Ford Quality Agent
I designed an internal agent for Ford quality engineers. It pulls multiple data sources into one answer, shaped by the question asked. It derives parts, root cause, and next steps from vehicle warranty data, and it contributed to ~$55M in savings.
Engineers now catch a recurring issue before the vehicle leaves the plant. The data sits in one place, so cross-functional teams pull what they need instead of booking a meeting to ask for it.
A Ford quality engineer spent the day finding data, not deciding with it. Claims, vehicles, suppliers, and plant repairs each lived in a different system.
The expensive decisions waited on that data: which vehicle line to focus on, which supplier to charge back, which vehicle function to hand to which plant. Each one meant someone assembled the numbers by hand, then held a meeting to pass them on.
I started by interviewing quality engineers about where their time went and what got in the way. What they told me shaped the product. The Quality Agent puts all of that data behind one conversation. An engineer asks the way they would ask a colleague, and the answer comes back as a sentence, a chart, or a table.
I built it as a working HTML prototype, tested it with engineers, and revised it with them before handing it to the build team. Engineers were going to act on what the tool told them, so I made every state show its work.
The prototype was not a throwaway. Ford already had a design system, so I read its guidelines straight out of Figma with Figma's MCP server instead of rebuilding them by eye from a screenshot. From those I made a small set of reusable HTML and CSS components: type, colour, spacing, buttons, chips, tables, and the chart frame. Every screen in this study is assembled from that set. Reviews stayed on the product because the styling was never in question, and the build team inherited components rather than a picture of them.
Launch was not the end of the work. The agent is live, and it contributes heavily to Ford's quality goals and to the warranty and recall savings behind them. I watch how engineers use it in Contentsquare, the product analytics tool. What I see there is what I change next.
All data shown is sample data from the visual prototype.
The empty state is a menu of real questions, grouped the way the work is grouped. Nobody has to learn how to prompt.
Packaged reports are slash commands. The agent asks for the one thing it cannot assume, then generates the rest.
The agent names its stage in the thread and in the header. Nothing happens silently.
A report is a document, not a chat reply. Excel and screenshot export sit under every table.
A trend question gets a chart with the reading written above it. A single-fact lookup gets one sentence.
