A customer comes in because their car shakes, and the inspection finds fifteen other things. Usually they see a long list of upsells. RecLink groups everything under the problem they actually came in with, so the fix makes sense. I designed it and built it as a working prototype, and it’s now in pilot.
A customer books in for one thing, such as the car shaking, the brakes grinding or a smell from the air vents. The technician's multi-point inspection (MPI) then finds everything else, from a torn CV boot to a clogged cabin filter. Each finding is a chance for more work, and the advisor usually has one phone call to explain all of it. Behind that call, three people work from the same inspection, each in their own tab of myKaarma's software.
RecommendationService advisorPrices the work and presents it to the customer
↔
PartsParts counterSources and prices the parts for each recommendation
↔
MediaTechnicianAttaches the photos and video that show the customer why the work is needed
What was wrong
A flat list reads like a sales pitch.
The current product lists every finding as its own row with its own price, in the order the technician typed them. The diagnosis that answered the customer's complaint and the extra work it uncovered appeared as unrelated lines. Every extra finding looked like an extra sale, and advisors had to explain the connection again on every call. Use the buttons to switch between today's tab and RecLink.
Today's tab: every finding is its own row with its own price, in the order it was typed. Nothing says which one answers what the customer came in for.
The whole Recommendation tab in the real prototype. Scroll inside the frame to see it all.FIG. 01
02 · What I decided
Every decision keeps the story together.
Grouping findings by concern is the core idea. These five decisions made it hold up in the advisor's, parts counter's and technician's tabs, and in what the customer receives.
decision/01 · structure
Group work by the customer’s concern.
In pilotInformation architectureDealer interviews
Every recommendation now sits under the concern the customer came in with. "My car is shaking while I drive" becomes a card, and the diagnosis, the road-test finding and the torn CV boot all sit inside it.
The card shows one total, so the advisor can say "the shaking is $953 to $1,048 to fix". That's the upsell, and it doesn't sound like one.
One concern card holds everything that answers it, with one total. The customer's own words are the heading.FIG. 02
decision/02 · regroup
Drag to regroup, reorder to tell the story.
In pilotInteraction designUsability tests
Technicians write findings in the order they find them, which rarely makes sense to a customer. Advisors can now drag findings between concerns and reorder them.
A preview shows the job as the customer will read it, so the advisor sets the story before the customer sees a price.
Pre-approved work stays at the top of its concern, so regrouping never buries what the customer has already said yes to.
Findings sit under the concern they answer, in the order the advisor chooses. Drag a row to move it.FIG. 03
decision/03 · pre-approved
Already approved looks already approved.
In pilotStatus designUsability tests
Work the customer agreed to at check-in looked like any other recommendation, so advisors asked for approval twice. Now it sits at the top of the concern with a green tick.
In testing, it changed how advisors opened the call: "you've already approved the diagnosis; here's what it found."
The Parts tab shows the same pre-approved line in the same green, with the filter counts the parts counter works from.FIG. 04
decision/04 · options
Good, better, best, priced side by side.
In pilotChoice architectureSupport tickets
A finding can carry several repair options, each with its own labour and parts. For a torn CV boot, that might be a new axle from the carmaker (OEM), a remanufactured one, or just a boot kit. The concern shows the range rather than one take-it-or-leave-it price, which makes the bigger job easier to say yes to.
Approve and decline now sit right beside each price, instead of in a pop-up.
Declined work stays visible, so dealers can follow up later.
A torn CV boot with three prices: a new axle from the carmaker, a remanufactured one, or just a boot kit. The customer picks one, and the advisor records it on the same row.FIG. 05
Parts for one recommendation, with internal notes the customer never sees.FIG. 07
03 · Where it is
It’s in pilot. Here’s how we’ll know it works.
There are no results yet. Each release goes out to dealers in randomly chosen waves. We compare each dealer with its own results from the weeks before, and judge the release on weeks 3 to 6 against a success bar agreed in advance.
An illustration of how each release is measured, using made-up data. It shows the % change in approval rate compared with each dealer's own four weeks before the release.
H1 · North star
Context sells
Grouped findings get approved more often. Measured by approval rate and upsell per repair order.
H2
Options lift the yes
Offering repair options gets more work approved. Measured by how many estimates have 2 or more options, and their approval rate.
H3
Sent sooner
Inspection results reach the customer sooner. Measured by the time from finishing the inspection to sending the MPI.
H4 · Guardrail
Nothing gets harder
Measured by support tickets per 100 repair orders and survey score.
Test sessions
13
8 dealerships and 5 internal groups went through the prototype in usability sessions.
Pilot dealers
Asked for it
Several asked when they could have it. The pre-approved band at the top of each concern was the change they mentioned most.
Methods: user interviews, usability tests, feedback from UserVoice (a product feedback tool), analysis of Zendesk support tickets, dealership visits, automated browser checks on every change, and a release dashboard that reads live data from Looker.
04 · How I built it
How I built it with AI agents.
RecLink was too big to build one AI prompt at a time, so I designed the workflow first: where feedback goes, what an agent can change alone, and where a person approves. Several steps became tools the wider team now uses.
Comment
Point at the problem in the prototype
In the prototype, press C, click any element and choose Fix or Add to design system. That creates a GitHub issue (a tracked task) with the component, screen and exact element attached, so an AI agent can pick up the fix without anyone having to describe where the problem is.
Propose
Gaps become design system proposals
When RecLink needed a component the design system didn’t have, it became a proposal with a live preview of the current and proposed versions side by side. Three designers must approve it before any code change is opened, and a person makes the final merge.
Check
Automated checks on every change
Browser checks run on each change, so the agent’s work is tested against the same rules every time instead of being checked by eye.
Report
Bugs arrive with their context
Developers and product managers report bugs from the page itself. Each report is pinned to the element it's about, with a screenshot and browser details attached.
Test
Check a study is fair before users see it
Testify, my usability-testing tool, plans the sessions and checks that two prototypes differ only in the thing being tested.
The rule throughout: AI drafts, and people decide. Three of these became tools with case studies of their own.
I'd write an automated check across the three tabs first. Nearly every bug was right in one tab and wrong in another, and a script comparing them would have caught most.
I'd move onto the design system sooner. One round of testers reacted to the unfamiliar look before the structure, and that cost a session.