AI: assistant, inline help and agents

Three different things share one rule: AI acts with your permissions, never above them, and it proposes rather than commits. Nothing an assistant produces is stored until a person takes the normal save path.

The assistant

Open the assistant from the top bar — or with ⌘J / Ctrl+J — and on a wide screen it docks beside the page rather than over it, so the record or form it is working on stays in view while it works; on a narrower one it slides in over the page. The panel travels with you through the app, so a conversation survives the navigation it performs. Ask it to build a form, wire logic, style the public form's page, edit a document or notification template, find submissions, or explain what a form does, and it works through the same product surface you would — showing each step it takes. A run of steps folds to one line once it succeeds and stays open while anything is still running or went wrong, so a long answer is not buried under the work that produced it.

A new conversation opens with three ways to begin that belong to the page you are on: a review of the form in the builder, the records still waiting on a form's list, a summary or the next step on a record. Drop a file anywhere on the panel to attach it. A turn that fails offers Try again, and a turn that finishes while the panel is closed leaves a mark on the button so the answer is not missed. The template editors have an Edit with AI button that opens the assistant on the template you have open; an edit it saves while you also have unsaved changes is raised for you to choose between, never silently overwritten either way.

It holds exactly your privileges. Every action it performs is an ordinary authenticated request made as you, so a tool it tries that you could not perform is refused for the same reason it would refuse you. It has no credential of its own and no way to name a tenant: an assistant cannot reach anything you cannot reach.

Deliberate limits, beyond your privileges:

• No deleting or archiving. Destructive actions stay in the interface, where they are deliberate.
• No administration. It cannot create users, grant roles or change billing.
• Navigation is allow-listed — it can only take you to known pages.
• It keeps no memory between turns. The conversation lives in your browser, which is what makes concurrent users safe.

Attach a PDF and it can both read it — rules, thresholds, questions — and turn it into a real, editable form.

Inline AI

Beyond the panel, AI appears where the work is. Each of these is a single call with a structured result, validated against your real data before it can drive anything:

In every case the result lands in the page's own state — prefilled values, an applied filter, a seeded report — and nothing persists until you save it the usual way. The server still re-runs its own engine on save, so an AI suggestion can never become an authoritative number by itself.

Insights

Every form has an Insights page: press Analyze — or set a daily or weekly refresh, at an hour in your own timezone — and you get a stored report on what that form's submissions are doing. The division of labour is strict. The server computes the statistics: submission trends, numeric profiles with outliers, category distributions, correlations between fields, and how long records sit in each workflow state. The AI then narrates those figures into ranked findings, each with a recommendation — and anything it claims about your data is checked before it is stored.

The check has teeth: when a finding targets a slice of submissions (“the West-region records still in Review”), that filter is validated against the form's real fields and then executed — the count on the finding is the query's answer, and View affected submissions opens the ordinary submissions list with exactly that filter, so the number on the card is the number the list shows. A finding whose filter named something the form does not have keeps its narrative but gets no count and no link. Reading reports needs the insight:read privilege; running an analysis or setting the schedule needs insight:run, separately, because analysing spends credits and a schedule commits to spending them regularly.

The same page can also train predictive models on the form's own history: pick a field — or the workflow outcome — and the features to learn from (free-text answers included: what someone wrote in "What happened?" can carry as much signal as any number), and a model trains on past submissions to score the open ones. An outcome model trains only on records that actually reached a final state, and predicts the ones that have not. Its accuracy is measured on held-out data and shown with the fields that drove it, and every prediction appears as a clearly labelled estimate beside the record — never among its real values, never in a calculation, never in an export of the record's data. Training is deterministic compute and costs no AI credits.

Predictions are then held to account. As records close, each one's last prediction while open is compared to how it actually ended, and the model card shows live accuracy in plain words — "62% right on the last 34 closed records — 95% at training" — with drift called out when reality has moved away from the training history. Retraining is one click, or a weekly toggle: the model keeps its identity, the new version simply takes over, and a retrain that fails leaves the old model serving rather than a gap. A second kind of model flags unusual records: train it on what normal looks like and an arriving submission that doesn't fit wears an "Unusual record" mark with the fields that made it odd. And on any record, Test a change answers the what-if — nudge a value and watch the prediction move, with nothing saved.

Agents

An agent is a standing job written in your own words against exactly one form: “when a claim arrives with an estimate over £10,000, summarise it and move it to Review”. Agents run on a schedule or when you start them, and every run is recorded with its outcome — a run in progress is visible while it works.

Three fences bound every agent, and they are enforced when a tool actually runs:

• An explicit privilege grant. An agent holds only the privileges you ticked when you created it — and you cannot grant an agent a privilege you do not hold yourself.
• One form is its whole world. A sibling form's submission is refused even when the underlying permissions would have allowed reading it.
• A run focused on one submission cannot write to another.

A run that has begun is never silently re-executed: its actions are not all reversible, so a run interrupted by a restart is marked failed for a person to look at rather than replayed.

Credits

AI work is metered in credits. Translating a form, analysing one into an insight report, filling one from a document, or turning a typical page into a template costs a handful — a dense multi-page reproduction spends more, at what it measurably cost; generating a whole form from a description costs more still; an agent spends credits as it works. Training a predictive model costs no credits at all: credits meter model reasoning, and training is deterministic compute. Credits reset monthly, and they are the one allowance that stops at the limit rather than degrading — a runaway loop should cost you nothing, and nothing here is ever billed as overage.

Availability

Every AI affordance is gated on the deployment being configured for it: where AI is not configured, these features render nothing rather than failing when clicked — and predictive models additionally require the deployment to run the ML compute service, hiding the same way where it does not. Agents are available from the Team plan; see billing.