What I got wrong building an AI chat

2023–2026 · UX, design and frontend

Before every AI product looked the same, I spent three years designing and building Sourcetable's AI chat, trying to make it more than a chat. I got a lot wrong.

Figma, React, Tailwind, shadcn/ui, Parcel

TL;DR I made the original design of Sourcetable’s AI Assistant at a hackathon in March 2023, and have kept designing and building the chat since, through 450+ changes. It became our most used feature, with 63k+ first-chat interactions a year. Getting people to use it well took a lot of tries: a box that covered the data it was talking about, a first message that sent itself, @-mentions to point it at the right data, and an edit menu nobody opened.


It was a normal startup week in March 2023. Everyone was starting to talk about AI, and it still felt like a novelty. Then came a hackathon that weekend, and the brief was to bring AI into the product.

Damn, how? The design board filled up fast, from autocomplete to a tool that pulled tables out of documents.

In the end we picked the chatbot. It was the simplest thing to ship as a v1, and it had the most room to grow.

A hand-drawn sketch of the first AI Assistant: a chat window titled AI Assistant floating over an empty spreadsheet, opened from a round sparkle button in the bottom right. Inside, two questions in gray bubbles on the right, each answered on the left after a sparkle
Spring 2023. The assistant floats over the grid, on top of the header row it is talking about.
A hand-drawn sketch of a spreadsheet with Jan and Feb typed in a Month column and selected, a dark Autocomplete button with a sparkle at the corner of the selection, and Mar, Apr and May suggested in gray under them
Autocomplete: type a few cells and the AI offers the rest.
A hand-drawn sketch of an invoice dropped on a spreadsheet, with a logo, item lines with prices and a total. It shines, and an arrow with a sparkle points from it to a table with Item and Price columns
A table extractor: drop in a document, get a table back.
A hand-drawn sketch of a spreadsheet with a Review column and a Mood column. The selected Mood cell holds =AI(A1) with a sparkle, and the cells under it read angry, meh and happy
AI in the cell: a formula that asks the AI, one row at a time.
A hand-drawn sketch of a dashed selection on a spreadsheet, with a dark bar that has a sparkle and asks 'Is this cleanup correct', and redo, no and yes buttons beside it
Cleanup in the grid: the AI tidies a range, then asks if it got it right.

Taking the box apart

Within a week it was in the product, a box floating over the grid. I designed it, and a lot of that week’s UI commits were mine: the loading message, the chip for the sheet you had selected, tables inside the answers, a copy button. Floating was the idea. I wanted an assistant that lived inside the spreadsheet, one you could make as wide as you needed.

But floating over the grid meant covering it. The box sat on top of the very data you were asking it about. By the end of May I docked it in a fixed sidebar on the right, for now. The CEO and I agreed to disagree there. I always wanted people to pick for themselves, floating or docked. And the sidebar didn’t really solve it. The chat stopped covering your data and started taking room from it instead. “For now” stuck. I never got the time to go back to it, because at a startup there’s always something bigger, so it stayed a sidebar, just a better designed one.

Another problem: it was too simple. There was so much more to add than a chat box, and it kept growing and growing. It learned to stream answers, take payment through the card modal when credits ran out, and listen through a mic.

A hand-drawn sketch of the v1 chat window up close, with red notes: 'Clippy? too much?' on the smiley avatar, 'feedback' on the thumbs, 'What data?' on the Selected chip, 'still working?' on the typing dots and 'stop it midway' on Stop Execution
March 2023. V1, piece by piece.

It ended up being the most important and most used feature we had, but we wanted more. We wanted charts. We wanted it to take control of the spreadsheet and automate the work, not just talk about it. And connectors were our edge: few products could pull in your data at all, so we wanted people asking questions across everything they’d connected.

Chatting with your data was still a new idea back then, and people didn’t get as much out of it as we hoped. So we started building new ways in.

