Shapes

People, AI and code

AI task

A model gives a typed answer about each record: extracted fields, a label, a summary, a translation or a sentiment.

Nodes › AIOpen in Neblex
Good for
Structured answers from AI
Role
Works on the records in a flow
Find it in
Nodes › AI

When to use it

  • Fields should be pulled from emails, PDFs or images into your structure.
  • Tickets or messages should be classified, summarized or translated.
  • Answers a model is unsure about should go to a person to check.

Reach for something else when

Example

Extract invoice number, date and total from each emailed PDF, and send invoices with low confidence to a person to check.

  1. When an email arrives
  2. AI task
  3. Branch
  4. Wait for form

How it works

Settings
Model provider, its Connection and the Model (any provider an AI agent can use, including Ollama or an OpenAI-compatible server through an On-Prem Worker); for Extract with AI the Fields to extract as a JSON Schema; for Classify the Labels, one per line; for Translate the language to Translate into; optional Extra instructions; Review below confidence (0 to 1); Ask again [2]; Temperature.
Content
The record's text field, or its document file ({filename, contentType, contentBase64}, as Read file and uploads hand files on): images and PDFs go to models that read them directly (Anthropic, Amazon Bedrock and, for PDFs, OpenAI and Azure OpenAI; images also to the OpenAI-compatible providers and Ollama), and text files are read as text. Without either, the whole record is the content. Content of any length runs: it goes to the model in parts and the answers are combined (fields merged, summaries summarized again, translations joined in order).
Output
The record with the task's fields added: extracted; label; summary; translation and detected_language; or sentiment and score (from -1 to 1). _ai tells you how it went: model, tokens, schema_valid, reasks, parts, model_reported_confidence (the model's own estimate) and review_needed.
Checking answers
Every answer is checked against its structure. One that does not match is asked for again with what was wrong, up to Ask again times. A record whose answer never matched, or whose confidence is below Review below confidence, gets _ai.review_needed = true: add a Branch on it that sends those records to a Wait for form step so a person checks them, and the rest straight on. After Classify, a Branch or Rules step on label routes each record.
Cost
Tokens are metered like an AI agent's and count toward the workspace's monthly AI budget; once it is spent, the step stops before its next part with the budget message. Replaying a run reuses the recorded answers without calling the model.
More in the guide

Add it to a flow

  1. Open the flow from Build › Workflows.
  2. Open Nodes and find AI task under AI, or search for it. Click it or drag it onto the canvas.
  3. Drag from the previous step's output handle to its input, then open its settings with ✎ Edit.
  4. Run the flow and click the step to see the records it received and produced.