AI underwriting platform for institutional real estate · design exercise, client confidential
AI Underwriting — Five Decisions, Not a Thousand Actions
Ten agents read 47 documents. The analyst should only see where their work stopped.
The challenge
The client builds a multi-agent system that underwrites real-estate acquisitions for institutional investors. The brief: design the core experience for an analyst working through an £80m UK logistics deal, and make the AI’s work trustworthy without making the analyst manage it. Behind the deal sit 47 documents, 10 specialist agents and 1,240 consistency checks. Every product in this space is tempted to show that machinery: agent timelines, activity feeds, a confidence score on everything. It demos well, and it hands the analyst a thousand things to supervise. But the analyst’s job is not to watch agents. It is to form a view they can defend at an investment committee. The question became: what is the smallest surface that keeps the evidence, the control and the consequence of every judgement?
What I did
1. Showed only where the work stopped. The system extracted 412 facts; 407 are uncontested and fold away in grey. Of 18 discrepancies it resolved 13 on its own, where one source is clearly authoritative. What reaches the analyst is a verdict line and five rows: the judgements no agent can make. The machinery is one collapsed line, not a dashboard.
2. Made provenance the structure of the table. Each row opens in place into three columns: what the document says, what the system made of it, and what you decide. A citation opens the actual page with the sentence highlighted. I left out confidence scores on purpose: “92% confident” hides the difference between an extracted fact and a multi-step inference, while a citation to page 34 promises nothing and shows the source.
3. Kept the consequence in the same frame as the choice. A pinned strip carries the live return. Swap the seller’s £8.50 rent for the comparable £7.60 and levered IRR drops from 14.2% to 12.8% as you watch, while the mandate line flips from “clears by 120 bps” to “short by 20 bps”. There is no recalculate button anywhere in the product.
4. Let the product disagree with the deal. A tenth agent reads outside the pack — planning registers, infrastructure, tenant news — and finds a consented development that shares the site’s only lorry access. On the case path the five judgements take the return to 12.0% against a 13.0% hurdle, so the primary action is not Approve. It is “Send to VP — recommend counter at £75.4m”. Leave the assumptions alone and it recommends proceeding.
The solution
A working browser prototype of the underwriting flow: a verdict, five decisions, cited evidence, editable assumptions, returns that recalculate from the analyst’s own choices, and a handoff to the VP with the recommendation computed from them. Accepting a risk unchanged needs a written reason and an owner, and that reason travels into the investment pack. I deliberately left out a chat panel, an agent-activity view, modals and section navigation. Each would have put the complexity back. The deal, tenants and figures are invented but internally consistent, and the arithmetic ties together on every screen. The brief allowed one working day.
What I’d do differently
I designed the analyst’s side and stopped at the handoff. The decision only matters if a VP can challenge it, so the reviewer’s view — what changed, where the analyst overrode the system, what to probe — deserved a rough sketch in the same day. And the central bet is untested: I found no research showing that citations make people trust AI output more. Next I would remove evidence access for some sessions and watch whether analysts open sources less over time. If they don’t, the product isn’t earning trust, just being tolerated.