Capabilities

Talonic includes an embedded AI agent accessible from every page via Cmd+I (Ctrl+I on Windows). The agent understands your workspace context and can inspect schemas, search documents, analyze extraction quality, explore cases, and build schemas — all through natural language. It serves as a conversational interface to your entire workspace, eliminating the need to navigate through multiple pages to find information or perform common operations.

The agent is context-aware, meaning it automatically knows which page you are on and what data is visible. If you open the agent from a document detail page, it already has that document in scope and can answer questions about its extracted fields, processing status, or classification without you needing to specify which document you mean.

The agent classifies every user message as either a question (answered with information) or a command (triggers an action). Questions are handled instantly with read-only access, while commands go through the impact-level system to ensure safety. The agent streams its responses in real time, so you can see reasoning unfold as it queries your workspace data.

There are important limitations to be aware of. The agent cannot access external systems or the internet — it only works with data already in your Talonic workspace. It cannot bypass permission boundaries, so team members with read-only (Viewer) access cannot use the agent to make changes. Long-running operations like full batch extractions cannot be triggered through the agent; those must be initiated from the relevant UI page. The agent also cannot modify field registry entries directly — those changes flow through the resolution process.

The agent is particularly effective for onboarding new team members. Instead of reading documentation about each platform feature, new users can ask the agent questions like "How many document types do we have?", "What schemas are available?", or "Show me our most common fields." The agent provides instant, contextual answers that help users build a mental model of their workspace. This reduces time-to-productivity for new team members from days to hours.

What the Agent Can Do

The agent has deep access to your workspace and can help with:

Example Agent Interactions

The agent excels at cross-cutting queries that would otherwise require navigating multiple pages. For example, you can ask it to summarize extraction quality across your latest job runs, identify which document types have the lowest confidence scores, or compare field coverage between two schemas. The agent queries the underlying data in real time and streams results, so complex analyses that would take several minutes of manual navigation are answered in seconds.

Schema creation through the agent is particularly powerful. Describe the fields you need in plain language — for example, "Create a schema for purchase orders with vendor name, PO number, line items, unit price, and total amount" — and the agent maps each field to the registry, identifies the best match for each, and creates a draft schema ready for your review. This is faster than manually searching the registry and adding fields one by one through the schema editor.

Behind the scenes, the agent uses the same internal APIs as the web interface. When you ask "Show me the documents processed this week", the agent queries the documents endpoint with a date filter. When you ask "What is my capture rate?", it reads the telemetry data from the dashboard service. This means the agent always shows the same data you would see in the UI — there is no separate data layer or cache that could show stale results.

curl "https://api.talonic.com/v1/documents?after=2026-05-01T00:00:00Z&status=completed" \
  -H "Authorization: Bearer $TALONIC_API_KEY"

{
  "data": [
    {
      "id": "doc_7f3a1b2c",
      "filename": "invoice_2026_0472.pdf",
      "type_detected": "Invoice",
      "status": "completed",
      "extraction_count": 1
    }
  ],
  "pagination": { "total": 156, "limit": 20, "has_more": true }
}

Spec Architect

When you ask the agent to help design or tune a Spec, it acts as a Spec Architect: it reasons over the real configuration catalog rather than from memory, and it can estimate how much human review a Spec will generate before you commit to it. This matters because the cost of a Spec is not just extraction — it is the number of cells a person has to review, and that is driven by which validation gates and triage thresholds you place in the rail. The agent grounds every suggestion in the actual primitives and their review-volume effect.

Review estimation comes in two touches. Touch-0 (estimate_review_volume) is an instant, free, read-only heatmap of which fields are likely to need review, computed from data already captured at ingestion — it returns a qualitative per-field risk band (low / medium / high), not a percentage, and is the right tool for a Spec that has not run yet. Touch-1 (estimate_review_rate) is the precise, warm estimate over a completed pipeline run: it replays the real triage decision over the run's real cells and validation verdicts, returns the actual review rate under a candidate triage config, plots the review-rate-vs-capacity tradeoff curve, and — given a target review rate — back-solves the capacity and risk cut that hit a sustainable review budget. Neither touch extracts or writes anything.

These capabilities run inside the agent loop; there is no separate REST endpoint to call them directly. GET /v1/agent/tools lists the tools the agent can use, including the Spec Architect set, and you reach them by asking the agent in natural language — for example, "Estimate the review rate if I cap review capacity at three cells per document" or "Which fields will need the most review on this Spec before I run it?".