Introducing Annotation Forms: Capture any human feedback without leaving Confident AI

What are Annotation Forms?

An annotation form is the set of questions a reviewer answers when they open an item in an annotation queue. You design it once in Project Settings → Annotation, and from then on every reviewer working that queue sees the same fields, in the same order, capturing the same data.

Each form is a list of fields, and each field has three parts:

• Field — the question the reviewer is asked, e.g. "Describe how good the AI response is."
• Type — how the answer is captured (more on the types below).
• Description — optional guidance shown to the reviewer, e.g. "Focus on characteristics such as response length, and conciseness."

Any field can be marked Required, so reviewers can't submit until the questions that matter are filled in.

app.confident-ai.comBuild an annotation form field by field in Project Settings on Confident AI.

The problem: review always ended up in a spreadsheet

You could already annotate on Confident AI — leave a thumbs up or down, give a 5-star rating, write an expected output. That covers the common cases, but it's a fixed set of fields. The moment a team needed to capture anything beyond them, they ran into a wall.

So what we kept seeing was this: the SME or PM does the part the product supports, then exports the data and does the rest in an Excel sheet. The severity rating, the failure category, the three checkboxes for what went wrong, the rubric score their team actually cares about — all of that lives in a spreadsheet next to the export, disconnected from the traces it describes.

That's the worst of both worlds:

1. The real review lives outside the platform. The judgments that matter most are in a file on someone's laptop, not attached to the traces, datasets, or evals they're about.
2. Nothing flows back. A score in a spreadsheet can't become a golden expected output, alignment data for an LLM judge, or a filterable signal in your dashboards. Someone has to re-import and re-code it by hand — so usually no one does.
3. No consistency across reviewers. Every SME builds their own columns, their own scale, their own definition of "good." There's no shared schema, so the results can't be compared or aggregated.

Annotation Forms fix this at the source: define the exact fields your reviewers need — once, in the product — so the entire review happens on the platform and the spreadsheet detour disappears.

Field types

When you add a field, you pick how its answer is captured. Forms support seven types, so you can match the question to the right input instead of forcing everything into text:

• Text — a free-form text answer. Best for open-ended explanations and qualitative notes.
• Number — a whole number answer.
• Decimal — a decimal number answer, for fractional scores and measurements.
• Yes / No — a simple yes/no toggle for binary judgments like "Did the AI answer the question?"
• Single choice — pick exactly one option from a list, ideal for categories like severity or sentiment.
• Multiple choice — pick one or more options, for tagging every issue that applies to a response.
• Criteria — score a single annotation criterion, capturing a structured rating against a definition you set.

Mix and match these across a single form — a Yes/No for "is it correct," a single choice for severity, a Criteria score for helpfulness, and a Text field for notes — to capture exactly the dimensions you care about.

Building a form

Forms are built in a drag-and-drop editor:

1. Open Project Settings → Annotation and create a New Form
2. Click Add to append a field
3. Write the question, pick a Type, and optionally add a Description to guide reviewers
4. Toggle Required on the fields that must be answered
5. Reorder fields by dragging the handle on the left of each field
6. Save

Switch to the Preview tab at any time to see the form exactly as a reviewer will — so you can sanity-check the wording and ordering before it goes live.

From annotations to evals

The point of structured annotations isn't just tidy data — it's that structured data flows directly back into the rest of Confident AI. Because every answer has a known type:

• Scores and criteria become quantitative signals you can filter, chart, and threshold on, just like automated metrics.
• Choice and yes/no fields become labels you can slice your traces by in the Observatory and Dashboards.
• Reviewed items become aligned data for your LLM judges and golden expected outputs for your datasets.

Human judgment stops being a comment nobody reads and becomes a first-class input to your evaluation workflow.

Get started

Annotation Forms are live on Confident AI now.

• Read the annotation documentation for the full reference on every field type and option
• Open Project Settings → Annotation in your project to build your first form

And keep an eye on the blog this week — we've got more shipping.