What Type of AI Agents Work Well in the Banking Sector?

The AI agents that work in banking are not the most capable. They are the ones that are the most accountable.

Across roughly 10,000 production runs on identical tasks, a structured, scenario-scoped workflow succeeded 15% more often than an autonomous agent: 82% against 71%. Drop the structured workflow to a model one tier down and it still beats the autonomous agent running on the best model available, at roughly 40 to 50% lower model cost. [3][6]

Every bank evaluating banking automation asks some version of the same question first: which agent is the most capable. We think that is the wrong question. Capability is not what fails in banking. Accountability is.

The sector is past the pilot stage and into the part where the numbers get checked. Evident's Q1 2026 tracker, which follows every publicly announced AI use case across the 50 largest banks in the world, found that nearly 1 in 3 new use cases were agentic, up from 15% the previous quarter, and that most of them cluster in product and service operations rather than enterprise-wide copilots. [1] The same tracker found that only 15 of those 50 banks disclose an aggregate return on their AI investment at all. [1]

Read those two findings together. Deployment is accelerating. Measurement is not. McKinsey's banking operations team puts the gap more bluntly: close to 80% of the financial institutions they work with in Asia report using some form of AI-led application, and a similar proportion globally report no significant impact on the bottom line. [2]

The wrong question

Buyers ask which agent is smartest. Production asks which agent is reliable.

We have run the comparison directly. Same tasks, same inputs, architecture as the only variable. Autonomous agents reached 79% task coverage with a 71% success rate. Structured workflows reached 93% coverage and an 82% success rate, succeeding 15% more often on identical work. Then the result that reframes the whole procurement conversation: a structured workflow running one model tier down still reached 91% and 80%, beating the autonomous agent on the best available model. On current published pricing, that tier costs roughly 40 to 50% less per token, and the structured workflow spends fewer tokens anyway, because it does not carry the full chain of context into every call. [3][6]

Structure was doing more work than model capability. That has a direct commercial consequence. If you are selecting agents on benchmark scores, you are optimising the variable that matters least.

Three traits that survive banking

The agents that hold up in a regulated operation share three properties. None of them is intelligence.

1. Scenario-scoped. The agent handles one recognisable situation, not a category of work. Not "customer support" but "customer disputes a duplicate debit on a closed card." Scope is what makes behaviour predictable enough to test.

2. Deterministic verification. Between steps, a rule-based check, not another model call. Schema validation, permission enforcement, parameter ranges. Extracted values are checked against internal records before anything is written, sent, or held. A second model can be talked out of its instructions. A rule cannot. This is the line between an agent that drafts and an agent that can be relied on: a response built on validated data can be defended line by line.

3. Escalates when uncertain. The uncertain path is designed in, with the data already gathered and the reasoning logged. An agent that never escalates is not more capable. It is less honest about its own confidence.

The scope paradox

There is a real objection to scenario-scoped agents, and it comes from McKinsey. Their banking operations partners warn about the "fingers and toes problem": automating only part of what an individual contributor does, which is why the previous automation wave underdelivered. They also list narrow point solutions as a reason AI impact plateaus in financial institutions. [2]

Both criticisms are correct, and neither is an argument for broader agents. They are arguments against narrow workflows. Scope the agent narrowly so its behaviour can be specified and tested, then cover the workflow end to end so no part of the case falls back to a person by default. In our banking deployments that means 180+ scenario-specific agents inside a single workflow, with a separate verifier agent checking safety and correctness on top of them. One broad agent per department gets you the fingers and toes problem. A handful of narrow agents and no workflow gets you the plateau.

Classic RPA (robotic process automation) is scoped narrowly too, but to a screen rather than a scenario. A UiPath, Automation Anywhere, or Blue Prism bot needs structured data and a stable interface, so free-text complaints and a redesigned portal break it. Narrow scope only works if the thing being scoped is a decision, not a click path.

The market appears to be converging on this. Evident found that specialist vendors outside the hyperscalers now account for 68% of banking AI deployments, concentrated in workflow-specific applications across credit, anti-money laundering, and treasury. [1] Gartner expects 40% of enterprise applications to be integrated with task-specific agents by the end of 2026, up from under 5% today. [4] The direction of travel is narrower, not broader.

Which AI agents work in banking operations?

The pattern across all four: the agent does collection, extraction, and preparation. The institution keeps the decision.

What does banking automation actually return?

Patched runs back-office operations for banks and NBFCs (non-banking financial companies): fraud ops, customer service ops, and reconciliation. Since only 15 of the 50 largest banks publish an aggregate AI return, it is worth stating exactly what we mean by ours. Each figure below is per workflow, measured against that workflow's own pre-deployment baseline, and covers the cost of the deployment rather than the cost of a wider programme.

