Skip to content
AI Automation

'Digital Colleagues': Goldman's Claude Agents Run $2.5 Trillion Of Operations, Lloyds Wants A Super-Agent, And The FCA Is Worried About A Faster Bank Run

Seven months after Goldman Sachs revealed it had embedded Anthropic engineers inside its technology teams to co-develop autonomous Claude agents for reconciliation, trade accounting and client onboarding - covering operations for $2.5 trillion in assets under supervision, with onboarding 30% faster - the picture of what agentic AI actually does inside a bank has come into focus. Accenture finds 57% of banking executives expect agents fully embedded in risk, compliance, audit and fraud within three years. Lloyds is building agent-creation tools for its teams and a 'colleague super-agent' for 21 million app customers. And the FCA has named the risk nobody wanted to say aloud: interacting agents moving deposits at machine speed could dramatically accelerate the probability and pace of a bank run. This is where bank agents are in September 2026 - and the model that is winning.

AlchmAI Editorial13 min read

$2.5tn

Assets under supervision whose operations Goldman's Claude-built agents manage, co-developed over six months with embedded Anthropic engineers

30%

Faster client onboarding in Goldman's tests, alongside developer productivity gains of more than 20%

57%

Of banking executives expecting AI agents fully embedded in risk, compliance, audit, fraud and transaction monitoring within three years (Accenture)

21m

Lloyds mobile app customers who are the user base for its agent-building tools and planned 'colleague super-agent'

The phrase Goldman Sachs chose was 'digital colleagues', and it has turned out to be the most accurate description of what agentic AI is actually doing inside banks in 2026. In February, the bank's chief information officer, Marco Argenti, told CNBC that Goldman had spent six months embedding Anthropic engineers within its technology teams to co-develop agents built on Claude that perform complex, rule-based work - transaction reconciliation, trade accounting, client vetting and onboarding - across operations for $2.5 trillion in assets under supervision. The tests showed onboarding 30% faster and developer productivity up more than 20%. Argenti said it was premature to expect job losses, though Goldman might cut the third-party providers it uses today as the technology matures.

Seven months on, that deployment reads less like a headline and more like a template, because the rest of the industry has converged on the same shape. Accenture's Top Banking Trends for 2026, published in January, found 57% of banking executives expecting AI agents to be fully embedded in risk, compliance, audit, fraud detection and transaction monitoring within three years, 56% expecting broad adoption in credit assessment, loan processing and KYC, and nearly half of banks and insurers already creating roles to supervise agents; McKinsey puts the net cost reduction from AI at up to 20%. The functions being automated are exactly the ones Goldman started with - and exactly the ones that have resisted automation for decades because they require processing large volumes of data against strict regulatory rules.

What Lloyds Is Building, And Why It Matters More Than Goldman

Goldman's deployment is the one that made news, but for most of the industry the more instructive case is Lloyds Banking Group, because Lloyds is a retail and commercial bank with 21 million mobile app customers, and its problem is the problem most banks have. In January its group head of AI, Rohit Dhawan, set out an agentic strategy across five areas - customer interactions, back-office operations, frontline colleague support, workplace assistance and software development - and described two things worth noticing. The first is agent-building tools that let teams create and share their own agents, which is the point at which agentic AI stops being a central technology project and becomes a capability the business owns. The second is a 'colleague super-agent' intended to automate routine tasks within compliance boundaries, with Lloyds' Responsible AI team ensuring agents operate inside guardrails with real-time monitoring and escalation protocols.

That combination - distributed creation, central governance - is the operating model the industry is settling on, and it has an important consequence for anyone building these systems. If hundreds of teams can create agents, then the control layer cannot live in each agent; it has to live in the platform, in the tools agents are allowed to call, the data they are allowed to reach, and the monitoring that watches every one of them. That is a very different engineering problem from building one clever agent, and it is the problem most banks are now actually working on.

