Agentic AI in Financial Services: What Is Actually in Production

A chatbot answers a question. An agent takes steps toward finishing a task: it pulls documents, checks them against rules, drafts an output, and hands off or acts. That difference is why banks treat agents differently from earlier AI. They don’t just recommend, they do. Teams exploring this space often start with AI Agents & Automation, where guardrails and human review are designed in from day one.

1. Software engineering. This is the clearest production story. Goldman Sachs has hundreds of AI coding agents writing production code alongside roughly 12,000 human engineers. The bank has said it expects to scale from there, and its CTO has described starting with hundreds of agents and potentially growing to thousands. Code is a good first domain because output can be tested, reviewed, and rolled back. Building this kind of capability usually means investing in AI Product Engineering alongside the agents themselves.

2. KYC and client onboarding. Onboarding is document-heavy, rule-bound, and slow, which makes it a natural fit. JPMorgan has described an agentic-first onboarding system for clients, aimed at cutting a process that took up to five days to under a minute. That is a stated target from an executive, so treat the exact numbers as ambition rather than audited results. Goldman is also working in this area: it has been developing Claude-powered agents with embedded Anthropic engineers for accounting, KYC, and internal operations.

3. Back-office operations. Trade accounting, reconciliation, and month-end close are repetitive, high-volume, and rule-heavy. Goldman’s agents are positioned as productivity boosters for back-office and developer roles rather than direct replacements. Expect most real deployments to look like this: narrow, supervised, and unglamorous.

4. Employee-facing assistants. Tools that help advisors and analysts find and synthesize internal information are the most widely adopted. Morgan Stanley reportedly reached 98% adoption among financial advisor teams. These are closer to copilots than fully autonomous agents, but they are the foundation many agent projects build on. Firms building similar tools typically need Generative AI & LLM Solutions grounded in their own data, so answers stay accurate.

Industry data suggests banks are further along than most sectors, but unevenly. An OutSystems survey of 320 banking IT leaders found that 52% rate their agentic capabilities as advanced or expert, while only 12% have moved 76% to 100% of their agentic projects from pilot to full production. That is a vendor survey, so read it directionally. A separate report cited in a compliance-governance article found that only 15% of CFOs at financial institutions say they are ready to deploy AI agents, with governance, traceability, and human oversight as the top barriers. Source: OutSystems.

So “in production” often means a few agents in a few workflows, not a bank run by agents. Closing that gap is where AI Strategy & Consulting helps: deciding which AI bets will pay off before you spend on them.

Based on public reporting, I haven’t found credible evidence of agents autonomously making credit decisions, executing trades without human limits, or moving customer money unsupervised at scale. The pattern is agents on internal, reviewable work first, and customer-facing or irreversible actions later.

Regulators are openly saying the old playbook doesn’t fit:

• The UK FCA’s chief executive has said existing frameworks weren’t built for autonomous agents, and requiring a human in the loop for every agent action is unlikely to be realistic.
• Bank of England roundtables heard from firms that traditional model risk validation wouldn’t be sustainable as agentic systems multiply, and that human-in-the-loop is being challenged.
• On October 15, 2026, the Financial Stability Board published final guidance calling for risk-based human oversight, with clear conditions for when an agent must seek approval, and more intensive oversight for higher-risk uses. Source: TechInformed.
• In the US, a Treasury-backed framework released in February 2026 translates the NIST AI risk framework into 230 financial-services-specific control objectives.

The shift is from “a human approves everything” to “controls, testing, and monitoring around outcomes, with humans at defined decision points.” That makes monitoring and deployment discipline critical, which is the focus of MLOps & AI Infrastructure: running models in production without surprises in cost, accuracy, or uptime.

1. Start where outputs are checkable. Code, reconciliations, and document review beat open-ended decisions.
2. Build the harness, not just the agent. Evaluation, routing, audit trails, and escalation rules are what let you scale.
3. Define approval points by risk. Low-risk steps run freely, and high-risk ones pause for a human.
4. Plan for concentration risk. Many firms relying on the same model vendors and clouds is a systemic concern regulators are watching.
5. Measure honestly. Separate pilot results from production results.

Most of the gap between pilot and production comes from choosing the wrong first workflow. Before funding an agent, a bank or fintech team can test each candidate use case against five practical questions.

• Can a person check the output quickly? Code can be tested and reconciliations can be compared against a ledger. If nobody can verify the result without redoing the work, the agent saves little time and adds risk.
• What happens when the agent is wrong? A wrong draft costs a few minutes of review. A wrong payment or credit decision can harm a customer and draw regulatory attention. Map the worst realistic outcome before deciding how much autonomy to allow.
• Is the data ready and permissioned? Agents act on what they can reach, so they need clean sources, access limited to what each task requires, and a clear record of which documents shaped each output.
• Which steps need human approval? Set approval points by risk rather than asking a person to sign off on everything. Reading documents can run freely, while actions that move money or change a customer record should pause for a person.
• Can you measure the result in production? Agree the baseline before launch: handling time, error rate and escalation rate are common starting points. Report production numbers separately from pilot numbers, because pilots usually run on easier cases and cleaner data.

A use case that passes all five is a good candidate for a limited rollout. One that fails on the first two usually belongs behind a human, with the agent preparing the work rather than finishing it. Starting there lets a team build the evaluation, logging and escalation tooling that later, higher-risk agents will depend on.

What is agentic AI in financial services?

Agentic AI refers to AI systems that take steps toward finishing a task, such as pulling documents, checking them against rules and drafting a result, rather than only answering questions. In banking it is used today mainly in software engineering, KYC and onboarding, back-office operations and employee-facing assistants.

Are banks using AI agents to make credit decisions?

Based on public reporting, not autonomously. Agents prepare documents, checks and drafts, while credit decisions, trading and customer outcomes still go through human approval and model risk controls.

What is the biggest barrier to using AI agents in banking?

Governance more than technology. Firms need evaluation, audit trails, approval points based on risk and monitoring around outcomes, which regulators such as the UK FCA and the Financial Stability Board are now asking for.

Where should a financial firm start with agentic AI?

Start where outputs are easy to check, such as code, reconciliations or document review, build the controls around the agent, and measure production results separately from pilot results.

Agentic AI in financial services is real, but narrower than the headlines suggest. It is strongest in engineering and operations, growing in KYC and onboarding, and still cautious near customer money and credit. The firms pulling ahead are the ones that did the unglamorous work of governance first.

Planning agentic AI for a financial services workflow?
Novatore builds secure, audit-ready technology for fintech.

You can explore all our services