[object Object]

The team designed the change-approval workflow as a tidy four-step path. The mining tool revealed it actually executes through 11 distinct paths, three of which loop back through approval, and 60% of requests rubber-stamp through a single approver who reads exactly nothing. The Zurich Process and Task Mining feature exists to close exactly this gap between process design and process reality — but the recommendations need a human filter or you automate away controls that mattered.

The Capability

The Zurich release includes Process and Task Mining — the platform reviews actual workflow execution logs across wf_context, sys_history_line, and task tables, identifies execution patterns (delays, loops, rework, dead branches), and recommends improvements ranked by frequency and impact. It closes the gap between the BPMN diagram in the SOP and the actual sequence of events on production records.

What It Finds

Tasks that take 10x longer than the median for their type — usually a queue or routing problem, occasionally a single fulfiller bottleneck. Loops where the workflow bounces between users (often “needs more info” cycles). Approvals that approve 95% of the time without comment, which are auto-approval candidates. Paths that never get taken in production despite being in the design — dead branches that confuse readers and add maintenance cost. Hand-offs between teams that consistently exceed SLA on the second team’s intake step.

Common findings categories:
  bottleneck_task — single step exceeds median by 10x
  approval_rubber_stamp — approval rate > 95%, mean comment length < 5 words
  rework_loop — record returns to a prior state more than once
  dead_branch — 0% utilization in trailing 90 days
  handoff_lag — first-touch time after assignment > 4h

Acting on Recommendations

Review mining output monthly with the process owner and an operations lead. Prioritize recommendations by frequency multiplied by impact (time, cost, satisfaction). Implement via workflow tweaks, Decision Table changes, or agentic playbook additions — small reversible changes preferred over rewrites. Track impact post-change for at least 30 days; did the inefficiency actually reduce, did anything else regress?

Limits

Correlation, not causation. Mining shows what happens; humans interpret why. Some “inefficiencies” are required regulatory checks that look redundant and are not. An approval that rubber-stamps 95% of the time may still be the audit-required attestation that prevents the 5% bad case from causing a reportable incident. Review carefully before automating away anything that touches compliance, security, or financial controls.

Implementation Sequence

Stand up mining against one high-volume workflow first — incident, change, or service request. Validate the mining tool’s findings against what the operations team already knows; if mining tells you what they would have told you, the model is calibrated. Expand to additional workflows once the operations team trusts the output. Mining everything on day one produces too many recommendations to triage and erodes credibility when most are noise.

What Changed in 2026

The Zurich release added agentic recommendations — the mining engine can now propose specific Flow Designer changes, not just narrative descriptions. Treat the proposals as drafts requiring human review; the mechanical change is correct, the business judgment of whether to apply it is not.

Common Failure Modes

Treating high-frequency findings as automatic implementation candidates without checking which ones the team has already considered and rejected for documented reasons. Acting on findings from sparse data — a workflow with 30 executions per month does not have enough signal for confident recommendations; require minimum execution counts before treating findings as actionable. Losing the historical comparison after changes — capture the baseline metrics before each change so the post-change measurement is meaningful.

What to do this week: run mining against your single highest-volume workflow and read the top five findings; mark which ones the operations team already knew — that hit rate tells you how much to trust the rest.

[object Object]
Share