The Risk The FCA Named

British banks have been unusually forward in retail trials - NatWest, Lloyds and Starling have all tested agentic systems with customers, on complaints handling, automated savings, budgeting and predictive spending - and the Financial Conduct Authority has responded by being unusually specific about what worries it. Its chief data officer, Jessica Rusu, put the general concern plainly: everyone recognises that agentic AI introduces new risks, primarily because of the ability for something to be done at pace. But the regulator's most striking observation was systemic rather than operational. Multiple interacting agents, each acting rationally for its own customer, could simultaneously shift deposits between accounts in response to the same signal - dramatically accelerating the probability and pace of a bank run.

“A bank run used to require people to queue. An agent that moves a customer's savings to the best rate the moment it changes does not queue, and neither do the ten million agents just like it.”

This deserves to be taken seriously by every firm building customer-facing agents, and it changes the design brief. An individually sensible agent - one that moves money to the best available rate - becomes collectively dangerous when there are enough of them acting on the same information at the same speed. The FCA also flagged the more mundane risks: agents hallucinate plausible but false responses, and executives may not deeply understand how the systems decide. Gartner's forecast frames the commercial version of the same caution - 40% of financial services firms using agents by the end of 2026, but more than 40% of projects abandoned by 2027 as costs escalate.

What Separates The Deployments That Stick

We build AI and workflow automation for financial firms, and with Goldman, Lloyds and the Accenture data as the backdrop, the pattern in the projects that survive to production is consistent enough to write down.

  1. 01They start in the back office, on the reconciliation-shaped work. Trade accounting, breaks, onboarding files, document review - high volume, rules-heavy, and verifiable in seconds by the person who used to do it. Goldman started there deliberately, and it is where the measurable return is.
  2. 02They keep the decision where the accountability is. The agent proposes, escalates or drafts; a named person with the authority to be wrong approves. Lloyds' 'within compliance boundaries' and Goldman's 'digital colleagues' are both ways of saying the same thing.
  3. 03They put the controls in the platform, not the agent. Once teams can build their own agents - which is where Lloyds is heading and every large bank will follow - the permission boundary has to be enforced by what tools an agent can call and what data it can reach, not by instructions inside each one.
  4. 04They instrument for the systemic question, not just the operational one. After the FCA's warning, a customer-facing agent needs rate limits, velocity checks and a view of aggregate behaviour across all agents, because the risk is not what one agent does but what all of them do at once.
  5. 05They embed the engineers. The single most under-reported detail of the Goldman story is that Anthropic's engineers sat inside Goldman's teams for six months. The agents that work are co-developed by people who understand the bank's processes and people who understand the models, in the same room. The ones bought as a platform and configured by a vendor are the 40% Gartner expects to be abandoned.

The Bottom Line

Agentic AI in banking has stopped being a question of whether and become a question of how, and by September 2026 the how is unusually clear. Goldman's Claude-built agents run reconciliation, accounting and onboarding across $2.5 trillion of supervised assets with onboarding 30% faster; Accenture finds 57% of banking executives expecting agents fully embedded in risk, compliance and fraud within three years; Lloyds is giving its teams the tools to build agents and governing them centrally, on the principle that AI handles volume and humans handle complexity. The FCA, watching British banks trial agents with real customers, has named the risk that matters most - millions of rational agents moving deposits at the same instant - and the design answer is aggregate monitoring, velocity limits and controls enforced by the platform rather than the prompt. The deployments that stick start in the back office, keep judgement with accountable people, put permissions in the platform, and are built by embedded engineers rather than configured from a vendor's menu. That is the workflow and AI automation work we do for financial firms in London, and the Goldman and Lloyds stories are the clearest evidence yet that it is the model the industry has chosen.

References & Further Reading

AI AutomationAI Automation LondonWorkflow Automation AgencyAgentic AIClient-Facing & Internal ToolsCompliance & Regulatory SystemsAI Agency UK
Share Email
AI

AlchmAI Editorial

Research and analysis, London

The AlchmAI team writes about the markets, technology and regulation we work with every day. We build trading platforms, real-time charts and AI analysis tools for brokers, prop firms and fintech teams from our office in Mayfair, London. Every article lists its sources. Nothing we publish is investment advice.

This article is general information and commentary. It is not investment advice or a recommendation to buy or sell any investment. Important information

Keep reading

All articles