For a while now, AI in credit has mostly meant scoring. A model looks at an application, produces a number, and a human decides what to do with it. That arrangement is starting to change, and not in the direction most people expected.
It's one thing for AI to tell you what the data says. It's another for it to actually do something with it. Pull the document, check the numbers against other systems, catch what's off, write up the findings, and pass the case to a credit analyst with most of the groundwork already done. That's the shift: AI is moving from advising the workflow to doing parts of the work.
You can see how fast banks expect this to move. Accenture's Top Banking Trends for 2026 report found that over the next three years, 57% of banking executives expect AI agents to be fully embedded in risk, compliance, audit, and fraud detection functions, and 56% believe agents will reach broad adoption in credit assessment, loan processing, and KYC specifically.
BNY is already showing what that looks like in action: the bank has built an internal AI platform called Eliza that lets employees design their own agents, and it's working toward 150 AI-powered offerings across its operations.
That's the direction. Here's where it actually shows results, where it doesn't yet, and what it means for the teams who have to make it work safely.
Where the gains are already real
McKinsey's own research on multiagent systems offers one of the few named, methodical field results rather than a survey estimate. The firm analysed a US bank that used a multiagent system to prepare credit memos and found credit analyst productivity gains of 20 to 60%, along with roughly 30% faster decision-making. That's not a chatbot answering questions faster. That's agents doing the actual preparatory work, document review, data cross-referencing, first-draft memo writing, so the analyst reviews a pre-assembled case instead of building one from scratch.
McKinsey's broader work on agentic AI in corporate credit is honest about how uneven this still is. In its 2025 Global Survey on the state of AI, nearly two-thirds of organisations said they haven't yet begun scaling AI across the enterprise. A separate McKinsey survey of credit risk leaders found that regional banks, not the largest megabanks, are actually furthest ahead on deployment, with productivity uplift the single biggest reason institutions greenlight a use case in the first place.
Look at those two findings together, and there’s a pretty obvious gap. Some lenders are already getting real value from agentic AI. Most others haven’t got there yet. The ones that close that gap first could have a serious edge in lending.
The governance question nobody's fully answered yet
This is no longer just a question of what AI agents can do. For ops, risk, and engineering teams, it's also a question of how to govern them.
In April 2026, the OCC, the Federal Reserve, and the FDIC jointly replaced the long-standing SR 11-7 model risk management framework with revised guidance, OCC Bulletin 2026-13. That new guidance explicitly states that generative AI and agentic AI models are "novel and rapidly evolving" and are, as a result, not within its scope. The regulators have said they plan to issue a request for information specifically on banks' use of generative and agentic AI, but for now, there's a real gap: the newest, fastest-moving layer of a bank's credit stack is the one layer the formal model risk framework doesn't yet cover.
That doesn't mean agentic AI in credit decisioning is unregulated in practice. The OCC's own May 2026 Semiannual Risk Perspective describes supervisory support for a "measured approach" to integrating generative and agentic AI into banking operations, provided human oversight stays embedded in the workflow, and notes banks may consider expanding agentic AI into material financial decisions over time. Regulators are watching closely. They just haven't finished writing the rulebook for what they're watching.
For a risk team, that's arguably harder to manage than a fully specified rule would be. You can't point to a single regulatory checklist and confirm you've satisfied it. What you can do is build the same discipline the old framework required, documented governance, an inventory of what each agent is doing, monitoring, and an audit trail that traces every action back to the rule and data point behind it, and apply it to agentic systems now rather than waiting for the rulebook to catch up.
Why this has to run inside the LOS, not next to it
This is the part that decides whether agentic AI in credit becomes a genuine advantage or a genuine liability: where the agents are allowed to act.
An agent that can pull data, draft a memo, or pre-approve a segment of applications is only as safe as the guardrails it's operating inside. If those guardrails live in a separate layer sitting outside the origination system, every agent action needs its own audit trail, its own policy check, its own way of proving to a regulator what happened and why. That's fragile even with one agent. It gets worse with five.
If the agent instead operates inside the LOS itself, against the same policy engine that governs every human-initiated decision, the audit trail is already there. The eligibility rules the agent has to respect are the same ones a human underwriter respects. When a regulator asks how a decision got made, the answer isn't "let us go check the AI layer separately." It's the same answer you'd give for any other decision the system made, because structurally, it is the same system.
That’s the bigger change here. The LOS isn't just where an application gets processed anymore. It's becoming the execution layer, the structured, policy-bound environment where an AI agent is allowed to act because every action it takes is already inside a system built to log, explain, and defend that action.
What that actually looks like inside FinBox LOS
It's easy to describe an "execution layer" in the abstract. Here's what it looks like once you get concrete about it, because FinBox LOS is built around exactly this idea in a few specific ways.
Ask AI for policy simulations. Before a risk team changes an eligibility rule or a pricing threshold, they can ask the system to simulate the change first, running it against real scenarios to see the impact before it ever touches a live borrower. That's the difference between finding out a policy change was too aggressive after a quarter of bad loans, and finding out before you ever ship it.
Convert documents to policies. A lot of what slows down compliance and risk teams isn't deciding what the policy should be, it's the manual work of turning a policy document or a new regulatory requirement into something the system can actually enforce. FinBox LOS can take that document and turn it into an executable credit policy directly, cutting out the manual configuration step that usually sits between "we decided this" and "the system is actually doing this."
MCP. Sentinel AI connects to external tools and systems through MCP, so an agent isn't limited to what's already inside the LOS. It can reach out, pull data, and take actions across other applications a lender's stack depends on, without every one of those connections needing to be hand-built and separately audited.
Agentic Studio. This is where the pieces come together into something a lender's own team can build and manage directly:
- AI agents that handle basic checks before a credit manager ever sees the file, with only genuine exceptions routed to a human. The product's agent library grows over time rather than shipping as a fixed set.
- Agentic workflows where an agent can execute multiple steps in sequence and decide what to do next, rather than running one fixed script regardless of what it finds.
- Tools agents can call on to fetch data, hit APIs, or run specific checks, the actions an agent needs in order to actually do something rather than just recommend something.
- Skills, reusable capabilities built for specific jobs like document analysis, income assessment, or interpreting a policy, so the same capability doesn't get rebuilt from scratch every time a new agent needs it.
Put together, that's the practical version of the governance argument above. The agent isn't operating in some separate AI layer nobody can fully account for. It's operating inside Agentic Studio, using tools and skills a risk team can see, against policies that were either simulated before launch or generated directly from the compliance document that created them, with exceptions still landing on a human's desk. The audit trail isn't something you build after the fact. It's a byproduct of how the system was built to work in the first place.
What this means for your teams, starting now
- For ops: the manual, document-heavy steps agents are already good at, extraction, cross-referencing, first-draft memos, are the first candidates to hand off. The productivity gains McKinsey found aren't theoretical anymore, and tools like document-to-policy conversion turn what used to be manual configuration work into something an agent does directly.
- For risk: the OCC's current guidance gap means you can't outsource your governance thinking to a regulator's checklist yet. Simulating a policy change before it goes live, and keeping every agent's actions inside one auditable workflow rather than a separate AI layer, is how you build that discipline in now rather than retrofitting it later.
- For engineering: the fewer separate systems an agent has to reach into, the smaller the surface area for something to go wrong, or to be unexplainable later. A structured environment like Agentic Studio, where agents draw on shared tools and skills instead of bespoke integrations per use case, keeps that surface area from growing every time a new agent gets added.
