
Workiz Copilot
AI-native workflows for field service teams.
- Company
- Workiz
- Timeline
- Oct 2025 – ongoing
- My role
- Senior & Lead Product Designer
Overview
Workiz is a powerful CRM that grew complicated: flows with too many steps, settings that are hard to find. Redesigning every flow would take years. AI gave us a cheaper path. Owners ask, Copilot does the clicking, and each capability fixes a usability problem without rebuilding the feature underneath.
I designed the original Copilot: the panel, its entry points and the AI patterns inside it. Designers in Communications and Marketing built their own AI features on it. I used it across the core domain for two jobs: getting information fast and getting daily work done.
How one designer covered it
- Keep up with weekly AI releasesDesigned in parallel with development and reviewed each release with the PM
- Reuse what already workedShared AI components (panel, chips, feedback, in-field generation) that other teams reused
- Polish in the real productFront-end prototypes in the Workiz app
- Synthesize research fastResearching and prototyping with Claude
- Miles Lewis128$15,600
- Jessica Ray141$14,250
- Thomas Kim163$12,980
- Alice Smith122$9,840
- David Brown109$7,310
Open report in Workiz
The challenge
Owners don’t want to learn a CRM. They want answers and finished work.
So the question was: can AI do the hard part of the work for owners, so the product feels simple without rebuilding it?
| Where we were | Where we had to get |
|---|---|
| Answers behind report builders and filters | Ask a question, get the answer and the report |
| Client history spread across pages | A summary when you need it |
| Estimates and messages written by hand | Professional text and images in one click, reviewed by the owner |
| AI features each with their own UI | One AI pattern any team can build on |
Goals
Business goals
From the Copilot launch brief: drive adoption and activation of Custom Reports, drive upgrades to Pro and Ultimate, and build trust and enthusiasm around Workiz AI.
User goals
We designed for owners and office managers, who needed to:
- Get information fast: an answer about the business or a client without hunting through pages.
- Support my workflows: finish estimates, invoices and messages faster and better.
Design objectives
- Bring AI to where the work happens, not only into a chat.
- Be clear about what the AI can do.
- Keep the owner in control: review, edit, approve.
- One AI pattern that other teams can build on.
Timeline: designing in parallel with the AI team
The AI team shipped experiments weekly, so I designed in parallel: they built a capability while I designed how it looks and behaves. The launch notes for Copilot and for AI automation messages both credit my design.
| Phase | Focus | When |
|---|---|---|
| Phase 1 | Build the panel: panel, entry points, Client Insight | Oct 2025 – Jan 2026 |
| Phase 2 | Learn from beta: more suggestions, honest copy | Mar 2026 |
| Phase 3 | Ship what worked: suggestions, reports from chat | Aug – Oct 2026 |
| Phase 4 | Show the work: trust patterns, in progress | Sep 2026 – now |

Sources: HubSpot Breeze, monday sidekick, Atlassian Rovo, Intercom Copilot and Jobber Copilot; AI design systems (Rovo UI, Cloudscape, Carbon for AI, Microsoft HAX, Google PAIR, Fluent 2); user interviews and the AI team’s customer calls; usage data (Amplitude) and Copilot thread logs; feature requests (Loop).
What owners told us
- Getting informationReports are hard to buildFilters and columns too advanced for many owners. They want to ask, not build: “Give me how many jobs we made during the last month, how was performance.”
- Getting informationCan’t trust the numberOwners recount AI answers by hand: “I go to the timeline, select the filters and I count manually.”
- Getting informationNumbers without meaningThe dashboard shows what changed, not why or what to do next.
- Daily workUnprofessional documentsShorthand and typos (“standard installation kit and labor”) reach the client.
- Daily workNo imagesMany items have no image, and finding one takes time owners don’t have.
- Daily workGeneric messagesAutomation messages are written once and stay generic.
Part 1. Get information fast
Client insight shipped with the panel in Phase 1. Reports from a question launched in October 2026. AI insights on the dashboard are in beta.
Reports from a question
Solves: reports are hard to build · can’t trust the number
An owner asks “revenue by zip code this year.” Copilot answers with a short insight and a report card that opens the full custom report. The number sits next to its source, so it can be checked in one click.
Data: 32% of accounts that got a Copilot report opened the full report within a week. Reports opened from Copilot about doubled in October, when report generation launched.
Client insight
Solves: no context on a call
When a known client calls, a pop-up summarizes recent activity, the active job, financials and a next step. It never covers the dialer.
Data: The widest reach: 52% of accounts using Copilot. 47% came back to it in a later month, the highest return of any capability.
