Financial services — wealth managementGrounded retrieval & workflow automation

An AI assistant that gives wealth relationship managers a third of their week back

Relationship managers at a private bank spent their weeks hunting for market views, building comparison decks and writing up meetings. We built an assistant that answers in seconds, builds the deck and logs the notes — about 35% of RM time unlocked, delivered in 40 days.

Four everyday requests from a relationship manager — what a rate cut means, a comparison of two funds, who to call first today, and speaking during a client meeting — go to one RM assistant. A request router picks the skill that owns each job and keeps the house view in every prompt; an Explain-It layer shows the sources and rules behind any answer. The skills are Quick Answer and Bullet Builder for cited answers in 10 seconds or less and talking points, Prep Deck for a three-slide Sharpe, beta and draw-down comparison, Call Prioritiser for the top five clients to call after a market move, and Smart Notes and Mailer for tagged notes and a recap email in four minutes or less. All of them read one house knowledge map of CIO research, house-view notes, around 400 fund fact sheets, news feeds and meeting transcripts. No client personal data enters the loop, and every number clicks back to the report it came from. Results along the bottom: about 35% more RM time, 70% less time searching, 80% less deck preparation and nine more calls per RM per week.
The RM asks the way they would ask a colleague. The assistant decides which skill does the work.

Project snapshot

Client
The wealth-management arm of a private bank, whose relationship managers advise its wealth clients
Industry
Financial services — wealth management
Business function
Relationship management and client advisory
Challenge
Relationship managers lost around 15 hours of a 45-hour week to admin: searching several sources for a market view, building portfolio comparison decks by hand at three hours each, triaging who to call after a market move and writing up every meeting. Only the top fifth of clients got a timely call.
Solution
One RM assistant over a knowledge map of the bank’s research, house-view notes and around 400 fund fact sheets. It answers market and product questions with source links, turns overnight news into talking points, builds a three-slide Sharpe, beta and draw-down comparison from two tickers, ranks the top five clients to call, and transcribes meetings into tagged notes and a compliant bilingual recap.
Result
About 35% more RM time: 70% less time searching, 80% less deck preparation and 100% of meetings auto-logged, turning into nine more proactive client calls per RM per week — delivered in 40 days across five sprints.

Key outcomes

≈ 35%
More RM time for clients
70%
Less time searching for market insight
80%
Less time preparing client decks
+9
Proactive client calls per RM per week

The client

A wealth business that sells advice, one conversation at a time

The wealth-management arm of a private bank, where relationship managers advise clients on markets, funds and portfolios. Its growth depends on how often those managers talk to clients, and how well prepared they are when they do.

Most of an RM’s week, though, went into getting ready to talk rather than talking.

The challenge

Relationship managers doing research, design and admin

Every market view was hunted down by hand, every comparison deck was built by hand, and every meeting was written up by hand. Market insight alone took around nine hours a week. A portfolio comparison deck took three. Notes and follow-up emails took another 45 minutes per meeting.

Where the week went

≈ 15 h
of a 45-hour week on search, prep and admin
Top 20%
of clients getting a timely call after a market move

When markets moved, there was no time to work out who was affected. The largest clients got a call. Everyone else waited.

Constraints

What we had to design around

  • Scattered sourcesThe house view lived in CIO research PDFs, chat threads, email and shared drives. In the RMs’ own words, they had to go to different places to extract the information — and pulling it by hand was double work.
  • One voiceA junior RM and a senior RM had to give a client the same view. The assistant had to speak the bank’s playbook, not improvise its own.
  • ProofAdvice to wealth clients gets audited. Every number had to link back to the specific report it came from, so an auditor could trace it in under two minutes.
  • Client privacyNo client personal data in the first release. The model had to stay contained within the information the bank gave it, with encryption, a consent toggle and a watch-list screen built in.
  • GrowthThe first release sat apart from the core banking platform. The same engine had to take live client positions later — and other business lines after that — without being rewritten.

Our approach

Start from the RM’s day, not from the model

We mapped the RM’s week first, stage by stage, and put an hour figure on each one. Market insight, messaging, client prep, triage, meetings, follow-up and audit each became a job with its own measure: time to insight, deck-prep minutes, the share of meetings logged.

