CASE STUDY
An AI relationship CRM for sales teams
A CRM where every health score, next step and summary says how sure it is, why, and how to fix it.
THE PROBLEM
Thread is a relationship CRM concept for B2B account teams. It scores the health of each relationship, suggests the next step, summarises every email and call, and keeps contact profiles up to date. The design problem was how to show an AI's guess so people trust it enough to act on it, and can push back when it is wrong.
THE DECISIONS
Each decision, why I made it, and what it cost.
Without a confidence level, a rep either trusts every prediction or ignores them all.
Each output carries three stitches and a word: confident, fairly confident or unsure. It adds an element to every row, so it stays small unless Thread is unsure.
Thin data still produces a confident looking number, so low confidence suggestions should wait for a person.
Unsure suggestions show in ochre, with Confirm and Not right instead of Accept and Dismiss. It adds a click, on purpose.
A list of field names is not a reason. A plain sentence lets a rep judge a suggestion in seconds.
Each output opens to one sentence of reasoning, the signals used and when it was generated. Every explanation has to be written in plain language, which is more work than listing factors.
If fixing the AI takes longer than ignoring it, people ignore it. If a fix disappears, managers cannot see where the AI is being overruled.
Accept, edit, dismiss and undo sit on the output itself, and an override is marked set by you. Every AI component now needs four states instead of one.
A rep starting the day needs a short ranked list, not a table sorted by score.
Today is a numbered queue ranked by urgency and then health, grouped into today and this week. Less is visible at once, so the full directory stays one click away.
A single score says where a relationship is, not where it is going. Sofia at 31 matters because she was at 58 eight weeks ago.
Each contact opens on an eight week health line with warm and cold thresholds, turning signal red when it cools. A line is less precise than a chart with axes, so the exact score and its change sit right beside it.
THE SCREENS
The manager's view of the team, and the same product on a phone.
THE FLOW
Every health score, next step and summary follows the same path. The branch that matters most is the one where Thread is not sure.
The rating is shown as three stitches and a word on the suggestion itself, so it is read before any action is taken.
Shown in the queue with Accept, Edit and Dismiss.
Held back in ochre with Confirm and Not right, so a person decides before anything happens.
"Why this?" opens a plain sentence of reasoning, the signals it used and when it was generated. It is optional, one click, and never blocks the action.
The suggestion is applied, with Undo still available.
The rep's version replaces it and is marked set by you, so the override stays visible.
It leaves the queue, with Undo still available.
Because overrides stay marked, a manager on Team health can find a cooling account, the rep behind it, and whether the AI's read is being trusted.
HOW I WOULD TEST IT
Thread is a concept and has not been tested with sales teams yet. These are the questions a first round would answer, and what each answer would change.
Five account executives clear a seeded Today queue in the prototype, with one wrong suggestion marked unsure and one marked confident.
If they accept the unsure one as quickly as the confident one, the confidence mark is too quiet and needs more weight on the row.
Ask each rep to explain a health score in their own words after opening it, with no introduction to the feature.
If the explanation is read back as a list of fields rather than a reason, the reasoning sentence is doing too little and the signals too much.
Seed an urgent account at position seven and see whether reps find it before the session ends.
If most miss it, the cut off at five is too aggressive and this week needs to surface higher.
THE DESIGN SYSTEM
Warm paper, dark ink and one signal colour. A serif for people and decisions, a sans for the interface.
Newsreader
Names, headlines, health scores and the AI's reasoning, which is set in italic so it always reads as Thread speaking. Optical sizing keeps it crisp from fourteen to seventy two pixels.
IBM Plex Sans
Interface text, recommendations, buttons and the small caps labels. Built for dense, functional screens, which is what a CRM is.
IBM Plex Mono
Scores, dates, deltas and chart axes, so figures line up in a column and read as data rather than prose.
HOW IT WAS DESIGNED
THE OUTCOME
