The loudest voices in agentic AI want you to believe your BPM estate is legacy to be replaced. The quietest risk is pretending nothing has changed. Both are wrong. The right question isn't "agents or BAW?" — it's "which of my processes earn autonomy, and which earn determinism?"
What BAW is still better at
Long-running, stateful processes with SLAs, escalations and compensation logic. Deterministic execution you can replay in an audit. Human task management with real governance around it. If a process is stable, regulated and high-volume, a workflow engine is not technical debt — it's the control plane. No agent framework today gives you what a mature BAW estate gives you out of the box: guaranteed state, guaranteed sequence, guaranteed evidence.
What agents are better at
The messy edges your BPM never absorbed: unstructured intake, exception triage, cross-system research, drafting responses, deciding which process to invoke and with what data. In most estates we assess, 80% of the operational pain lives in the 20% of work BAW never modeled — the swivel-chair work between processes, not the processes themselves.
Pattern 1 — wrap
Keep the BAW process as the system of record. Expose it to agents as a tool — via REST APIs or MCP — and let the agent handle intake, context-gathering and exceptions, then start or advance the process. The process stays auditable and untouched; the agent removes the friction around it.
Low risk, fast value, no migration. This is where most regulated organizations should start — and IBM's own direction supports it: watsonx Orchestrate's agentic workflows are explicitly designed to combine deterministic steps with agent autonomy in one governed flow.
Pattern 2 — rebuild
Some processes deserve more. If a process is mostly routing decisions between humans, if its structure changes faster than you can redeploy it, if the BPMN diagram was always a fiction the operators worked around — rebuild it as an agentic workflow: deterministic checkpoints where compliance demands them, agent autonomy in between.
Rebuilds pay off precisely where the original process automated the form of the work rather than the work itself.
The four-question framework — per process, not per estate
- How stable is the happy path? Has the core flow changed in three years, or three times this year?
- What's the cost of a wrong decision? Reversible inconvenience — or a regulated, reportable event?
- How much of the work happens around the process rather than in it — intake, triage, chasing, re-keying?
- How deep is the existing investment? Integrations, audit history, trained users, tuned SLAs.
Stable + regulated + deep → wrap. Volatile + reversible + shallow → rebuild. Everything else → wrap first, rebuild selectively.
A portfolio decision, not a platform decision
The estates that modernize well don't pick a side. They run the four questions against every significant process, wrap the majority, rebuild a deliberate minority, and retire what neither deserves. That's the difference between a modernization roadmap and a platform bet.
We run this assessment against real BAW estates — process by process, with effort estimates attached. It's usually the fastest two weeks of clarity an automation team can buy.