Across the three, the return runs from 2.7x to 6.1x. We are not going to average them, and the spread is the more useful number anyway, because it tells you where returns come from. Reconciliation is already partly automated in most institutions, so the remaining gain is narrower. Customer operations is high volume and almost entirely manual, so the same architecture returns more than twice as much. If a vendor quotes you a single multiple that holds across every workflow, ask them what the denominator was. Treat 3x as the bar. Below that, a deployment rarely justifies the time, money, and change management it costs. Two of these three clear it comfortably; reconciliation sits just under, which is an argument for sequencing it behind the other two rather than leading with it.

Two supporting figures from the same deployments: 86% of support emails are now handled end to end, with the remainder going to human review by design, and turnaround on a live deployment has moved from more than three days to under an hour.

Speed is where the second-order effect shows up. As one fraud and risk lead on a live deployment put it, "false positives dropped while escalation quality improved, which is exactly what risk teams need." When evidence assembly is consistent, the cases that reach people are the ones that deserve people.

The honest summary

Structured agents cost more to design upfront. That is real, and the objection deserves a straight answer: specifying scenarios one at a time is more work than pointing a capable model at an inbox. What has changed is the ratio. AI-assisted development has cut the cost of building structured pipelines, while the cost of debugging silent failures at production scale has not moved at all.

That does not mean scripted automation goes away. Where a bank already runs RPA bots against stable interfaces, the practical answer is not to rip them out but to put an agent above them. The agent reads the unstructured input, decides what should happen, drives the bot, and checks the result. The bot keeps doing the deterministic execution it is genuinely better at, and the agent supplies the judgment and the monitoring that classic RPA never had, noticing the moment a portal changes rather than three days later. RPA orchestrated by agents is a faster path to production than replacing either one. The case for agentic AI for banking operations starts where free text, judgment, and changing interfaces do.

Structured agents still route edge cases to people. That is the feature, not the gap. A banking workflow that claimed to resolve every case without review would be a red flag to any regulator reading it.

And accountability does not transfer. The institution owns every response, every hold, and every reply that reaches a customer, which is why each step has to be traceable. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. [5] Model capability is not on that list.

Banking does not need agents that can do anything. It needs agents that can explain everything.

Patched builds structured workflow AI for banking and financial services operations, from fraud ops and service ops to reconciliation. If you're comparing agent architectures for a banking workflow, we're easy to find.

Citations

[1] Evident Insights, AI Use Case Trends in Banking, Q1 2026, published May 2026. Tracks every publicly announced AI use case across the 50 banks in the Evident AI Index. Agentic applications reached 31% of new use cases, up from 15% in Q4 2025; 38% of Q1 use cases reported outcomes and 15 of 50 banks disclose aggregate AI ROI; specialist vendors outside the hyperscalers account for 68% of deployments.

In plain terms: an independent scoreboard of what the world's biggest banks have actually deployed. It is the source for how fast agents are arriving and how rarely their returns are published.

[2] McKinsey, The paradigm shift: How agentic AI is redefining banking operations, 27 Feb 2026. Abhilash Sridharan: nearly 80% of financial institutions in Asia report using AI-led applications while a similar proportion globally reports no significant bottom-line impact; narrow point solutions listed as a cause of plateau. David Deninzon: the "fingers and toes problem" as the reason earlier automation underdelivered; 50 to 60% of bank FTEs (full-time equivalents) tied to operations.

In plain terms: two McKinsey banking partners explaining why most bank AI programmes have not shown up in the numbers yet. It supplies both the adoption paradox and the strongest counter-argument to our own position.

[3] Patched, The Case Against Autonomous Agents. Roughly 10,000 runs across 150 outcome scenarios: autonomous 79% coverage / 71% success; structured 93% / 82%; structured one model tier down 91% / 80%.

In plain terms: our own controlled comparison, same tasks and inputs, architecture as the only variable.

[4] Gartner, press release, 26 Aug 2025: 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% today.

In plain terms: the analyst view that the market is moving toward narrow, task-scoped agents rather than general ones.

[5] Gartner, press release, 25 Jun 2025: more than 40% of agentic AI projects cancelled by end of 2027, on escalating costs, unclear business value, and inadequate risk controls. Same release: Gartner estimates only about 130 of the thousands of self-described agentic vendors are real.

In plain terms: the failure forecast, and the reasons given. None of them is that the models were not smart enough.

[6] Published model pricing, checked 28 Jul 2026. Anthropic lists Claude Opus 4.8 at $5 per million input tokens against Claude Sonnet 4.6 at $3; OpenAI lists GPT-5.5 at $5 against GPT-5.4 at $2.50. One tier down is roughly 40 to 50% cheaper per token before batch or cache discounts.

In plain terms: the arithmetic behind the cost claim, taken from the providers' own list prices rather than an internal estimate.