Agent typologies

The FlowX.AI agent catalog publishes pre-designed AI agents for banking, insurance, and logistics. Behind the industry-specific names, every agent in the catalog is a variation of a small set of recurring typologies - reusable pipeline shapes you can build yourself with integration workflow and Agent Builder nodes. This page describes each typology and how to implement it with the documented node palette.

​

The typologies at a glance

---

​

Document extractor and validator

The most common typology in the catalog: KYC pack validators, completeness checkers, cross-document consistency agents. When to use: a document set arrives (identity documents, contracts, invoices, claims) and you need structured fields out of it, plus a decision on whether the result is good enough to proceed. How to build it: a Document Extraction or Extract Data from File node extracts fields against a schema, then a Condition node gates on the extraction confidence - high-confidence results continue automatically, low-confidence results route to a review branch. See Fan-out extraction for the multi-document-type variant, and AI comparison and reconciliation for cross-checking extracted data against a system of record.

​

Grounded generator

The catalog’s report writers: SAR drafts, case narratives, credit memos, customer letters. When to use: the output is a document or narrative that must be grounded in verified content - policy texts, case data, templates - rather than generated freely. How to build it: a Context Retrieval node pulls relevant chunks from a Knowledge Base, then a Custom Agent node generates the draft from those chunks. Constrain the output with a Response Schema so downstream nodes receive predictable structure, and turn on the Personal Information Guard where the input may contain PII. See Knowledge base RAG for the retrieval setup.

​

Classifier and router

Complaint root-cause identifiers, alert triage screeners, exception handlers. When to use: free-form input (a message, an alert, an email) must be understood and dispatched to the right specialized handler. How to build it: an Intent Classification node classifies the input into up to 10 intents and routes each to its own branch in a single node - including a fallback branch when nothing matches. For classification against a custom label set with custom gating, use a Text Understanding node followed by a Condition node. See Intent classification and routing.

​

Voice intake

Call transcribers, first-notice-of-loss agents, voice-driven notification generators. When to use: the input is audio - customer calls, voice messages, dictated notes. How to build it: a Speech to Text node transcribes the audio, then a Text Understanding node extracts the entities, topics, or sentiment you need from the transcript. See Speech to Text.

​

Predictive scorer

Risk analyzers, churn predictors, KPI forecasters, failure-prediction agents. When to use: a numeric or categorical prediction drives the decision - a score, a probability, a forecast - and you gate the flow on a threshold. How to build it: FlowX orchestrates the pipeline around your scoring model rather than replacing it. Enrich the input with REST Call or DB Operation nodes, call your model (or a third-party scoring service) with a REST Call node, then gate with a Condition node on the returned score. Deterministic scoring logic can also run directly in a Script node. See Hybrid AI and business rules for combining AI outputs with auditable deterministic logic.

​

Privacy-guarded extractor

Identity and KYC extraction agents that must not leak personal data downstream. When to use: documents contain PII and the extracted output crosses a trust boundary - external models, logs, or downstream systems that shouldn’t see raw identifiers. How to build it: the document-extractor pipeline above, with the Personal Information Guard enabled on the AI nodes’ input and/or output to redact personal information. See Personal Information Guard.

​

Knowledge-validated extractor

Board-authority validators, contract-term checkers, coverage validators. When to use: extracted data must be checked against domain reference content - policy rules, product terms, jurisdiction requirements - not just against a schema. How to build it: extract with a Document Extraction node, retrieve the applicable reference content with a Context Retrieval node, then have a Custom Agent node compare the two and return a structured verdict with reasons via its Response Schema.

​

Relationship reasoner

Ownership-structure mappers, treaty cross-referencers, entity-network agents. When to use: the question is about how entities relate - who ultimately owns what, which contracts reference each other - rather than about a single document. How to build it: there is no dedicated graph node; model the reference relationships in a Knowledge Base, retrieve the relevant slice with Context Retrieval, and let a Custom Agent node reason over it, returning the reconstructed structure through a Response Schema. For relationship data held in your own systems, enrich with DB Operation or REST Call nodes first.

​

Human handoff

Escalation wrappers that route edge cases to a person. When to use: any of the typologies above produces a low-confidence or high-stakes result that a human must approve. How to build it: a Condition node detects the escalation case, and the surrounding BPMN process presents it to a person as a User Task - the agent workflow and the human step share the same process context. See Integrating agents in BPMN processes.

---

​

Combining typologies

Catalog agents rarely ship alone - each lists companion agents, and the same is true when you build your own. Typical combinations:

---

​

Related resources