SALESFORCE × HCI/D STUDIO · 2025 · 9-person team · 4 months

Agentic Search

We rebuilt search on Slack.com so a buyer could figure out whether Slack fits their team without leaving for Google — AI answers layered over traditional results, never forced on anyone.

Sponsor: Salesforce Design14 interviews5 concepts discarded89% heuristic passMy lane: research + rapid prototyping
01 · The Problem

Buyers arrive with a question. Slack.com answered with a list of links.

Search on Slack.com was an indexing tool: type keywords, get pages. Meanwhile the site's AI assistant lived somewhere else entirely, as a separate system nobody found. But the people deciding whether Slack gets bought arrive with intent, not well-formed queries, and when the site couldn't answer, they left for Google and often didn't come back.

"Why can't I just google it? It's much easier and it gives me a short summary."

P7 · SMALL BUSINESS OWNER
02 · The Research

Fourteen interviews. Four patterns.

Founders, team leads, and IT decision makers: the people who actually decide whether Slack gets bought. Four behaviors kept surfacing. The whole design is a response to them. Tap one for the voice behind it.

"Why can't I just google it? It's much easier and it gives me a short summary."

P7 · Small business owner

"The pricing page felt like there was too much information for a quick glance."

P4 · Director and founder

"I am neutral about AI. I don't mind it if it helps me accurately."

P9 · Small business owner

"Looks like there are more tabs in search than the top navigation bar."

P2 · Team lead
03 · The Pivot

Our sponsors called the first direction "a magic recipe without evidence."

They were right. Mid-project, the audience got cut down twice: from "people who use websites" to "people deciding whether to buy Slack," then to two specific personas: small business owners and enterprise researchers. Every feature decision got easier after each cut, and the concepts that couldn't survive the narrower audience got dropped.

Discarded 01

Hover-to-reveal AI on keywords

Undiscoverable in testing. Almost no one hovered, and the few who did couldn't tell what was AI and what was a tooltip.

Discarded 02

AI summary in a modal

Tested well visually, but every interaction required an extra click to start the conversation. Sponsors flagged it: friction without payoff.

Discarded 03

A separate AI chat history button

Sponsors pushed back. If the conversation is already on the page, scrolling up is the chat history. A dedicated button just added UI.

Discarded 04

Standalone AI chatbot in a sidebar

Preserved the original problem: search and AI as two separate systems the user had to choose between.

Discarded 05

Pre-categorized search dropdown

Users didn't share our categories. The predefined buckets fought their actual mental model of what they were looking for.

04 · The System

AI as a layer, never a mode.

A hybrid results page: AI summary on top, traditional results underneath, and a conversation that follows the user across page navigation. Five screens carry the whole system. Click through them.

slack.com
05 · The Proof

Every stage had its own test.

Observation, interviews, rapid concept testing, and a custom evaluation framework built with our sponsors. The closer to the final, the more we relied on tools we made ourselves.

01 · Contextual inquiry
14
participants observed

We watched people use Slack.com cold, with no task and no script. Most reached for the nav, fell back to search when stuck, and abandoned the site when search didn't help. This is where the project's actual problem got named.

02 · Brainstorm to test
7
features prototyped and stress-tested

Seven features prototyped in Figma Make and put in front of users and our Salesforce sponsors. Some, like contextually rolling prompts, landed immediately. Others got cut by mid-October.

03 · Concept directions
3
directions narrowed to one

Three end-to-end concept directions, each tested against the same query set. The split-AI-and-traditional concept advanced. The other two stayed in the deck as foils.

04 · Custom heuristic evaluation
89%
of our own heuristics passed

Nielsen wasn't built for hybrid AI-search systems. So we wrote our own twelve heuristics with our sponsors: AI/traditional distinction, source citation, conversation persistence, misspelling recovery, and more. Most ratings landed at 4 or 5 out of 5.

8-person Salesforce design panel · 11 questions fielded · 0 that broke the design

"How do users tell apart AI-driven from traditional search — and are they ever forced to use AI?"

AI and traditional are intentionally adjacent, not blended. Traditional stays keyword-based and filterable; AI is opt-in, conversational, cited. Suggested queries are entry points, not triggers. The user is never forced into a chat to get a basic fact.

06 · What I Took From It

Two things I keep coming back to.

There was always a faster version of every decision: ship the toggle button instead of resolving the toggle problem; tab the AI off to one side and call it done; force users into a chat to get an answer. Each shortcut would have been faster to build and would have rebuilt the exact problem we were trying to solve. The work was mostly in noticing those shortcuts and not taking them. The collapse arrow on the AI panel is the project in miniature: easy to leave out, and everything else only works because it's there.

Our research was diffuse until our sponsors pushed us, twice, to cut the audience down. First from 'people who use websites' to 'people deciding whether to buy Slack.' Then from that to two specific personas: small business owners and enterprise researchers. Every feature decision got easier after each cut. Feature scope followed audience scope for free.