Morgan Stanley

Morgan Stanley Research publishes daily, and the Research Portal held all of it. Yet clients and analysts still ran on email, phone and Symphony, billed per interaction, conversations scattered across tools with no way to hold a group thread.

AskResearch and the portal were two ways into the same library, one built for browsing, one built for asking. As Senior Design Lead, remote from June 2020 to December 2022, I owned UX strategy across both and built the design system that outlived them.

These are findings: what clients and analysts told us, not outcome metrics. Communication was analyst-initiated and ran over email (the team's estimate put 75 to 90 percent of it there), and most of what arrived was small: “80–90% of client inquiries are short and require 2–3 sentence responses from the team.”

Analysts named the missing increment: “we got the articles, emails, calls, podcast, but no chat.” Bloomberg, the benchmark everyone cited, was found limiting (a new window per group chat, two invitees at a time), and lead analysts wanted triage, not the firehose: “allow my team to help respond.” Synthesis clustered the findings into three themes: engagement and access, chat interaction, research integration.

Synthesis collapsed the findings into three themes: engagement and access, chat interaction, research integration. Each was sharpened into how-might-we statements the squad could ideate against, specific enough to sketch to, open enough to argue with.

The workshop programme ran across the research organisation and fed both products, AskResearch and the Public Appearance Tool, which is why the artefact boards carry PAT themes alongside the assistant's.

Ideation ran in timed rounds, with tech, marketing, analysts and clients in the same room. Dot-voting did the filtering, and what kept surviving the votes was transparency: answers that show their source.

The surviving ideas became wireframe flows, concrete enough for the squad to write the feature list against. Structure was decided while it was still cheap to change, and every grey box from these rounds shipped as a screen in the final designs.

The wireframes are also where the cut happened. An early concept put a personalised “for you” stream on the portal home; it tested as noise, since clients wanted the latest from the analysts they already followed. The stream became a follow list and an alert, and the home stayed a library door.

The candidate features, grouped into epics, went back to the room for a final round of dot-votes. Three to four epics made the MVP; everything else was sequenced behind it. The vote set sprint zero in motion, and the road to MVP was built hand-in-hand with the product owner.

The citation is not metadata, it is the interface. Discovery fails in the last inch, at the moment someone has to trust an answer enough to use it, so every answer names the report it drew from and links it. Where the model is least certain, exactly where the source matters most, weak answers degrade into a guided search rather than a confident guess.

Inside a human conversation the assistant is summoned, not switched to: type @AskResearch, autocomplete confirms the mention, and it answers in-thread, in front of everyone. When the question outgrows the model, the same thread escalates to the analyst or the sector team, and the meeting gets booked from there instead of over email.

The assistant shipped where the conversation already lived: a docked rail on the portal that collapses to a launcher, the mobile app, and inside Symphony. The figures say more than another paragraph would.

Every colour token shipped with its WCAG contrast ratio against the page ground, movement never travels by colour alone, and the type scale holds a twelve-point floor for citations. The citation chip is the component I fought for longest: an AI answer without a visible source reads as a guess, and in a wealth context a guess is a liability.

The core set went out as a Figma library with tokens and docs, and the system was adopted across teams, cutting design-to-dev time by twenty percent. That is the number on this project I am most proud of, because it kept paying after I left the room.

The Research Portal and AskResearch shipped, and the Public Appearance Tool followed inside the Analysts section in 2022: three products over a thirty-month engagement. A relationship that had run on billed interactions moved onto surfaces the firm owns, in front of a client base of more than a hundred thousand.

The assistant took the bottom of the inbox first: the short, repetitive inquiries the research had estimated at eighty to ninety percent of what clients send, the kind that need two or three sentences and a link. Each one the assistant resolved was an interruption an analyst never took, hours handed back to the research clients were actually paying for.

Clients stopped trading time for answers. What used to cost a billed phone call, or arrive as an email a day later, was answered at the moment of need, with the source attached. Meeting prep stopped being a last-minute hunt through documents; the thread already held the reports, the earlier answers and the context.

One thread replaced the five apps the conversation had been scattered across: Symphony, Skype, Slack, email, phone. Because the conversation now lived in one place, its history became institutional memory. A new joiner could read what a client had asked and what they had been told instead of reconstructing it from other people's inboxes.

The citation-first model is what made the assistant viable in a wealth context: an answer that names its source is one compliance can live with and a client can check, and the weak-answer degradation protected that trust by refusing to guess exactly where a guess would be costly. The sector group room let a lead analyst scale through their team, triaging a shared thread instead of answering everything alone.

The design system outlived the engagement. The twenty percent cut in design-to-dev time compounded across every team that adopted the library and every surface that shipped after it, the Public Appearance Tool included. What I took with me: an AI answer is only as good as its citation, and if I ran this again I would prototype the weak-answer states first, not last, because the moments where the system is least sure are where trust is won or lost. That lesson is in Ripplux now, where every recommendation shows its evidence.