On the sixth of February 2026, CNBC reported that Goldman Sachs had spent the previous six months working with embedded engineers from Anthropic to build autonomous AI agents, powered by Anthropic's Claude model, to handle two specific categories of back-office work: the accounting of trades and transactions, and the vetting and onboarding of new clients through the Know Your Customer and anti-money laundering verification processes [1]. The agents are projected to reduce client onboarding timelines by approximately thirty per cent. They operate in the workflow layer, handling the extraction, comparison, and preliminary categorisation of structured and unstructured data, with human analysts retained to manage exceptions that fall outside the rules the agents have been trained to apply [2]. Goldman described the initiative as an efficiency project and stated that it expected efficiency gains rather than near-term job cuts, using the technology to speed processes and limit future head count growth. This was, as such announcements invariably are, an accurate description of the immediate intention and a significant understatement of the structural implication.
I do not raise this observation to criticise Goldman Sachs, which is doing precisely what a well-managed institution should do when presented with a genuine productivity opportunity, or to impute bad faith to an announcement that was probably drafted by people who believe every word of it. I raise it because the distinction between "near-term" and "structural" is one that banking has encountered before, at each of the major technological inflection points in the industry's history, and the pattern of those encounters is instructive in ways that the near-term framing consistently fails to capture.
What the ATM Did, and Did Not Do
The canonical example is the automated teller machine. ATMs were deployed in American bank branches beginning in the early 1970s, and by 1990 there were approximately 100,000 of them operating across the country, rising to approximately 352,000 by the turn of the millennium [3]. The prediction at each stage of this expansion was that the machines would displace bank tellers, whose primary function, the mechanical processing of cash deposits and withdrawals, could be performed by the machine at lower cost and with greater availability. The prediction was wrong, for a period that lasted roughly two decades. The number of bank tellers employed in the United States rose from approximately 500,000 in 1990 to approximately 527,000 in 2004, in a period during which ATM numbers quadrupled [4]. The explanation is now well established in the labour economics literature: ATMs reduced the cost of operating a bank branch, which enabled banks to open more branches, which required more tellers to handle the expanded branch network and the non-mechanical aspects of retail banking that the ATM could not perform. Automation of the mechanical task enabled expansion of the human task; the net effect on employment was close to zero for a sustained period.
The labour economics literature refers to this pattern as the "lump of labour fallacy," after the erroneous assumption that there is a fixed quantity of work to be done and that automation therefore necessarily reduces the quantity available for human workers [5]. The ATM story is the definitive banking-sector demonstration that the fallacy is indeed a fallacy: when a technology reduces the cost of performing a task, it typically increases the total volume of that task being performed, creates new complementary tasks, or both. The net employment effect depends on the elasticity of demand for the service and the degree to which the automated component is genuinely substitutable for the human component across the full range of functions involved.
Why This Time Is Different: The Nature of What Is Being Automated
The reason I am not entirely reassured by the ATM analogy, in the context of Goldman's February announcement, is that the analogy rests on a categorisation of the relevant human work that no longer applies. The ATM automated the mechanical execution of a transaction: it accepted a card, verified a PIN, disbursed cash, and printed a receipt. None of these operations required, in any meaningful sense, the application of judgment. A bank teller doing the same task was not exercising professional discretion; she was performing a sequence of steps that any competent person could be trained to perform in a few hours. The ATM replaced that sequence with a machine, and the teller was freed to perform the aspects of branch banking that genuinely required human engagement: explaining products, managing complaints, exercising discretion on unusual requests, and providing the basic social function that retail banking had always included.
The Goldman AI agents are automating something categorically different. Trade accounting, at the level of complexity that Goldman Sachs operates, is not a mechanical task. It requires the interpretation of transaction records that may be ambiguous, incomplete, or internally inconsistent; the application of rules whose interaction is not always straightforward; and the exercise of judgment in cases where the rules run out or where the data does not neatly fit the categories the rules were designed to accommodate. Client onboarding through KYC and AML verification is similarly not mechanical: it involves the assessment of counterparty risk, the interpretation of documentation that may be provided in multiple languages and formats, the recognition of suspicious patterns that do not correspond to any simple checklist, and the exercise of discretion on cases that fall into grey areas. These are not execution tasks in the ATM sense. They are reasoning tasks, applied in structured domains with high volumes of structured and unstructured inputs. They are, as the American Banker observed in its coverage of the announcement, "data-intensive reasoning problems" [2].
The distinction matters because the lump of labour fallacy argument for the ATM rested on the observation that the machine replaced only the execution, while the human retained the judgment. When the machine replaces the judgment as well as the execution, the argument loses its force. The freed human analyst who "handles exceptions" is performing a function that is, by definition, residual: it is the category of cases that the agent could not resolve, and the expectation, across every technology deployment of this kind, is that this category will shrink over time as the agent is trained on the exceptions it previously failed to handle.
The ATM freed the teller from the mechanical task while leaving her the judgment. The agentic AI handles both. What remains for the human analyst is the exception, and exceptions, by design, diminish.
The Scale That Makes Agents Necessary
The context in which Goldman's agent deployment should be understood is not primarily one of cost reduction but one of volume. Global private credit assets under management exceeded three and a half trillion dollars by the end of 2025, a figure that represents roughly a tripling over the preceding decade, driven in significant part by the constrained lending capacity of regulated banks under the capital requirements that the Basel III Endgame was, until this week, expected to tighten further [6]. JPMorgan Chase announced a fifty billion dollar commitment to private credit partnerships in the same period; Citigroup and Apollo announced a twenty-five billion dollar joint lending programme [7]. The operational consequence of these arrangements, in terms of trade accounting, documentation verification, and counterparty onboarding volume, is substantial. The human workforce capable of processing these volumes at the requisite speed and accuracy does not, at current training and recruitment rates, exist in sufficient quantity. The agent is not replacing workers who might otherwise have been hired; in many cases, it is filling a gap that no available workforce could fill.
