The Intelligence Layer

Three dates mark the arrival of native AI in professional finance. On 19 May 2026, CAIS, the leading alternative investment platform for independent financial advisers, announced an integration with Anthropic's Claude built on the Model Context Protocol, embedding alternative investment intelligence directly into advisers' existing workflows.³ It was an early signal of where the industry was heading: not AI as a separate destination, but AI wired into the platforms where work already happens.

On 10 September 2026, OpenAI launched ChatGPT for Financial Services, a tailored offering for financial institutions combining built-in financial data with its frontier model's reasoning, shaped by design partnerships with Morgan Stanley and Evercore and aimed initially at the work of investment bankers and equity researchers.¹⁴

Four days later, on 14 September 2026, Anthropic launched Claude for Financial Advisors, a suite of connectors and workflow skills linking the assistant to the custodians, portfolio platforms, CRMs and planning tools advisers use daily. Launch integrations and partners named in the announcement and contemporaneous reporting include BlackRock, Charles Schwab, Vanguard, Addepar and iCapital, with the suite built in partnership with firms including Dynasty Financial.²⁵⁶

The sequence matters less than the direction. Within four months, the two leading frontier AI laboratories and the largest alternatives distribution platform in the independent adviser market all shipped products that put AI inside the daily workflow of the front office. This is not a pilot phase. The partner lists, which read as a roll call of the largest asset managers, custodians and platforms in wealth management, indicate production deployment intent.

The new products converge on the same cluster of front office tasks. Anthropic's announcement describes connectors and skills for the research, preparation and documentation work that occupies an adviser's day: meeting preparation, portfolio reviews and client communication drafting, with portfolio tracking among the automated workflows.²⁶ OpenAI's offering targets research, financial modelling and customised client materials for banking and research teams.¹ CAIS's integration lets advisers query fund data, evaluate manager performance and surface portfolio insights without leaving their primary workspace.³⁷

Compliance screening sits inside this wave as well. When client notes, meeting records and portfolio data flow through a single assistant connected to the CRM and the custodian, suitability checks and documentation standards can be applied continuously rather than sampled retrospectively. That is a genuine operational gain, and it explains why the largest platforms in wealth management signed on as launch partners rather than observers.²⁵

The front office, in short, just got faster. An adviser who spent Monday morning preparing for the week's client meetings now walks in with preparation already done. A banking analyst who spent nights assembling comparable analysis now supervises a first draft. These are real productivity gains, and they will compound as the connectors deepen.

But notice what each of these workflows consumes: data. Meeting preparation consumes CRM records, holdings and history. Portfolio reviews consume positions, transactions and performance. Compliance screening consumes documents, notes and communications. Every one of these AI outputs is only as good as the underlying records, and in private markets those records do not live in the front office. They live in the middle and back office, in fund accounting systems, administrator portals, data rooms, custody feeds and reporting packs.

Nothing in the September announcements touches the systems where private markets data is actually produced and maintained. Fund accounting still runs on ledger-centric platforms, many implemented a decade or more ago. Reconciliation between administrators, custodians and managers still runs on spreadsheets and email. Capital call and distribution notices still arrive as PDFs. Portfolio company data still moves through quarterly packs assembled by hand. Governance records, consents, valuations and audit trails still sit in disconnected systems, each holding a partial and slowly diverging copy of the truth.

We documented the cost of this fragmentation in The Fragmented Fund: data inconsistency across point solutions, intelligence decay as context is lost between systems, governance gaps where no system holds the full record of a decision, and inflated total cost of ownership as firms pay to reconcile their own data back to itself. None of those costs are addressed by an adviser-facing assistant. They are, if anything, amplified by one.

The reason is straightforward. Front office AI increases the volume and velocity of questions asked of the data. An adviser who can query a portfolio conversationally will ask ten questions where they previously asked one. Every question is answered from whatever data the assistant can reach. If the underlying records are fragmented, stale or unreconciled, the assistant does not fix that. It industrialises it, producing fluent, confident outputs on unreliable foundations, faster than any review process can catch.

This is the asymmetry at the heart of the current moment. The industry has just invested, through its platform partners, in dramatically accelerating the consumption of private markets data. It has not yet invested equivalently in the production, governance and connection of that data. The gap between the two is where the next generation of infrastructure will be built.

