Citi’s AI Agents: A Glimpse into Banking’s Future

Banks have long talked about automating tasks, streamlining workflows, and improving productivity. But Citigroup’s latest move takes that talk several steps further. In September 2025, Citi began a pilot deploying agentic AI inside its internal systems, giving employees the ability to initiate complex multi-step tasks via a single prompt. This isn’t just automation. It’s an experiment in letting AI act semi-autonomously across multiple internal systems. It signals what many in fintech have predicted for years: the next generation of banking tools, including advanced software to manage investments, will operate less like helpers and more like coworkers.

The Pilot: What Citi Is Trying

Citi’s pilot involves about 5,000 users within its internal teams. The platform is built on Citi Stylus Workspaces, its proprietary AI workspace, now upgraded to support agentic workflows.

Here’s how it works in practice:

• A user issues a prompt (e.g. “Summarize key metrics for client X, compare to peers, and send me a presentation draft in Spanish”).
• The agent chains tasks: it pulls data from internal and external sources, runs comparisons, generates narrative, translates, and formats.
• Human oversight remains: the agent’s output is reviewed before being finalized or sent.

Citi also uses a combination of models – including Google’s Gemini and Anthropic’s Claude – inside the system, with controls in place to cap cost overruns from long or complex tasks.

Moreover, in parallel with this, Citi has rolled out or expanded AI tools for its developer workforce. It has deployed agentic capabilities to its developers – automating simpler tasks like patching, upgrades, and code rewriting – using an AI agent named Devin from Cognition.

Why This Matters

1. From Tools to Agents

Traditional AI augmentation helps with single tasks (e.g. summarizing, translating, scanning). Agentic AI links tasks and systems together automatically. That shift unlocks more powerful workflows and might reduce friction in how work gets done.

2. Efficiency & Scale

By collapsing multiple steps into one prompt, Citi aims to accelerate internal workflows and free human attention for higher-value decisions. As one internal communication suggests, the goal is to reduce the cycles between idea and execution.

3. Data & Infrastructure Demands

To enable agents, Citi must have tight integrations, reliable APIs, feature stores, version control, observability, error catching, and fallback logic. Any weak link means agents will fail or cause errors. This pilot is also a stress test for Citi’s underlying architecture.

4. Risk, Oversight & Governance

Agentic AI introduces new layers of risk: bias in models, mistaken outputs, leaks of internal data, uncontrolled chaining of actions, regulatory scrutiny. For agents acting on internal systems, banks will need guardrails, logging, rollback ability, and transparency.

In financial services, where decisions can affect capital or legal liability, trust and auditability matter as much as performance.

Challenges & Open Questions

• Cost control: Agents chaining long tasks may trigger heavy computation or multiple external calls, ballooning cost.
• User adoption: Employees must trust agents enough to delegate tasks. If agents make errors, confidence breaks fast.
• Model updating: Agents must stay current with internal policies, data schemas, regulatory rules.
• Explainability: When an agent makes or suggests a decision, employees or compliance teams must understand why.
• Security & Scope: Which systems can the agent access? What permissions? How do you prevent rogue behavior?

What the Industry Thinks

Citi isn’t alone in betting on agentic AI. In its internal “Citigroup GPS” reports, the bank frames agentic AI as pivotal to the rise of a “do it for me” economy in finance – where AI doesn’t just respond, but acts.

At the Citi Gen AI Summit 2025, leadership discussed how agents might assist wealth advisors by combining market data analysis and personalized recommendations. The emphasis was on “agents helping humans” rather than fully autonomous systems.

Meanwhile, in its developer domain, Citi’s use of agentic AI (Devin) is framed as a layered approach: agents handle relatively simple, repeatable tasks, and human developers continue managing higher complexity work. Other companies looking for AI services, may opt for the best software development firms like S-PRO.

What Firms Should Watch / Act On

• Start small: test agents on low-risk internal tasks before moving to customer-facing or decision-making ones.
• Build infrastructure now: integrations, API maturity, logging, safety nets.
• Design with rollback: never let an agent execute something irreversible by default.
• Observe user behavior: watch which prompts succeed, which fail, and where agents frustrate users.
• Regulator-ready: maintain versioning, audit logs, model snapshots, and human override paths.

Wrapping Up

Citi’s agentic AI pilot isn’t just experimentation – it’s a signal. The bank is placing a bet that the future of work in finance will reside not in manually driven dashboard clicks, but in agents that think, plan, and act.

If the pilot succeeds, we may look back at this moment as a turning point: from human doing to human orchestrating.