Lock the data, open everything else

A chart tool that cannot be trusted with the data is not a chart tool.

The job that started it

Earlier this year, one of our collective — Ellen Drewes — took on a job most designers would quietly dread: a major intergovernmental client needed more than a hundred data visualizations for a single report cycle, on a timeline that made the usual workflow impossible. The usual workflow, in this case, was designers rebuilding every chart by hand in Illustrator from a slurry of Excel files, Word docs, PowerPoint decks, and numbers pasted into emails.

The obvious move was to reach for AI. The obvious problem was that off-the-shelf models happily hallucinated values and changed data silently, with no flag, no warning, no audit trail. Every revision round quietly compounded the risk. For a report that would be read by policymakers, that wasn't a productivity issue. It was a credibility issue.

So she built GraphWitch instead. The working principle was one line long: lock the data, open everything else.

Lock the data, open everything else

That single constraint reorganized the whole product. The AI agent was wrapped in strict guardrails — chart type, format, palette, typography, layout, all fair game. The numbers, never. Users upload almost any filetype, images included, and the tool extracts the data and constructs the visualization. The extracted values sit in an editable panel next to the chart, so a client correction lands at the source instead of being retyped by a designer at midnight.

The rest of the product followed from that same discipline. A style guide builder generates an editable brand system from a single image — a screenshot, a photo, a client's deck — with typography and color variables you can push around on the fly. Every chart type the team asked for lives one dropdown away. Export presets are tuned to the formats reports, decks, and socials actually ship in, not the generic sizes a default template ships with.

The result: a manual workflow that used to take three weeks finished in three days, revisions included. The modest goal going in — cut manual work by half — got quietly obliterated.

Why the constraint is the product

The part we keep coming back to isn't the speed. It's the shape of the constraint. AI in production doesn't fail because the models are weak. It fails because teams treat 'the model can do this' as permission to let it do everything. The teams that ship reliable AI tools are the ones who can articulate, in one sentence, what the model is not allowed to touch.

Lock the data, open everything else. Lock the diagnosis, open the treatment. Lock the facts, open the interpretation. The specifics change; the pattern doesn't. Guardrails aren't a limitation on the product — they're the reason anyone trusts it enough to use it twice.

After the client project shipped, Ellen published GraphWitch to the open web with no marketing. Real users started showing up from more than a dozen countries, running long sessions, coming back. That part is a separate story. The lesson we took from the build is the one above.

If you're putting AI into a workflow where the output has to be trusted — reports, diagnostics, financial analysis, anything a stranger will read and act on — start by naming the thing the model is not allowed to invent. Design the product around protecting it. Then let the model loose on everything else.

That's when the tool starts doing the work.

Get in touch: thecollective@memeador.com