← All epics

What data annotation actually needs from its tools

I run an annotation company, so I have opinions about annotation tooling that are shaped by watching real people use it for eight hours a day.

The buyer's checklist is not the annotator's checklist

Buyers ask about model coverage, export formats, audit trails and integrations. All of that matters. None of it is what determines whether a project finishes on time.

Annotators care about:

  1. Throughput — how many items per hour, and whether the tool fights them.
  2. Fatigue — whether the interface makes mistakes more likely by hour six.
  3. Clarity — whether the guideline is visible without leaving the screen.
  4. Recovery — whether a misclick costs ten seconds or ten minutes.

The details that pay for themselves

  • Keyboard-first navigation, with shortcuts visible and remappable.
  • Guideline reference pinned beside the canvas, not in a separate tab.
  • Instant undo that also works after a save.
  • Progress that reflects real remaining work, not just completed items.

These are not glamorous. They are the difference between a project that hits deadline and one that quietly slips while everyone reports green.

Training is the product

The strongest annotation operation is not the one with the best software. It is the one where newcomers become productive fastest. That is why our learning centre gets as much attention as the tooling — the bottleneck is rarely the interface, it is the onboarding.

If you are building annotation tooling, spend a week sitting next to annotators before you write a line of code. Then delete half the features you planned.