The quiet redistribution inside Anthropic's finance push

On 5 May, Anthropic announced ten new Claude agents aimed squarely at the back, middle, and increasingly the front office of financial services: pitchbook construction, month-end close, credit memo drafting, KYC refresh, earnings-call summarisation, fund accounting reconciliation, and a handful of others in the same register. The same week, the company unveiled a Forward Deployed Engineer joint venture with Blackstone, Goldman Sachs, and Hellman & Friedman, a vehicle designed to put Anthropic's people physically inside the firms paying for those agents. The framing, repeated across the trade press, was that this is part of the company's pre-IPO enterprise revenue push.

That framing is true. It is also the least interesting thing about the announcement.

What's actually being sold here. The ten workflows Anthropic named are not a random sampling of finance work. They are, almost without exception, the specific tasks that junior bankers, associates, fund accountants, KYC analysts, and credit risk staff do for the first three to seven years of their careers. Pitchbook construction is what a first-year analyst at Goldman or Lazard does at 2am. Credit memo drafting is what a credit analyst at a mid-market lender does for years before they're trusted to underwrite. Month-end close is what an army of fund accountants at every administrator and asset manager grinds through, every month, forever. KYC refresh is what entire offshore teams in Bangalore, Manila, and Belfast were stood up to handle over the last fifteen years.

Anthropic did not pick these workflows because they were technically the easiest. It picked them because they are where the headcount is, which is where the cost is, which is where the willingness-to-pay is. The pitch to a CFO is not subtle: here are the ten line items on your org chart that you can shrink.

The pitch to a CFO is not subtle: here are the ten line items on your org chart that you can shrink.

Who actually loses. The deal sheet is silent on this, as deal sheets always are. But the people whose work these agents do are not abstract. In US investment banking alone, the analyst and associate cohort across bulge bracket, elite boutique, and middle-market firms numbers in the tens of thousands; the broader junior population across buy-side, accounting, and capital markets infrastructure is several multiples of that. Outside the US, the picture is starker because of where this work was offshored. Genpact, WNS, TCS, Infosys BPM, EXL, and the captives that JPMorgan, Goldman, Morgan Stanley, and the European banks built in India and the Philippines employ hundreds of thousands of people doing precisely the work Anthropic just productised. The Nasscom estimate of Indian BFSI BPM employment alone runs above 1.2 million when you include captives.¹

I want to be careful here. The agents do not, today, replace these workers wholesale. What they do, and this is what every Forward Deployed Engineer engagement of the last eighteen months has actually delivered, is compress the headcount needed per unit of output. A team of forty fund accountants becomes a team of twelve overseeing agent runs. A KYC refresh queue that needed two hundred analysts now needs forty reviewers. The work doesn't disappear; the people doing it do.

This is the pattern that the labour economist David Autor and his collaborators have been documenting for a decade in other sectors: the question is rarely "will the job exist" but "how many of them, at what wage, with what bargaining power."² The answer in finance back-office work is going to be: fewer, lower, less.

Why the FDE joint venture matters more than the agents. The agents themselves are a product release. The Forward Deployed Engineer joint venture with Blackstone, Goldman, and H&F is the actual mechanism. Anthropic learned, watching Palantir, that selling enterprise software to financial firms is not a software sale, it is a consulting engagement dressed as a software sale. You cannot ship a credit memo agent and expect a regional bank to deploy it. You ship the agent and a team of engineers who sit in the bank's offices for six months wiring it into the bank's systems, retraining the credit team to supervise it, and writing the memos for the audit trail that will satisfy the OCC.

The joint venture is how Anthropic scales that motion without putting two thousand engineers on its own balance sheet pre-IPO. Blackstone gets first access to deployment inside its 250+ portfolio companies. Goldman gets the same inside its own four divisions and, more importantly, inside the firms it advises. Hellman & Friedman gets to deploy across the financial services platforms it owns or is about to own.

The structure is elegant from the buyer side. When a Blackstone portfolio company restructures its finance function around Claude agents and lays off three hundred people, the cap-table benefit accrues to Blackstone's LPs. The model revenue accrues to Anthropic. The layoffs accrue to the portfolio company, where they sit on a P&L that nobody outside the deal team will ever read.

The pre-IPO logic, properly named. Anthropic is not pursuing financial services because finance is where AI does the most good. It is pursuing financial services because it is the highest-margin, highest-willingness-to-pay vertical in the economy, and because the IPO comp narrative requires booked enterprise revenue at scale. The trade press has reported the company targeting an enterprise revenue run-rate that justifies a valuation north of $300bn at listing.³ You do not get there on developer subscriptions. You get there by displacing labour costs in finance, law, and consulting, three industries whose unit economics are built on hourly billing of expensive humans.

I find it hard to believe that the people writing the S-1 are not aware of this. The whole point of the financial-services push, from a capital-formation perspective, is that the dollars Anthropic is going to capture are dollars currently flowing as wages. That is not a side effect. It is the investment thesis.

