Why BPM Should Orchestrate and Agents Should Decide

BPM for Control, Agentic AI for Intelligence

Most real GenAI automation use cases are landing in the Intelligent Process Automation (IPA) space. This involves tools like Kofax, Abbyy FlexiCapture for case initiation in the form of documents, calls received by a call centre agent, alerts triggered by monitoring systems, ECM for document storage, and a BPM layer for process orchestration.

Why BPM Still Matters

Every enterprise has a BPM layer for rule-based process orchestration, which is used for:

• Clear end-to-end process orchestration

• Clear auditability and traceability

• Highly rule-based, predictable processing

An architecture pattern that can be used by a case management system can look like this:

BPM → LLM → Parse LLM Output → Define Next Step

But is this the right approach for GenAI automation?

The Role of LLM Orchestration Frameworks

Multiple LLM orchestration frameworks are available in the market, with LangChain being one of the most widely used. These frameworks provide the following flexibility:

• Chains of sequences that can be dynamically defined by the LLM

• LLM calls completely decoupled, making switching service providers plug-and-play

• Context and memory management

• Easy retry logic in case of failure

BPM can manage state very efficiently, and it can call LLMs. It can even retry calls. So should we not use LangChain or an LLM orchestration framework at all?

Decoupling BPM and LLMs

Here is my take on this:

• BPM and LLM calls should be completely decoupled

• BPM should orchestrate the process that is traceable, predictable, and accountable

• BPM should call an AI service, and this is where a LangChain-like framework should be implemented

LangChain should take the input and process the document. Dynamic decisioning and probability-based statistics should be calculated here. This is where Agentic AI should come into the picture.

LangChain should send back:

• Confidence score

• Recommended next step

• Reasoning / explanation

BPM should then decide the next step based on the confidence level. BPM should use a human-in-the-loop feature for accountability.

Claim Management Example Architecture

BPM: Case Initiation (Process Orchestration)

Handled by BPM (e.g., IBM BAW,Camunda ,JBPM)

A case is created when:

• A claim is submitted via a portal

• An email is ingested via ECM

• Document is scanned via Kofax abby lex capture

BPM responsibilities:

• Create case ID

• Store metadata for supporting documen

• Attach supporting documents in ECM

Key point: No AI logic here. BPM makes a REST call, not a direct LLM call. BPM remains deterministic. AI logic stays isolated.

LangChain: AI Orchestration (Core Intelligence)

Now LangChain takes over.

Document Loading and Chunking

LangChain:

• Fetches PDFs from ECM

• Splits text intelligently (claims, invoices, reports)

RAG (Retrieval Augmented Generation)

LangChain retrieves:

• Policy documents

• Claim rules

• Past similar claims

This context is injected into the prompts.

Agentic Workflow:

• Extract claim details

• Validate mandatory fields

• Check policy coverage

• Assess risk level

• Generate structured JSON output

Output Validation and Retry

LangChain:

• Enforces JSON schema

• Retries if output is invalid

• Falls back to another model if required

Example output:

{ "claimAmount": 12000, "coverage": "YES", "riskScore": 0.32, "confidence": 0.86, "missingDocuments": [] }

LangChain → BPM Response

LangChain returns decision support, not the final decision.

{ "recommendation": "AUTO_APPROVE", "confidence": 0.86, "explanation": "Policy active, no exclusions, low risk score" }

Rule-Based Decisioning in BPM

Now BPM does what it does best.

Example BPM rules:

• If confidence ≥ 0.85 → auto-approve

• If confidence < 0.85 → human review

• If documents are missing → request information

No probabilistic logic inside BPM.

BPM: Human-in-the-Loop Review

If routed to a human:

• Reviewer sees:

  • AI recommendation

  • Confidence score

  • Explanation

Reviewer actions:

• Accept

• Override

• Request more documents

All actions are:

Logged Audited SLA tracked

Finally:

• Update claim status

• Notify customer

• Archive documents in ECM

• Close case