Scout by Factors.ai · M Munawar Ahmed

Rebuild: Pitched to our CTO, built on an open-source chat library
April 2026: Launch

A campaign performance dashboard Scout generated as a JSX page, with spend, ROAS and a campaign table — A campaign dashboard Scout wrote and rendered from one request.

The problem

Scout is Factors.ai’s AI assistant for account intelligence. It began as a hackathon prototype on a rapid prototyping framework. That was the right call for a demo, but every new feature meant another iframe, and getting iframes to talk to each other was slow and fragile. Features shipped late and broke often. It was not ready for customers.

The pitch

I pitched our CTO on replacing the prototype with our own React chat app, built on an open-source chat library.

The tradeoff: we gave up the framework’s speed for quick experiments. In return we got full control of the interface, and we didn’t spend months rebuilding message lists, input handling and streaming from scratch. Our time went into the parts people actually see.

Deep dive 1: Running AI-written code safely

Problem Scout can answer with a full HTML or JSX page, like a dashboard or a deal timeline. People want to see that page working inside the chat. But the code is written by a model, so the app can’t trust it.

Decision Every generated page renders in a sandbox, isolated from the app and the user’s session. Preview and Code tabs let people check what was produced. Outputs can span multiple files and carry versions.

Result People get working, interactive pages inside the chat, and a bad output can’t break or reach into Scout.

A deal journey timeline rendered live from AI-written HTML, with Preview and Code tabs — AI-written HTML, rendered live in a sandbox. Preview and Code sit side by side.

Deep dive 2: One answer, many views

Problem The right chart depends on the question, and the model doesn’t always pick it. Asking again just to see a table wastes time.

Decision Each data answer keeps its data separate from how it’s drawn. People switch between chart and table, and between bar, column and line, without a new request. References on the answer open the same data in either view.

Result People pick the chart they want in one click. The numbers match across every view.

Scout answering with a line chart, with Chart and Table tabs and a chart type menu open — One answer, three chart types and a table. No new request needed.

Deep dive 3: Showing the work

Problem People won’t act on numbers they can’t check, and a long silent wait feels broken.

Decision Scout shows its reasoning steps as they happen, collapsed once the answer arrives. It asks a clarifying question when a request is ambiguous instead of guessing. Every answer lists the reports it used.

Result People can trace each answer back to the reports it used. Long requests show progress as Scout works.

Scout's reasoning steps for a dashboard request, including a tool call — Reasoning steps, including tool calls, show up as Scout works.

Two source references on an answer, each with chart and table shortcuts — Every answer lists the reports it used.

Deep dive 4: Streams that survive a dropped connection

Problem Long answers can take a minute or more. A closed laptop or flaky Wi-Fi used to mean losing the answer and starting over.

Decision Streaming is resumable. When the connection comes back, the answer picks up where it left off instead of restarting.

Result Long answers are no longer all or nothing.

Design system and tooling

I own Scout’s design system and led the move from antd to shadcn, so components match our design instead of fighting a library’s defaults. I moved the build to Vite, Storybook 10 and Tailwind 4, and set up hot reload that saves about 2.5 to 5 developer weeks a year. I also built MCP server tooling so AI coding assistants work with our components and conventions, not generic guesses.

Outcome

Scout went from an internal demo to a launched product in April 2026. Answers come with reasoning and references, people can share a chat by link or export it as a PDF, and a prompt library helps new users start.

A share dialog with Private and Team Access options and a Copy link button — Share a chat privately or with your team.

A prompt library with Ask, Routines and Agents tabs and prompt cards — The prompt library gives new users a place to start.

Looking back

When a tool starts slowing everyone down, say it out loud. The prototype was great for demos, but every new feature fought it. Pitching the rebuild with honest tradeoffs, instead of working around it, is what got Scout to customers.