30-40%
of commercial lending resources are still consumed by manual data entry, quality checks, and document review conducted across multiple fragmented systems
Source: KPMG, “Accelerating Commercial Lending with Artificial Intelligence,” 2022
The commercial lending automation gap
Consumer lending is already highly automated, yet commercial lending is not. KPMG estimates that 30-40% of commercial lending resources are still consumed by manual data entry, quality checks, and document review conducted across multiple fragmented systems.
The gap is not a technology problem alone. Commercial lending resists automation because behind every decision sits a judgement call that score-only systems struggle to make. Borrowers vary widely in structure, financial profile and funding purpose, and many applications sit in the grey area between a clear approve and a clear decline.
The case a credit score can’t read
Consider this: a manufacturing business applying for £450,000 of growth funding. Its file shows two missed direct debits in the last quarter. A score-only system sees a conduct red flag and routes the case to decline or a slow manual referral queue.
An experienced underwriter sees something different. The missed payments coincide with the business’s largest customer paying 60 days late, account conduct has fully recovered since, the order book is growing, and the funding purpose directly addresses the concentration risk that caused the wobble in the first place.
WHAT THE SCORE SEES
- Two missed direct debits
- Conduct risk
- Decline or manual review
WHAT AN UNDERWRITER SEES
- A major customer paid 60 days late
- Account conduct has since recovered
- The order book is growing
- The funding addresses the original concentration risk
The signal is real, but it is temporary and explainable, not structural. That difference matters to the decision, and it is the kind of difference a credit score cannot express.
Funding Xchange (FXE), a UK-based credit decisioning specialist and SME lending marketplace connecting small businesses to more than 70 lenders and providing white-label credit decisioning software to over 80 bank brokers, set out to solve this problem. Working with Cloud Combinator, an AWS Partner specialising in agentic AI, FXE built a multi-agent underwriting platform on Amazon Bedrock AgentCore that reasons through ambiguity the way experienced underwriters do, while keeping humans in control of every decision.
Why it matters at scale
At FXE’s scale, with thousands of applications flowing monthly across more than 80 brokers, even small inconsistencies in underwriting compound into material risk. Two underwriters reading the same conduct signal differently is an anecdote on one case and a portfolio problem across a year of decisions.
The cost is also human. Every grey-area case that sits in a queue waiting for a reviewer is a broker waiting for an answer and a small business waiting for funding. And every insight a senior underwriter brings to a tricky case is used once, then buried in the file — never systematically reused across the portfolio.
01 • CONSISTENCY
Small inconsistencies become portfolio risk
Two underwriters reading the same conduct signal differently is an anecdote on one case and a portfolio problem across a year of decisions.
02 • SPEED
Grey-area cases wait in the queue
Every grey-area case waiting for a reviewer is a broker waiting for an answer and a small business waiting for funding.
03 • EXPERTISE
Senior insight is used once, then buried
What a senior underwriter brings to a tricky case is used once, then buried in the file — never systematically reused across the portfolio.
Agents that argue with each other
The platform works like an experienced credit team. A profiling agent triages each application, specialist agents examine the borrower, account conduct, trading performance and property factors where relevant, and an arbiter weighs the findings into a structured verdict and formal credit paper for human review. Straightforward cases move through a lighter-touch path, while complex cases receive deeper investigation, so cost and processing time match case complexity.
The most distinctive element is the debate stage. On mixed-risk cases, two AI agents are deliberately set against each other: each agent builds the strongest version of its own position, a steelman rather than a strawman, one for the risk and one for the mitigation. This structured adversarial reasoning is how experienced credit committees already work: weak arguments are stress-tested before a decision is made, so the position that holds, does so on the evidence rather than because nobody put the opposing case properly.
For our manufacturing case, one debate agent presses the conduct concern: two missed payments is two missed payments. The other counters with the late-paying anchor customer, the recovered conduct, and the funding purpose. The arbiter weighs both positions to produce a verdict that records not just the conclusion but the reasoning on both sides.
Nothing is decided by any single agent’s initial assessment.