This is a genuinely different situation from the one that has typically been invoked in discussions of automation and banking labour. When the electronic trading revolution of the 1990s and early 2000s eliminated most of the jobs on the trading floors of the major exchanges, it did so by replacing human intermediaries whose function was price discovery and order matching, tasks that electronic systems could perform more quickly, more cheaply, and with greater reliability. The displacement was rapid and concentrated, and the banking industry absorbed it primarily through expansion into new product areas rather than through retraining the displaced traders. The volume dynamics in the current case are, if anything, more acute: the combination of increasing transaction complexity, regulatory documentation requirements that have grown substantially since 2008, and the expansion of non-bank lending activity that feeds back into bank balance sheets through warehousing and co-investment arrangements has generated a documentation and verification workload that is, by any reasonable assessment, not manageable at scale with human-only workflows.
Accountability and the Limit of Delegation
The question that Goldman's announcement does not address, and that neither Goldman nor Anthropic has any obvious incentive to address prominently, is the question of accountability. When a human analyst misclassifies a trade, fails to identify a suspicious counterparty, or exercises judgment incorrectly in a KYC assessment, the error is attributable to a person with a name, a professional licence, and a legal status that makes the attribution meaningful. The regulatory framework for anti-money laundering compliance, in particular, is built on the assumption that specific individuals bear specific responsibility for specific determinations. A Money Laundering Reporting Officer exists precisely because the law requires a named person to make the ultimate judgments in the compliance chain.
The introduction of AI agents into this chain does not eliminate this requirement; it complicates it. Goldman has stated that human analysts remain responsible for managing exceptions, and the legal position is presumably that the human analyst who reviews and approves an agent's output bears the relevant professional responsibility. But the practical reality of a system in which an agent processes the routine cases and a human reviews only the exceptions is that the human's effective oversight of the agent's reasoning is limited to the subset of cases where the agent's output is most likely to be correct, namely those where the case is sufficiently straightforward to fall within normal parameters. The genuinely difficult cases, by contrast, are the exceptions, and whether a human analyst reviewing a flagged exception has sufficient visibility into the agent's reasoning process to provide meaningful oversight rather than merely formal approval is a question that neither the technology nor the regulatory framework has yet resolved [8].
What to Watch
The Goldman-Anthropic deployment is not, in itself, the transformative event that will answer the structural question. It is an early indication of a trajectory. The relevant variables to monitor over the next several years are not the efficiency metrics that Goldman will report, which will almost certainly be favourable, but the following: whether the category of "exceptions requiring human review" contracts over time, as the agents are trained on their own edge cases; whether the regulatory frameworks governing AML and KYC compliance are updated to specify what constitutes adequate human oversight of agent-assisted determinations; and whether other institutions develop comparable capabilities through their own AI partnerships, creating a competitive dynamic in which agent deployment becomes a table-stakes requirement rather than an optional efficiency initiative.
If the first of these things occurs, and on current trends it probably will, the staffing implications will be different from those that the ATM deployment produced, because the ATM's expansion of branch networks was driven by a demand elasticity for retail banking services that the broader population was willing to act on. It is less clear that a thirty per cent reduction in the time required to onboard a new institutional counterparty will generate a proportionate expansion in the number of institutional counterparties being onboarded. The volume driver for private credit and institutional banking is not primarily the speed of onboarding; it is the availability of capital and the risk appetite of investors. Faster onboarding may capture some marginal volume that slow onboarding was previously foregoing, but the effect is unlikely to be of the magnitude required to offset a substantial reduction in the headcount engaged in the underlying process.
Bagehot wrote, in his observations on the City of London, that a bank's most important asset was neither its capital nor its connections but the accumulated judgment of the men who operated it, judgment that was the product of years of close observation and could not be quickly manufactured or easily replaced. He wrote this in 1873, before the telegraph, before the ticker tape, before the electronic ledger, before every prior wave of automation that banking has absorbed and, thus far, survived with its essential functions intact. Whether the absorption continues, and whether the human judgment that Bagehot identified as the irreducible element of banking survives the present wave in a form that is genuinely human rather than merely nominally so, is the question that Goldman's February announcement places before the industry. It will not be answered quickly. It will be answered, and the answer will matter considerably more than the thirty per cent onboarding efficiency gain with which the conversation began.
- CNBC. "Goldman Sachs Taps Anthropic's Claude to Automate Accounting, Compliance Roles." CNBC. 6 February 2026. cnbc.com
- American Banker. "Goldman Equips AI Agents to Do Trade Accounting, Onboarding." American Banker. February 2026. americanbanker.com
- Humphrey, David B. "Why Do Estimates of Bank Scale Economies Differ?" Federal Reserve Bank of Richmond Economic Review. 1990. See also: ATM Industry Association historical deployment data.
- Bessen, James. "Toil and Technology." Finance and Development, IMF. March 2015. imf.org
- Autor, David H. "Work of the Past, Work of the Future." AEA Papers and Proceedings, Vol. 109. May 2019. aeaweb.org
- Alternative Credit Council; Preqin. "Private Credit: Understanding the Asset Class." AIMA & ACC Global Private Credit Report. December 2025.
- Boston Institute of Analytics. "The Rise of Private Credit: How Investment Banks Are Competing With Shadow Lenders in 2025." BIA Analysis. 2025. bostoninstituteofanalytics.org
- PYMNTS. "Goldman Sachs Lets AI Agents Do Accounting and Compliance Work." PYMNTS.com. February 2026. pymnts.com