AI Assistant

Aptli's AI assistant helps you get answers and get records filled without hunting through menus. It comes in two shapes: a chat panel on desktop where you ask questions and let the assistant fill the form or filter the list you're looking at, and a set of voice-and-photo capture flows on mobile that turn what a field worker says and photographs into a draft report, validation, or stock search.

The assistant is always included in Aptli — there's no separate license for it. It can still be switched off deployment-wide by an admin, and it needs a network connection because every request goes to the configured AI provider.

The assistant never changes your data on its own. Everything it produces is a suggestion — a drafted answer, a proposed form patch, a set of filters, a captured record. A human always reviews and clicks Apply or saves through the normal write path. The AI cannot create, edit, or delete records itself.

AI Chat (desktop)

The chat is a slideover — a side panel you open with the sparkles button in the top navigation bar. Type a question in the composer at the bottom and press Enter (or tap send). The thread persists as you move between pages; it's a single panel mounted for your whole session, not a per-page widget.

The assistant does not query the database on its own. Instead, each page offers up the data it's showing — the records in a list, the form you're editing, the features you've selected on the map — and the chat sends a compact, filtered slice of that context along with your question. A context picker at the top of the panel lets you choose which bundle to share. The amount of context is proportional to the task: a status question carries far less than a screen full of features.

Because context is scoped to what your own session can already see, the assistant can never surface data your role restrictions hide from you.

Ask about selected map features

On Field Records, select features the normal way — tap them, or draw a lasso to grab everything in an area — then use the sparkle affordance to open the chat with those features already loaded as context. Ask things like "summarise the work orders covering these poles" or "which of these sites are below minimum stock?" The assistant answers from the features you selected; it does not pull in outside map data on its own.

Let it fill the form

When you have a record form open, the assistant can propose a form patch: describe what you want in plain language and it returns filled-in field values for you to review. Fields you can't edit are never touched, and every value is checked against the form's real fields before it's offered — so a hallucinated field or an out-of-range option is dropped rather than applied. You still click Apply to accept the patch.

Let it filter the list

On list surfaces that support it, ask the assistant to narrow what you're looking at — "show only overdue work orders in Zone B" — and it sets the list's filters and sort for you instead of writing form fields. As with form-fill, any filter it can't match to a real control is silently ignored.

Rate limiting

Each user has a per-hour token budget (default 150,000 tokens, configurable by an admin) measured over a rolling one-hour window — not a fixed number of questions — so many short questions cost far less than a few large ones. When you exceed the budget you're asked to wait; it refills as the hour rolls forward. The same budget covers chat and the mobile capture flows below, so AI spend stays bounded no matter which feature drives it.

Provider and model settings

Admins choose the AI provider and model inline in the chat panel's settings (visible only to holders of the appSettingSchemasModify right). If AI responses are unavailable or degraded, contact your admin.

Voice-and-photo capture (mobile)

On mobile, the assistant powers three field-capture flows. Each turns dictation and photos into a draft you review and submit — never a silent auto-save.

Report capture — dictate what you did and snap photos; the assistant cleans up the transcript, checks that it actually describes work, finds the nearby job or work order to associate it with, and drafts a report with the completed work and consumed materials filled in. You confirm and submit.
Validation capture — photograph the work and speak your verdict; the assistant maps your already-formed pass/fail/needs-revision decision to the correct status and tidies the note. It never forms its own opinion of the work from the photo — the call is always yours.
Find stock (photo search) — photograph an item and the assistant suggests identifier tags, then runs a stock search against them, so you can locate something you can't name.

Works with AI off, too. If an admin has disabled AI or no provider is reachable, these flows still run — your dictation becomes the draft's description verbatim and you fill in the rest by hand. The capture flow degrades to manual entry rather than disappearing.

Importing files with AI column mapping

When you import a CSV or XLSX file through the Data Transfer window, the assistant can help match your file's column headers to Aptli's fields — suggesting a mapping you review, adjust, or override before applying. For GIS-specific formats (GDB, SHP, GeoPackage, DXF, DGN), see GIS File Imports — those go through a separate pipeline.

Tips and limitations

AI needs a connection. Every request goes to an external provider; the assistant can't answer while you're working from device with no signal.
AI never modifies data. All outputs are suggestions; saving, creating, or deleting always requires an explicit action from you.
Context is bounded. Very large selections may be trimmed before being sent. If an answer seems incomplete, narrow your selection or be more specific.
Nothing is stored server-side. Your chat history lives only in your own browser — there's no server-side transcript, so it won't follow you to another device and an admin can't review it.
Provider outages fail gracefully. If the configured provider is down, the assistant shows an error instead of blocking your work.