charles Schwab and Anthropic to Bring Claude to Independent Registered Investment Advisors", 14 September 2026.
• Charles Schwab, "Schwab Reports Monthly Activity Highlights", 15 September 2026 (August 2026 data).
• OpenAI, "Introducing ChatGPT for Financial Services", 10 September 2026.
• Linux Foundation, "Linux Foundation Announces the Formation of the Agentic AI Foundation", 9 December 2025.
• Model Context Protocol, official documentation.
• US Securities and Exchange Commission, Division of Examinations, 2026 Examination Priorities, 17 November 2025.
• Federal Reserve Board, FDIC, OCC and NCUA, joint proposed third-party risk management guidance, 11 September 2026.
• IOSCO, Artificial Intelligence in Capital Markets: Use Cases, Risks, and Challenges, Consultation Report CR/01/2025, March 2025.
• Federal Reserve Board, 2026 FOMC calendar.
• Brynjolfsson, E., Li, D. and Raymond, L., "Generative AI at Work", Quarterly Journal of Economics, 140(2), 2025, pp. 889–942. NBER Working Paper 31161 version.
• Dell'Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K., "Navigating the Jagged Technological Frontier", Harvard Business School Working Paper 24-013, September 2023. Working paper, not peer reviewed.
• Kim, A., Muhn, M. and Nikolaev, V., "Financial Statement Analysis with Large Language Models", arXiv:2407.17866, July 2024. Preprint, withdrawn by the authors in February 2025 and never peer reviewed. Cited only as an example of unverified evidence still in circulation.
• Investment Adviser Association, 2026 Investment Adviser Industry Snapshot, 3 June 2026 (2025 data).
• WealthManagement.com, "Claude for Financial Advisors Sparks Debate Among Analysts, Tech Providers", 15 September 2026.
• Financial Advisor magazine, "Kitces: Spending 80% Of Your Time With Clients Is A 'Dystopia'", 22 July 2026.
• Market data for The Charles Schwab Corporation (NYSE: SCHW) via stockanalysis.com, close of 15 September 2026.

This article is for information only. It is not investment advice and does not recommend buying or selling any security. Market levels are as at the close of 15 September 2026 and will have changed. Statements about future adoption, revenue or regulatory outcomes are estimates, and are distinguished in the text from verified facts.