Case study Financial Services / AI Solutions, Microsoft Copilot Studio, Power Platform

How a UK Specialist Bank Tested AI-Powered Mortgage Underwriting on Copilot Studio

One Orchestrator and seven specialist agents on Microsoft Copilot Studio take on the reading, cross-checking and arithmetic of a mortgage case. Approval, decline and policy exceptions stay with the authorized underwriter.

Headline result 7 agents specialist agents, one per underwriting stage, run in sequence
0 figures LTV, ICR and stress test computed in the generative layer
1 decision owner the authorised underwriter keeps approval, decline and exceptions
Industry Financial Services
Region United Kingdom
Client A UK specialist bank
Solutions AI Solutions, Microsoft Copilot Studio, Power Platform

Key takeaways

  • One Orchestrator and seven specialist agents on Copilot Studio, run in sequence, one per underwriting stage.
  • LTV, ICR and stress-test figures come from deterministic Power Automate logic, not the generative layer.
  • A Case Number carries the workflow; each agent retrieves its documents from SharePoint, not attachments.
  • Approval, decline and policy exceptions stay with a named, authorized underwriter.

THE CHALLENGE

The reading was crowding out the judgement.

A UK specialist bank wanted to know whether AI could prepare a mortgage underwriting assessment without ever making the lending call. Every case arrives as a stack of valuation reports, credit files, bank statements, portfolio schedules, Land Registry extracts and KYC checks.

The underwriter locates the figures by hand, reconciles them across documents, calculates LTV and ICR, checks each finding against lending policy, and catches what’s missing before it reaches a credit paper. Underwriting is regulated, so an AI that quietly fills a gap with a plausible number is worse than no AI at all.

Before

  • Every assessment assembled by hand, one underwriter per case
  • LTV, ICR and stress test worked out manually
  • Missing or contradictory data caught by reading alone

THE APPROACH

Split the work the way the assessment already splits.

We split the assessment the way it already splits: one Orchestrator agent as the underwriter’s entry point, seven specialist agents behind it, running in sequence so each stage builds on what the last one found. The build sits on Microsoft Copilot Studio and the Power Platform.

SharePoint holds case evidence and approved policy knowledge in separate layers. A Case Number travels down the chain, so each agent retrieves its own documents instead of depending on conversation attachments. LTV, ICR and stress-test figures come from deterministic logic in Power Automate.

  1. Underwriter enters Case Number

    Case evidence sits in SharePoint folders keyed to that number.

  2. Orchestrator + 7 specialist agents

    Run in sequence, each stage building on the last one’s findings.

  3. Authorised underwriter

    Reviews the consolidated assessment and makes the lending decision.

THE OUTCOME

The lending decision sits where it always did.

The assembly work in underwriting moves to a controlled agent workflow. The lending decision stays with the authorised underwriter. Time shifts from locating and transcribing figures to testing the assessment and pricing the risk, and the same seven checks run on every case, in the same order. Benchmarks for throughput and accuracy come from running against live case volume.

Outcome
  • Seven specialist agents run in sequence on Copilot Studio
  • LTV, ICR and stress test kept out of the generative layer
  • Every finding labelled by origin for audit
  • Approval, decline and exceptions stay with the underwriter

The assembly work sits with the agent workflow. The lending decision sits where it always did.

Is the reading crowding out the judgement in your underwriting team?

IWConnect builds AI agent workflows on Microsoft Copilot Studio and the Power Platform that take the assembly work out of regulated processes and leave the decision with the person accountable for it.

More proof

Related case studies.

Have a problem that looks like this? Let's talk.

Tell us what you are trying to connect or trust. We will show you the proof that fits and the team that can help.