What happens when the quiet, labor‑intensive parts of Wall Street meet the newest wave of agentic AI? Goldman Sachs is about to find out.
For the past six months the bank has embedded engineers from Anthropic inside its technology teams to co‑develop autonomous agents powered by the Claude model. Those agents are being trained to handle dense, rules‑driven tasks — trade and transaction accounting, reconciliation, client due diligence and onboarding — the kind of back‑office work that has resisted automation for decades.
Marco Argenti, Goldman’s chief information officer, frames the project simply: these are "digital co‑workers" for functions that are scaled, complex and process intensive. That description captures both the promise and the awkward truth: the bank is not trying to replace expertise overnight but to collapse the time those workflows take, freeing humans for higher‑value decisions.
Real work, not just toy demos
Early experiments at Goldman began with an autonomous coding assistant called Devin. Engineers quickly realized Claude’s strengths went beyond mere code generation. Argenti and his team saw the model reason through multi‑step problems and apply rule sets to messy datasets — the very capabilities needed for accounting and compliance.
Anthropic’s enterprise push — including products like Claude Cowork that execute computer tasks for white‑collar workers — has accelerated interest from institutions that prize control over data and governance. Goldman says it is still in the early stages, and while agents are expected to launch soon, no hard date has been given.
Efficiency without an immediate pink slip
Goldman’s leadership is explicit about intentions. CEO David Solomon has spoken of a multiyear reorganization centered on generative AI that seeks to constrain headcount growth. But Argenti emphasizes that the initial goal is capacity — speed, accuracy and better client experiences — rather than immediate job cuts. In practice, however, increased automation could limit future hiring and make some third‑party providers redundant.
That subtle distinction matters to employees and investors alike. When Anthropic unveiled coworking tools recently, the market reacted: enterprise software stocks tumbled on fears that agentic platforms could displace long‑standing revenue streams.
Part of a larger shift in finance
Goldman’s move is not an island. Large banks are experimenting with hundreds of AI use cases, from customer service chatbots to internal idea generation. Some firms choose to build internal agent layers — a trend that gives them tighter control over sensitive finance workflows — while others stitch together external models.
CFOs and finance teams are approaching these tools cautiously, sequencing deployments where risk and control can be managed. In structured, rules‑based areas such as reconciliation and regulatory checks, AI is already showing real utility.
Governance, audits and ethics
Automation of compliance work raises its own ironies: you train a machine to enforce rules, but you then need robust processes to audit the machine. That has renewed attention on ethical guardrails, benchmarking and bias testing, especially when decisions affect clients or regulatory reporting. Industry efforts to set consent‑forward and audit‑friendly standards are gaining traction, and organizations are looking to formal benchmarks to ensure models behave as intended — a discussion that echoes recent initiatives around ethical AI benchmarking.
What else could agents do?
Argenti hinted at broader possibilities: from drafting investment‑banking pitchbooks to monitoring employee activity. Those use cases cut across convenience and contention. Automating pitchbooks could turbocharge deal teams; automating surveillance opens thorny legal and morale questions.
Goldman also sees a path where agents reduce reliance on external vendors. That would compress costs but increase internal responsibilities for security, model tuning and regulatory compliance.
A cautious, experimental cadence
If there is a theme, it is incrementalism. Goldman is piloting, embedding engineers, tuning models and defining review gates. The bank is betting that well‑engineered agents can turn traditionally slow, document‑heavy chores into near real‑time operations without throwing out the human judgment those functions sometimes require.
For readers tracking the broader technology trend, agentic systems are already appearing beyond finance: from consumer booking assistants to enterprise workflows. The rise of agentic booking tools and plug‑ins shows how quickly these capabilities are spreading across domains, reshaping expectations about what software should do for users and staff alike (see recent developments in agentic booking technology).
Goldman’s experiment will be an important data point. If the bank can safely compress onboarding and reconciliation times without losing control, other institutions will follow. If it missteps — either operationally or legally — the episode will be a cautionary tale about the real costs of automating judgment.
Either way, the office that used to hum with spreadsheets and phone calls is about to get a new kind of colleague — one that demands new rules for oversight, auditing and, yes, accountability. The questions now are not only whether the technology works, but how firms will make it trustworthy. For a glimpse at the kind of governance conversations this unleashes, look to efforts around ethical benchmarking in AI that are starting to shape enterprise practice (/news/sony-fhibe-ethical-ai-benchmark).
The machines Goldman is building won’t be dramatic replacements. They'll be quiet, relentless processors of paperwork, and in that low, steady work lies the most disruptive change of all: the slow erasure of time as a cost. Whether that turns into better service, leaner payrolls or both will depend on choices the bank makes in the months and years ahead.