Why adversarial reasoning produces better results
This debate stage is not an unproven technique invented for the AI age. Dialectical reasoning, where a thesis is stress-tested by the strongest available antithesis to arrive at a provably more robust synthesis, has underpinned critical decision-making in law, governance, and business, for centuries.
Credit committees, audit challenges, and boardroom “devil’s advocates” all draw on the same principle: decisions that survive structured opposition are demonstrably stronger than those that don’t.
In underwriting, this matters because the cost of a weak decision is asymmetric. An under-challenged approval creates credit loss, whereas an under-challenged decline rejects a viable borrower and weakens the broker relationship. Both outcomes erode portfolio quality. Both are preventable if the reasoning is contested before the verdict is reached.
Agentic AI provides the ability to apply the same depth of rigorous adversarial scrutiny to every case, at any volume. Lenders can grow their application throughput without proportionally growing their underwriting bench. Decision reviewers’ time shifts from constructing the argument to validating it; a higher-value activity yielding more consistent outcomes across the portfolio.
AWS’s published guidance on operationalising agentic AI in production reinforces this pattern. The AWS Prescriptive Guidance for Operationalizing Agentic AI emphasises building trust through “traceability, runtime controls, and explainability to earn stakeholder trust”, whilst the AWS Well-Architected Agentic AI Lens recognises that autonomous agent operations demand “human oversight patterns purpose-built for autonomous operations.”
FXE’s architecture satisfies both: the debate stage provides explainability through recorded adversarial reasoning, whereas the human-in-the-loop review enforces oversight — keeping the outcomes auditable and defendable to regulators.
In a market where borrowers and brokers increasingly expect immediate, well-reasoned decisions, this architecture delivers what a purely manual process cannot: the quality of a senior credit committee, delivered consistently at the pace of a digital platform.
Humans stay in control, and the system learns from them
Every AI-generated credit paper goes to a human underwriter for review. When the reviewer strengthens the reasoning, the platform compares the two versions and captures the difference as structured data rather than leaving it prose inside an edited document. Recurring reviewer corrections are aggregated weekly into proposed rule changes, and a senior underwriter approves, rejects or edits each one before it enters the knowledge base.
- AI-generated credit paper
- Human underwriter review
- Compare & capture changes
- Proposed rule changes
- Senior underwriter approval
- Knowledge base
The system learns from human judgement, but never autonomously changes the underwriting logic.
Every step — from report generation through review, comparison and approval — creates an auditable trail.
This design choice is key for a regulated environment, where decisions must be explainable as well as repeatable and justifiable. That approach lines up with the eight dimensions of responsible AI that AWS sets out, including transparency, explainability, governance and controllability. The next time an application arrives with missed payments behind a late-paying anchor customer, the AI-generated report weighs the recovery evidence the way FXE’s best underwriter would, in the draft it puts in front of a reviewer rather than in any decision it takes on its own.
Voice of the customer
FXE’s CTO, Matt Collinge, sees the impact in practice:
Funding Xchange
The collaboration, expertise and dedication of the Cloud Combinator team were vital in helping us bring this agentic assistant to life. This is a real workflow change — not minutes shaved off admin — and it is live in commercial underwriting today.
The results
For lending in the £200k to £5M range, a bank today takes 25 days on average to reach a sanction, and an underwriter spends around 7 hours working through each application. The agentic platform cuts time to sanction to a guaranteed 24 hours and reduces the underwriter’s review and credit paper write-up from 7 hours to under 30 minutes.
24 hours
Time to sanction, now guaranteed
down from 25 days
< 30 min
Underwriter review and credit paper write-up
down from 7 hours
Beyond speed, FXE gains consistency across underwriting outputs, reuse of underwriter expertise, and a governed mechanism for continuous improvement that regulated use cases demand.
Conclusion
By combining Amazon Bedrock AgentCore with event-driven AWS services, FXE built an agentic underwriting platform that handles ambiguity through structured specialist analysis and adversarial debate, while a human-in-the-loop workflow captures reviewer judgement and feeds approved changes back into the knowledge base.
The question for commercial lenders is no longer whether AI will enter the underwriting room. It is whether it arrives as a black box, or as a governed, auditable system that earns the trust of regulators, brokers, and underwriters alike.
