AI

QBO orchestrates AI, not the other way around. A model never decides on its own what to read or what to change. A workflow step, a scheduled job, a menu option, or an API call invokes an AI method on a specific record. QBO chooses the model, sends only the data that record’s security allows, and saves the answer as ordinary QBO data: a document template, a label, a score, a form, a message, or new records. From there your rules, workflows, worklists, and reports treat it like any other data, and it sits inside the same audit trail.

What AI does in QBO

Building block	AI method	Where the result lands
Documents	Classify a document into one of your document templates	The document’s template
	Transcribe an audio file	The document’s description (searchable text)
Any record	Apply labels, such as language or sentiment	Labels on the record
	Review a record (and any documents you pass) and write a memo	A message on the record
	Update fields from what a model reads in a document	The record’s own fields
Scoring	Extract typed values from text, or directly from the original PDF or image	Score items, each with a confidence, plus a weighted confidence for the score
Imports	Turn a document into records, including nested child records	New records, through the standard import path
Forms	Fill in a form from the data on its parent record	Form fields and answers
Messaging	Draft a reply to an inbound email	A reply message, for a person to review and send
Reports	Turn a plain-English question into a query	A single read-only SELECT in the query editor, checked against your schema without running it
Queues	Send large volumes to a provider’s batch service	Results applied record by record as the batch completes

Key concepts

Concept	In QBO	What it does
Agent	AiAgent	A named configuration: provider, model, instructions, retry count, and the configuration entry that holds the API key. Agents can inherit from a base agent.
Provider	IModelResolver per provider	The model vendor behind an agent: Claude (Anthropic), Gemini (Google), OpenAI, or Amazon Bedrock.
Instructions	Agent instructions or a template	The standing guidance for a task, kept in the agent or in a message or import template so power users can change it without code.
AI engine	AIScoreEngine, AiModalEngine, AiImportEngine	Plugs AI into scoring and imports, configured on score and import templates.
Confidence	Score.Confidence, ScoreItem.Confidence	How sure the model is of each extracted value, rolled up as a weighted average.

Guardrails built in

• Your security applies. AI methods run as the calling user, on records that user can reach.
• People stay in the loop where it matters. Email reply drafts open for editing; nothing is sent unless someone (or a workflow you configured) sends it.
• Read-only reporting. Generated SQL is checked to be exactly one SELECT (no SELECT … INTO) and validated against your database schema without running it. A query that can’t be validated comes back with a warning, and nothing runs until a person runs it.
• Costs are visible. Token usage can be posted to a ledger, split by input, cached input, output, and reasoning tokens.

Choose your view

Use the buttons under the page title. End user covers the AI options you’ll see on documents, messages, and reports; Power user covers agents, templates, and engines; Developer covers the API, batching, and cost tracking.

Walkthroughs that use this

• Classify a document and extract its data with AI
• Draft replies to inbound email with AI
• Review a loan’s documents for a prepayment penalty