Moderate and categorize submissions with AI

Once a form is busy, the bottleneck stops being collection and becomes triage. A few hundred submissions a week is more than a human wants to read top to bottom, and the signal - the bug report, the hot lead, the harassment to escalate - hides in the noise. Three AI features, all on Pro plans and above, do that triage automatically and at receive time.

• Moderation scores each submission for how likely it is to be spam, going beyond structural defences to catch human-written junk.
• Auto-categorization labels each submission by intent - support, sales, bug-report, and so on - so you can filter and route.
• Insights cluster the open-text answers across many submissions into themes, so you can read a summary instead of 800 verbatims.

These build on the spam-filtering layers in the stop-form-spam guide - AI moderation is the top layer there - but go further: where moderation decides is this spam, categorization and insights decide what is this about and what are people telling us. Everything below is surfaced in the dashboard, the API, and over MCP for agent workflows.

On Pro and above, every submission runs through a language-model moderation classifier that reads the submission body and returns a single score between 0.0 and 1.0, rounded to two decimals - higher means more likely spam.

Unlike a honeypot or CAPTCHA, which only catch automated submissions, moderation catches human spam: manually-typed junk with real keystrokes behind it. SEO link drops, copy-pasted bot pitches, harassment, and phishing attempts all score high. The score also carries a category when applicable - harassment, solicitation, phishing, nonsense - which is shown next to it on the submission detail page.

A rough reading of the score range:

The submission detail page shows a coloured pill (green/yellow/red) for the score, and the form overview shows the average score over the last 30 days next to the spam-rate chart so you can watch for drift. The AI moderation doc is the full reference.

The model is not tuned per form - the score is what it is. What you control is the moderation threshold (default 0.75): submissions scoring at or above it are filed in the spam folder, and everything below lands in your inbox. The threshold is the single knob, and it lets you choose how aggressively to act on the score.

Tune it on the form's edit page under AI moderation:

• Lower it (e.g. 0.55) if you would rather over-flag and hand-review - useful for a form where missing a real message is worse than reviewing a few extra.
• Raise it (e.g. 0.90) if you only want the model to catch obvious garbage and prefer to see borderline submissions in your inbox.
• Set it to 1.0 to keep the score visible without acting on it at all - a good way to observe the model before you trust it.

Crucially, flagged submissions are filed, not deleted, so a false positive costs a click to recover rather than a lost customer. If the score consistently misses a category of spam you care about, do not fight the threshold - layer a custom rule on top to catch that specific pattern. The stop-form-spam guide covers how moderation fits alongside the cheaper layers.

Moderation tells you what to ignore; categorization tells you what each real submission is about. On Pro and above, every submission gets an auto-generated category - a single short label - at receive time. It is the difference between an inbox of 500 unread submissions and an inbox you can filter to "only the bug reports from this week."

The model picks from a constrained vocabulary so you can rely on the label. The default set is:

support, sales, partnership, bug-report, feature-request,
feedback, hiring, press, spam, other

If the model is not confident, it picks other rather than guessing. The category surfaces as a coloured pill in the submission list, on the metadata sidebar of the detail page (next to the moderation score), and as a filter dropdown in the inbox. It is also part of the API submission resource and can be filtered on the list endpoint, so a downstream system can route on it. The auto-categorization doc has the API shape and query examples.

The default vocabulary fits a general contact form, but a specialised form - a job application, a beta signup, a support escalation queue - usually wants its own labels. You can override the category list per form on the form settings page under Categories.

Custom categories are validated: between 1 and 20 per form, lowercase letters/numbers/dashes only, and a maximum of 32 characters each. The model is re-prompted with your list on every submission, so new submissions use the new vocabulary immediately. Existing submissions keep their original labels - you do not retroactively relabel them just by changing the list.

When you do want to relabel history - because you changed the vocabulary, imported submissions from elsewhere, or a new model version released - the MCP server includes a categorize-submissions prompt that takes a form ID and a date range and re-runs categorization across the range. It confirms the count, shows a sample of what would change, and only applies once you approve; re-running on already-categorized submissions overwrites with the latest pick. The auto-categorization doc covers the constraints and the batch flow.

Moderation and categorization act on submissions one at a time. The third feature works across them. When you have hundreds of open-text answers - "what is the main reason for your score?", a feedback box, a survey response - reading them all is where manual analysis traditionally dies.

AI insights cluster open-text responses into themes, surface representative quotes per theme, flag sentiment shifts against previous periods, and produce a summary you can paste straight into a decision document. On paid plans, the weekly digest does this continuously for always-on forms and surveys, so an emerging complaint theme reaches you in days rather than at quarter-end. Top categories with counts is one of the digest's sections, tying insights back to the categorization above.

Treat the output as a first pass, not a verdict: spot-check the themes against the raw quotes before you present them. The model accelerates the reading; the judgement about what matters stays yours. The post on AI insights for form responses at scale walks through a real analysis.

All three features are gated together on the ai_features plan flag:

• Free - the moderation classifier does not run (the score is absent), category is always empty, and insights are off.
• Pro and above - every submission is scored and categorized at receive time, and insights and the weekly digest are available.

There is no per-submission charge - moderation and categorization run on every submission as part of the plan. Downgrading to Free leaves existing scores and categories on past submissions but stops processing new ones.

The AI output is available everywhere you work with submissions: the UI (coloured pills, filters, the insights page), the API (the moderation score and category are on every submission resource, and you can filter the list endpoint by category), and MCP (the list-submissions and get-submission tools return the score and category, and the categorize-submissions prompt re-runs labelling). That makes the same triage available to a human in the dashboard and to an agent over MCP.

Ready to sort your inbox? Start free to build the form, then upgrade to Pro to turn on moderation, categorization, and insights. The send-submissions-to-your-CRM guide shows how to route the now-categorized submissions onward.