We then interviewed RMs from junior to top performer and collected the 50 questions they asked most. Those golden questions became the test set. A skill was finished when it answered real ones quickly, with sources an auditor would accept.

The solution

Six skills, one knowledge map, one place to ask

  1. Quick Answer — Ask any market or product question and get a house-approved answer with three source links. Hover-tips explain Sharpe, beta and the rest, so a junior RM can explain them as well as a senior one.
  2. Bullet Builder — One click turns overnight news and house-view notes into client-ready talking points, in the bank’s style.
  3. Prep Deck — Type two fund tickers and get a three-slide comparison — Sharpe ratio, beta, draw-down — drawn from around 400 fund fact sheets. The assistant writes a short “why this matters” note beside each chart.
  4. Call Prioritiser — After a market move, a daily list of the top five clients to call, with a script for each. Alerts are throttled to three a week so they stay worth reading.
  5. Smart Notes and Smart Mailer — The RM speaks; notes appear, with tasks and client interests tagged. One click drafts a bilingual recap with compliant phrasing and the disclaimer already in place.
  6. Explain-It — Click “Why?” on any answer to see the sources and the rules behind it. The same trail is what audit queries.
Five sprints over 40 days. Sprint one, understand and gather: RMs from junior to top performer interviewed, 50 golden questions collected, research and fund sheets loaded into a search-ready, read-only vault. Sprint two, sketch and prove: Quick Answer, search over the house research with source links on every answer, and Bullet Builder sketched with RMs. Sprint three, add muscles: Bullet Builder turning news into talking points, the Prep Deck wizard, and the Call Prioritiser with RMs defining what urgent means. Sprint four, capture and draft: the team sat in live client meetings to set note tags, then built Smart Notes and Smart Mailer; a compliance dry run confirmed zero personal data and that the guardrails held. Sprint five, bake-off: speed tuned, KPIs logged, and a mixed panel of junior and senior RMs scored the tool for ease of use. Day 40: the assistant is live, answering the top 50 RM questions in 10 seconds or less, comparing products, building decks, drafting talking points and logging meetings.
The golden questions came first, so every skill was built against something an RM actually asks.
40 daysfrom kick-off to a working assistant, delivered in five sprints

Responsible by design

Four rules for an assistant that talks to advisers

  • Ground

    Answers come from the bank’s research, not the model’s memory.

    A knowledge map links reports, house-view notes, fund sheets and transcripts, so an answer connects facts instead of guessing at them.

  • Cite

    Every number clicks back to its source.

    Each answer carries its source links, and Explain-It shows the rules applied. An audit-trail query takes under two minutes.

  • Contain

    No client personal data in the loop.

    Call ranking runs on a synthetic client table tagged only by sector and geography. Transcripts are redacted, and a compliance dry run confirmed zero personal data before release.

  • Stay on message

    The house view is in every prompt.

    Prompts embed the bank’s house-view rules and follow-up emails use approved phrasing, so a client hears one consistent view whichever RM they speak to.

The second rule is the one RMs asked for. A link to the report behind a number is what lets an RM repeat it to a client with confidence.

Results

What changed for the RMs

  • About 35% more RM time — roughly 15 hours of a 45-hour week that used to go on searching, preparing and writing up now goes back to clients.
  • 70% less time searching for market insight, with the top 50 RM questions answered in 10 seconds or less, each with its sources.
  • 80% less deck preparation, from two tickers to a three-slide comparison and one-pager.
  • 100% of meetings auto-logged, with a bilingual recap and action list out in four minutes or less — about 25 minutes of admin saved per call.
  • Nine more proactive client calls per RM per week — about 450 a year — and 4–6% organic growth in assets under management per RM per year from more frequent contact and sharper talking points.

RMs stopped spending their mornings getting ready to call clients and started calling them.

Beyond the numbers

What else changes

Junior RMs sound like senior ones. Hover-tips on Sharpe and beta, the house view in every answer and approved phrasing in every email close the gap between a first-year RM and a seasoned one.

And because the engine reads whatever feeds it, live client positions are an addition, not a rebuild — turning the same skills into per-client briefs and data-driven alerts, and later opening them to other business lines.

If your advisers spend more time preparing than advising

The first question is where their hours go. That is answerable from a week of their own questions, before anything gets built.