AI insights on the dashboard (beta)
Solves: numbers without meaning
A sparkle on each widget opens Copilot with what changed, three key observations and actions that link to the page that fixes it. The dashboard becomes a starting point for action.
Part 2. Support their workflows
Polish, item images and automation messages are part of the Copilot beta, behind a feature flag.
Polish estimates and invoices, and add item images
Solves: unprofessional documents · no images
One click rewrites item titles and descriptions into clear, client-friendly text. The owner reviews and applies each one, so nothing reaches a client unchecked.
Items without a picture get a realistic generated image in about 20 seconds, on estimates, invoices or the price book. The owner picks it before it’s attached.
Data: polish is the most-used capability, 24% of all threads. 58% of accounts that polished approved the edits; about 1 in 3 came back in a later month. 37% of accounts using Copilot generated images; about 1 in 3 came back. The in-document button is used by about 4x as many accounts as the same action in chat.
Automation messages
Solves: generic messages
When an owner sets up an automated text or email, AI drafts or refines it for that step, with subject and tone set for the trigger.
Data: 14% of accounts using Copilot; 1 in 5 came back, which fits how rarely owners set up automations.
What feedback changed
The team planned to route every AI action through the chat. I argued short text should be edited where it lives, with an in-field AI component (generate, revise, rate). Usage agrees: the in-document image button gets about 4x the chat.
Part 3. The UI decisions behind AI
Every capability sits on a small set of AI patterns I designed once. Communications and Marketing designers built their AI features on them. I checked them against five assistants (HubSpot Breeze, Atlassian Rovo, monday sidekick, Intercom Fin, GitHub Copilot) and AI design systems (Rovo, Cloudscape, Carbon, Microsoft HAX).
One pattern, reused everywhere
- Panel beside the work. Docked right, pushes the page instead of covering it.
- Entry points where the work is. “Ask Genius” in the top bar, plus buttons on client, job and report pages.
- Suggestions to start. Welcome chips, because an empty box doesn’t say what to ask.
- Chips suggest, buttons act. Anything that changes data is a button.

| Decision | Why |
|---|---|
| Suggestions as a start | An empty box doesn’t say what to ask |
| References | Owners recount numbers by hand. Data answers link to the report they came from; 32% of accounts that got a report opened it within a week |
| System status | Thinking states and an activity log of what Copilot did |
| Review before it lands | Owners pick polish edits and images; “Do it for me” confirms first. 58% approved polish edits; polish threads run 5 messages |
| Feedback | Thumbs up/down on every answer, in chat and in the field |
| Easy correction | Regenerate, edit or rewrite |
| Familiar output | Answers become a custom report, a CSV, estimate items |
One morning, end to end
Before his first job, an HVAC owner needs last month’s numbers and two estimates out. The same morning, two ways:
Before: Build a report with filters, recount the number, type every item description, search for product photos, write the follow-up text.
With Copilot:
- Ask. One question returns the answer and a report card.
- Check. One click opens the report behind it.
- Polish. Shorthand becomes client-ready text he approves.
- Illustrate. Images for items without pictures, picked by him.
- Send. The AI-written follow-up text is already set up.
He gets through it without opening a report builder or writing a description.
Results
Owners don’t just try Copilot. They come back to it.
Results
- Owner, appliance repair, seeing an estimate polished and illustrated“Oh my god, that’s pretty crazy. That’s pretty good.”
- The most-clicked prompts are profit and revenue questions: owners use Copilot to understand their business.
- 3 in 4Accounts came back the next month (Aug → Sep 76%, Sep → Oct 75%)
- 4xMore accounts click a suggested prompt after my welcome suggestions (6% → 26%)
- 47%Of accounts that used client insight came back in a later month
- 58%Of accounts that polished a document approved the edits
Accounts, not events. Amplitude and Copilot thread logs, 2026. Shares are of accounts using Copilot.
What’s next
Next milestone: trust. About half of help chats end without an answer. The missing patterns above are the plan.
- Polish core flows in front-end code: polish, images and the report card.
- Show the work on every data answer.
- Control matched to risk: AI-edited markers, revert and approvals.
- Operational Assistant: today’s to-dos when Copilot opens.
- Show AI working inside the text itself, like automation messages, not only in the Copilot panel.
- Mobile: Copilot is web-only today.
What I learned
- AI can fix usability cheaper than a redesign. One-click polish made documents professional without rebuilding the estimate editor.
- The best AI is tied to a moment. Features that show up on a call or an estimate beat open-ended chat.
- Design in parallel, then make it a system. Fast releases worked because each landed on shared patterns other teams could build on.