The natural response to the front office wave is procurement: buy the assistant, sign the connector, ship the integration. The September launches are explicitly designed to make this easy, meeting advisers and bankers inside the tools they already use.¹² For an individual adviser or a deal team, that is rational. For a platform or a fund, it compounds a structural problem.

Each new AI touchpoint is another system that reads the firm's data, reasons over it, and produces outputs that live somewhere else. The diligence memo drafted in one assistant, the portfolio review generated in another, the client note filed to the CRM, the compliance flag raised in a third tool: each is a fragment of institutional intelligence, and none of them enrich a record the firm owns. This is intelligence decay at machine speed. The fragmentation we described in The Fragmented Fund was created by human copy-paste between tools; assistants perform the same scattering automatically.

There is also a strategic exposure that platforms in particular should weigh carefully. Protocols like MCP make it trivially easy for an external assistant to query a platform's data.³ That is excellent for the adviser's experience. It also means the reasoning, the context assembly and the compounding usage intelligence happen on the model provider's side of the connection. The platform supplies the raw material; someone else's system learns from every interaction. Platforms that go down this path without their own intelligence layer are choosing to be data suppliers to other companies' compounding assets.

And the governance obligations are arriving regardless. As we set out in The Regulated Fund, the EU AI Act's high-risk regime applies from August 2026, with parallel expectations forming across the FCA, SEC, MAS, SFC and ASIC. Regulators will ask how AI-assisted outputs were produced, on what data, under whose oversight. An estate of disconnected assistants, each with its own access pattern and no shared record, is close to the worst possible answer to that question.

The resolution of the asymmetry is a distinct layer of infrastructure that the private markets stack has never had: an intelligence layer that sits between the systems of record and the systems of reasoning. Its job is threefold. First, connection: it assembles a single, structured, continuously reconciled representation of the firm's funds, deals, positions, documents and decisions from the ledgers, administrators, data rooms and platforms where those facts are produced. Second, governance: it carries permissions, provenance, residency and oversight as native properties of the data, so that any reasoning over it inherits the firm's controls. Third, service: it exposes that governed representation to whatever consumes it, whether that is an internal workflow, a client-facing experience, or an external AI assistant arriving through a protocol like MCP.

Each property matters because of what the September launches made clear. Connection matters because the assistants are only as good as the data they can reach. Governance matters because regulators and institutional clients will demand evidence of how outputs were produced, and because no platform should hand its compounding intelligence to an external party by default. Service matters because the reasoning systems will keep changing: the model that advisers use in 2027 may not be the one they use in 2026, and a platform whose value lives in its own governed layer survives that turnover intact.

This is not an argument against the front office products. They are genuinely useful, and they will improve. It is an argument about sequence and ownership. Firms that adopt assistants first and infrastructure later will spend the intervening years generating intelligence they do not own, on data they cannot fully trust, under obligations they cannot yet evidence. Firms that build or adopt the intelligence layer first get the same assistants, running on better data, inside their own perimeter, compounding an asset they keep.

Distribution and access platforms face the sharpest version of the choice. The CAIS integration shows the playbook: expose fund data through MCP and let a frontier assistant do the reasoning.³ It is fast, it delights advisers, and it quietly cedes the layer where usage intelligence compounds. The alternative is to operate an intelligence layer of their own, so that when an adviser's assistant calls, it calls into governed infrastructure the platform owns, and every interaction enriches the platform's asset rather than the model provider's. Platforms that move early here will set the terms on which the assistants plug in.

Fund administrators and custodians sit on the raw material the entire wave depends on. Their ledgers and records are the ultimate source of truth that every front office output traces back to. That position is an opportunity if their data is connectable and machine-readable, and a liability if it is not, because clients will increasingly route around sources their assistants cannot reach. The administrator that offers a governed intelligence layer on top of its books becomes more embedded, not less, as AI spreads.

Funds and asset managers should treat this as an infrastructure decision, not a software purchase. The question is not which assistant to buy; the September announcements guarantee there will be several good ones. The question is whether the firm's decisions, diligence and reporting live in a layer the firm governs, from which any assistant can draw and to which every output returns. That is the architectural position that survives model turnover, satisfies the incoming regulatory perimeter, and compounds.

The front office wave of 2026 will be remembered as the moment AI became ordinary in professional finance. The firms that define the next decade will be the ones that understood what the wave left untouched: the layer beneath, where data becomes intelligence, and where intelligence becomes an asset. Every private markets platform will need one. The only open question is who owns it.