Banking · AI research agent
Deep Research Agent
An AI agent that writes client and market briefings for bank relationship managers, with a source behind every claim. I led UX research and product design from a blank page.
- Role
- Lead, UX research and product design
- Users
- Relationship managers, analysts, compliance
- Made
- 60+ screens and a component library
- Platform
- Web and mobile, in English and German
A banker must defend a briefing to a client. A blank wait and an unsourced draft leave nothing to defend.
I made the agent’s work visible while it runs and every fact checkable: five named stages and a source on every claim.
60+ screens, key ones in five states, in English and German, from one library. No outcome metrics to share.
The problem
The agent can gather a meeting’s research in minutes, but nobody trusted it: while it ran the screen was blank, and when it finished every fact sounded equally sure.
- Asks
- “What do I need before my meeting?”
- Gets
- A briefing with no training needed.
- Asks
- “Can I trust this number?”
- Gets
- A source next to every claim.
- Asks
- “How did it get there?”
- Gets
- Every step on screen, a person at both ends.
My part
I led UX research and product design, from a blank page to 60+ screens and a component library. Research defined three readers and the question each asks before trusting a briefing.
How a briefing gets made
Four steps, all visible. The agent never works out of sight, and nothing reaches the meeting without a source.
- 1Pick a typeMarket, client, portfolio or KYCSections set
- 2BriefOne required fieldClient and sources
- 3Agent runsFive named stages, on screenA draft, step by step
- 4ReportA source on every claimChecked by a person
- Uses
- 4Market, client, portfolio and KYC, one flow
- Stages on screen
- 5Nothing runs out of sight
- Missing sources
- FlaggedNever filled with a guess
Key decisions
The idea behind every decisionTrust comes from being able to check: the agent’s work is visible while it runs, and every fact can be traced afterwards.
Start from a briefing type, not a blank prompt
Most AI tools open on an empty prompt, which assumes you know what to ask. Here you pick one of four briefing types and see its sections first; all four share one flow.
- Instead of
- An empty prompt box
- Chose
- Four briefing types, sections shown up front
| Report | Type | Status | Sources | Updated |
|---|---|---|---|---|
| NINordhafen Insurance AGYearly review | Client Briefing | Running | 3 of 5 | 2 min ago |
| DIDACH insurance marketQ3 outlook | Market Briefing | Ready | 14 | Yesterday |
| BFBrenner Family OfficeHalf-year check | Portfolio Insight | 1 gap | 9 | Mon |
| ALAlpine Logistics SEOngoing watch | KYC Agent | Watching | 6 | Weekly |
Errors explained at the field, not in a banner
The agent needs one input, what the briefing is for, so that’s the only required field. If it’s empty, the message sits under it and the disabled button says why.
- Instead of
- An error banner at the top of the form
- Chose
- One required field, the message right under it
Show the agent’s work while it runs
Minutes of blank screen read as a black box. The screen names five stages, the current step and every source as it’s read. A stage that needs a person turns amber.
- Instead of
- A spinner and a wait
- Chose
- Five named stages, every source as it’s read
- Outline drafted5 sections · 09:42
- Client file.pdf14 pages read
- Bloomberg6 data series
- NewsVision12 articles, 3 used
- 18 claims checked17 sourced, 1 gap
- Market data is 26 h oldRefresh before adding sources?Refresh
A source on every claim
A confident draft is a risk if nobody can tell which facts are solid. Every claim links to its source, and a claim without one is marked as a gap, not filled with a guess.
- Instead of
- A polished draft with no sources
- Chose
- A source tag on every claim, gaps marked
- Ask how the reinsurance review changes next year’s budget4
- Follow up on the cyber cover question from the last call5
- Confirm last year’s premium figure No source
One page to take into the meeting
The report is for working in; the meeting needs one page. Export briefing sets it as a printable page: what matters at a glance, four talking points, the open items, and every line still numbered to its source. The gap travels with it.
- Instead of
- The working report, exported as it is
- Chose
- One page for the meeting, sources kept
Nordhafen Insurance AG
Yearly review · Thursday 14 November, 10:00
Summary
Nordhafen renewed its commercial property book in March1 and is reviewing its reinsurance limits before the yearly review4. The relationship is stable, with one open question on cyber cover.
Before the meeting
- Renewed the property book in March1
- Reviewing reinsurance limits4
- New CFO since February2
- Open question on cyber cover5
Recent news
The group named a new chief financial officer in February2. Listed insurers in the DACH region gained 6% over the quarter3.
Open items
- Ask how the reinsurance review changes next year’s budget4
- Follow up on the cyber cover question from the last call5
- Confirm last year’s premium figureNo source found, so it is marked, not made up.
Sources
- Client file PDF, page 4
- NewsVision News feed, 12 Feb
- Bloomberg Market data, today
- Meeting notes DOCX, 3 Oct call
- Client emails Mail, 2 threads
Catch a broken feed before a run does
An expired feed key would break a run halfway. All files, feeds and links live in one hub with a live status, so a broken source shows up first.
- Instead of
- Finding a broken feed when a run fails
- Chose
- One hub with a live status per source
| Name | Added by | Date | Used in | On |
|---|---|---|---|---|
| Regulator noticesLink · finma.ch/… | AKA. Keller | 12 Mar | 14 reports | |
| Industry reports6 links | JDJ. Doe | 4 Mar | 6 reports | |
| Client file.pdfPDF · 14 pages | JDJ. Doe | 28 Feb | 2 reports |
One component set for two languages and the phone
Every screen exists in English and German from the same components, so a change lands in both. On the phone, Live Prep runs the same five stages.
- Instead of
- Separate screens per language and device
- Chose
- One component set for both languages and the phone
- Renewed the property book in March1
- Reviewing reinsurance limits4
- New CFO since February2
- Open question on cyber cover5
Results
60+ screens, each key screen in five states, in English and German, from one component library. No outcome metrics can be shared.
How success would be measured
- Prep timeMinutes from opening a briefing to having it ready for the meeting.
- Edits before useHow much a relationship manager changes in a briefing before relying on it.
- Failed runsRuns that break on a missing or broken source.
What I’d change
-
Running The steps have engineer names
“Semantic Grounding” means nothing to a relationship manager. It names what the system does, not what the reader gets.
FixLead with plain words, like “Adding sources”, and keep the technical name as a detail.
-
Mobile Mobile was a guess
Nobody tested the phone screens with relationship managers first.
FixTest one rough screen before building the set.
Numbers: no outcome numbers can be shared; counts come from the design file.
About the screens: rebuilt in HTML from the design file. Client names, dates and figures are examples.
