A prototype and an argument
A model categorises 35 messy spend lines in one call. Working out which of those decisions to trust is the design problem.
Why this exists
Ignite’s writing on data quality says that AI handles classification at scale, with human review applied throughout to keep accuracy high.
That sentence contains a design problem. A reviewer looking at a screen of machine decisions has two ways to fail. Open every one, and the review costs more than doing the work by hand. Accept everything, and the review is theatre. Neither failure is about the model. Both are about the interface.
So I built the review layer, and wrote down what I think it owes you.
Honest framing
Ignite already ships a spend classification agent. I have never seen its interface. There are no public screenshots, no documentation, and no demo video.
This is not a redesign and it is not a proposal. I do not know what their screen looks like, and I am not going to write as though I do. This is the same problem worked independently, from what they have said publicly, so that I have a point of view about it rather than an opinion about a product I have never used.
Saying that outright is better than hoping nobody notices. Where the work has a boundary, the boundary is the interesting part.
What the prototype does
35 spend lines carry the defects a real export carries. Inconsistent vendor casing, near-duplicate supplier names, Norwegian and English in the same column, invoice references standing in for descriptions, and lines that bundle two kinds of spend together.
One call to google/gemini-2.5-flash through the Vercel AI Gateway returned a category, a confidence, and a one-line reason for each line.
The prototype replays that recorded run, so it opens instantly and costs nothing to look at. It says so on screen, and there is a control to trigger a fresh call if you want to watch one happen. A live call takes about 27 seconds and 1.4 cents, and buys you nothing the recording does not already show.
Some of the answers are wrong. That is the point, and none of them are wrong on purpose. I did not author a single incorrect answer, because a review interface designed against output you wrote yourself is designed against nothing. The model is small and fast, chosen because this endpoint is public and a run has to cost a fraction of a cent. Worth knowing when you look at the mistakes, so you read them as a cheap model’s and not a frontier model’s.
The lines and the prompt both live on the server, so the request body is empty. Nothing a visitor sends reaches the model, which removes prompt injection and unbounded token cost from a public, unauthenticated endpoint in one move.
That is a decision about where a boundary sits, not plumbing.
Six principles
High confidence collapses. Low confidence surfaces. A table that renders all 35 rows identically has decided that every machine decision deserves the same scrutiny, which is both untrue and expensive. Directing attention is the interface's job.
This holds only if confidence tracks correctness. A model that returns 0.95 for everything leaves nothing to triage. A model whose confidence is uncorrelated with being right directs attention at random, which is worse than not directing it at all. So it is something to measure before designing on top of it, not something to assume.
Measuring it changed the design. The first model returned maximum confidence on 30 of 35 lines, which is a small model being asked for a number it cannot produce rather than a model that is sure. Confidence here is a three-value band instead of a percentage, because that is a judgement a model can actually make, and because collapsing high and surfacing low never needed more precision than that. A decimal would only have invited thresholds drawn over accuracy that was never there.
When checking a suggestion costs more than fixing it, people rubber-stamp. That is a rational response to a bad ratio rather than carelessness, and no amount of interface copy talks anyone out of it.
So the correction control sits in the row. No drill-down, no modal, no second screen. A modal here would add cost to exactly the action you most want people to take.
An AI-assigned category stays visibly AI-assigned after a human accepts it. Accepting is not authoring.
In six months someone will ask why a supplier is categorised the way it is. “a person reviewed this” and “a person decided this” are different answers to that question, and an interface that lets acceptance erase the difference has destroyed information that mattered. Ignite say publicly that every suggestion is visible and every change is logged. Provenance that survives acceptance is what makes that true in the interface rather than only in the database.
Print a justification under all 35 rows and within a minute nobody is reading any of them, including the three that matter. Always-on explanation becomes wallpaper.
The one-line reason is available on every row and appears when it is asked for. Keeping it out of sight is what keeps it worth reading.
Accepting all high-confidence rows at once is where the time saving lives, and any honest version of this interface has that button. It is defensible only with an undo that is visible without hunting and cheap to reach.
A confirmation dialog is not a substitute. People click through those without reading, particularly the fifth time. An undo that still works several actions later respects how the tool actually gets used.
A term the model misreads once, it misreads every time it appears. Fixing the same mistake twelve times is data entry, not review.
So the interface lets you correct the pattern once, shows which rows that will change before it changes them, and leaves it visible afterwards that a pattern was corrected rather than a row. At 35 lines you could live without this. At 35,000 it is the entire job, which is why it is the principle I would cut other features to keep.
One of the six lines it catches came back high confidence and is collapsed out of sight. Correcting the pattern reaches a row that triage by confidence alone would have passed straight over, which is the case for this principle in a single interaction: confidence tells you where to look, and it is not the same question as where the model went wrong.
How this was built
Cursor and Claude Code, in a day. Next.js on Vercel, one route handler, one model call.
The repository is public, including the seeded data and the prompt. The parts of this that would otherwise be claims are checkable.
flyary.app is a bigger one, shipped end to end the same way.
It carries no explanatory labels. If a control needs a caption to make sense, it is the wrong control.