Four ways in

Getting people in took five tries. The early versions taught first: a product tour, then a welcome survey that went slide after slide, asking your role, your company’s size and the tools you used before you saw a single cell.

The latest version, which I shipped in July 2026, skipped all of it. You landed straight in a live workbook, blurred behind one window, with four ways in. Connectors came first, because they were our edge. If you had nothing to connect, you could drop in a file, start from a sample dataset, or just ask.

A new Sourcetable account: the workbook blurred behind a window asking 'What do you want to analyze today?', with a prompt box and a file button, connector cards for Google Analytics 4, Google Sheets, Google Drive, Microsoft 365 and PostgreSQL, and five sample datasets
October 2026, a new account. The workbook waits behind one window with four ways in.

Whichever way you came in, the first message sent itself. Pick a sample and the data loaded at once, confetti fell, and the chat was already on it: “Explore this workbook: summarize it and chart the key trend.” Then it answered with tables pulled from the sheet, a summary and a chart of the monthly trend. From Sign up to an analyzed workbook took less than 30 seconds.

The sample workbook with the AI Assistant sidebar full of its first answer: expense breakdowns by department, then a Monthly Trend Analysis with bullets on 2024 performance, an early 2025 surge, a mid-2025 revenue freeze and a September 2025 recovery
Part of the first answer, the trend read straight off the sample.

Point at the right data

Getting people to connect their data was half of it. Once it was in, the chat had to know where to look. On the v1, a chip under your message answered “What data?” with the cells you had selected. A connector could hold a whole database. All the chat had for that was a Connectors button that dropped a name into your message as plain text.

In May 2026, @ became the way in. Type @ and every connector opens as a tree, down to its tables and columns, next to your files and tools. Nothing connected yet? The menu offered to add a connector right there.

A chat box reading 'Query my @click' above an autocomplete that lists a ClickHouse connector, its sales and visits tables, and the order_total and order_date columns
June 2026, from the ClickHouse connector announcement, with sample data. Type @ and the connector opens down to its columns.

The first version was already in when I took over its UX, and it was rough. Typing didn’t feel like it does in the products people already use, and backspacing felt odd. Complaints piled up in our support inbox. Honestly, that was the good part. It meant people were actually using it.

So I redesigned and rebuilt it (tests included). I changed how the arrow keys and clicks moved through the menu and what @ actually did, and added more suggestions. You could collapse a connector even when it matched, a mention could hold one space, and picking a connector stopped swallowing the next word you typed. This one stuck, and it put connectors right where people were already typing.

An edit menu nobody used

AI Edits let you edit spreadsheet content with AI without opening the chat. It was an amazing feature, and nobody used it.

We tried it in four places, and I designed all of them: a dot inside the cell, the menu bar, a cell placeholder and the one I put the most into, a separate glowy button. Nobody complained, which by now you know is a bad sign. On demos and user calls, nobody clicked it, not even the glowing button, as obnoxious as it was. Sheets and Excel have the same problem, and they’ve tried some of the same spots.

The header menu was gone about seven weeks after it shipped, and we stopped iterating once it became part of the cell placeholder. It should never have left the chat. There was a gap there, and I’m still not sure what should have filled it.

Stop it midway

People change their mind mid-answer, so the chat had to stop when told. Escape stopped an answer while it streamed, and so did a cancel button in the composer. Messages the app sent for you, like a new user’s first one, waited in a queue that drained only after the current answer finished, so nothing got dropped and nothing went out twice. A slow image upload could still fire a send twice, a race that only showed up on hotel wifi, and I fixed it.

If I started over

I’d start from the data, not the chat. Answers, charts and insights would be things you keep, ready to export or edit.

It would always show what it’s doing, and you could steer it midway instead of only stopping it. When it edits the spreadsheet, you’d know, every time.

You’d pick where it lives, floating or docked.

And in the code, the streaming and queue logic would move out of the component into one hook, with the cancel and upload-race paths tested first.