What the regulators are not yet doing. The OCC, the FCA, the PRA, ESMA, and the SEC have all issued, over the last two years, model risk management guidance that nominally applies to agentic AI in regulated workflows.⁴ The guidance is real, and the compliance teams at every bulge bracket firm are taking it seriously. But the guidance is almost entirely about model behaviour, hallucination, bias, explainability, audit trail. It is not about labour.

There is a reason for this: financial regulators do not have a labour mandate. The OCC supervises safety and soundness; the FCA supervises market conduct and consumer protection; the SEC supervises disclosure and investor protection. None of them are the body that asks "what happens to the credit analysts." That body, in the US, is the Department of Labor, which does not supervise banks. In the UK, it is split across BEIS-successor functions and ACAS, neither of which has any visibility into a Goldman deployment of Claude.

This is the gap that has always existed in financial-services labour, and it is the gap the FDE joint venture is built to exploit. There is no regulator whose job is to ask whether the deployment of a credit memo agent across a $50bn loan book should require a workforce transition plan. So nobody asks.

The offshore workforce question, specifically. I want to dwell on this because the conversation about AI and finance jobs has been almost entirely conducted as if the workers in question are graduates of Wharton and LSE. They are not, mostly. The largest concentration of people doing the work that Anthropic's ten agents now do sits in Bangalore, Hyderabad, Pune, Manila, Cebu, Belfast, Glasgow, Wrocław, and Budapest. These are workers who took these jobs because the jobs were stable, paid above local median wage, and represented entry into the global financial services labour market.

The offshore BPM industry was itself a previous wave of this same redistribution. Work was moved from London and New York to lower-cost geographies in the 1990s and 2000s, with the same arguments about efficiency and the same silence about who paid. That redistribution at least transferred wages from one set of workers to another. The current wave does not. The wages displaced by Claude agents do not become wages somewhere else; they become Anthropic revenue and customer margin.

The wages displaced by Claude agents do not become wages somewhere else; they become Anthropic revenue and customer margin.

This is the distributional fact that I think is being systematically underplayed. The previous offshoring wave was zero-sum across geographies; the current wave is positive-sum for capital and negative-sum for labour. It is not the same story.

The counter-arguments, engaged. There are serious people who will read what I've written so far and push back along three lines. I want to take each one seriously.

The first is that this work was always going to be automated, and Anthropic is simply executing what spreadsheets, ERP systems, and earlier RPA waves already started. There is something to this. The credit memo workflow has been partially automated for fifteen years; pitchbooks have template engines; KYC has had rules-based automation since the early 2010s. But the prior automation waves left substantial human judgement work in the loop, and the headcount was preserved because the tools augmented rather than replaced. The Claude agents, deployed with FDE wiring, are explicitly designed to close that loop. The difference is not marginal; it is qualitative.

The second is that the workers displaced will move into higher-value supervisory roles, and the net effect on the workforce will be benign. This is the standard CFO talking point and it is the one I find least convincing. The supervisory roles exist, but they exist in a ratio of perhaps one to ten or one to twenty against the displaced positions. The arithmetic does not work. If a fund accounting team of forty becomes a supervisory team of twelve, twenty-eight people do not become supervisors. They become unemployed, or they take a worse job.

The third is that finance work is not particularly meaningful and its compression should be welcomed on its own terms, that we should not mourn the disappearance of pitchbook drudgery. I have some sympathy for this view as an aesthetic matter and none for it as a labour matter. The people doing the work are not doing it because it is meaningful; they are doing it because it pays their rent. Whether the work is intrinsically valuable is a question for them, not for us, and the answer to "your job was not meaningful anyway" is not one any worker has ever wanted to hear.

What to watch. Three things, over the next eighteen months.

First, whether the FDE joint venture's deployments are accompanied by any disclosed workforce transition commitments. Blackstone, in particular, has said publicly it intends to deploy Claude agents across its portfolio. The portfolio companies employ, in aggregate, more than 700,000 people. The question of how many of those jobs are in scope is one that Blackstone's LPs, pension funds, sovereign wealth funds, endowments, are entitled to ask. They mostly will not.

Second, whether the offshore BPM contracts get renegotiated or simply not renewed. Genpact and WNS both have multi-year contracts with major banks that expire in 2027 and 2028. The shape of those renewals will tell us, more clearly than any Anthropic announcement, what the actual displacement curve looks like.

Third, whether any financial regulator decides that workforce displacement at this scale is, in fact, a safety and soundness question. There is a credible argument that a bank that compresses its credit analysis function from 200 humans to 30 humans plus agents has materially changed its risk profile, and that the OCC should care. So far, nobody has made that argument formally.

What's at stake. Anthropic's 5 May announcement is being read in the trade press as a product release and a revenue milestone. It is both. It is also the clearest signal yet of how the next phase of AI deployment is going to be financed, by capturing wages currently paid to several hundred thousand finance workers, routed through a joint venture structured to insulate everyone except the workers themselves from the political cost of doing it.

I do not think this is inevitable. I think it is a set of choices being made by a small number of well-capitalised actors, and I think it is worth naming them as choices while they are still being made.

Footnotes

¶