Run a Lean Waste Critique Against a Mapped Process

Workflow area: Process Modeling & Mining · Tier: Foundation · Hands-on time: ~25 minutes

Waste analysis has a cold-start problem: staring at a swimlane asking "what's wrong here?" produces either nothing or your existing prejudices. Running the map through the Lean categories systematically is exactly the kind of exhaustive, checklist-driven pass AI does without fatigue — it will interrogate every handoff and every approval, including the ones you stopped seeing years ago. The danger is symmetrical: the model will happily condemn a "redundant approval" that exists because of a regulatory finding in 2019. So the output is a register of candidates, each one validated with the person who owns the step before it is called waste in any document that circulates.

When to use — and when not to

• Use it after a process map is SME-validateda critique of an unvalidated map (see pb-02's straw-man caveats) compounds fiction with judgment.
• Use it with cycle-time or volume data in handwaste claims with numbers ("approval adds 4.2 days median") survive scrutiny; waste claims without numbers start arguments.
• Use it before solution designan improvement project that skips waste analysis automates the mess as-is; this critique is the case for redesign-first (the theme pb-36 and pb-37 build on).
• Skip it when the process is about to be replaced wholesalecritiquing a process scheduled for decommission is archaeology, not analysis.
• Skip it as an ambushrunning a waste critique on a team's process without their knowledge and presenting conclusions to their leadership is a trust-destroying move; involve the process owner from the start.

The recipe

1. 1.

   Stage the inputs. You need the validated process map as text (actor — step — system, numbered; export from Miro/Signavio or transcribe), plus whatever performance data exists: cycle times per step, volumes, error or rework rates, staffing. Data can be rough — label its quality honestly.

2. 2.

   Run the categorized critique.

   Copy

   You are a Lean-trained business analyst reviewing a validated
   process map for waste.
   Process: [NAME + ONE-LINE PURPOSE + VOLUME, e.g. "vendor
   onboarding, ~150/month"]
   Map:
   [PASTE NUMBERED STEPS: actor — step — system]
   Performance data:
   [PASTE CYCLE TIMES / VOLUMES / REWORK RATES, with a note on data
   quality, or "none available"]
   
   Sweep every step and every transition against these categories:
   waiting, rework/defects, overprocessing, handoffs, motion
   (system-switching), inventory (queues/backlogs), underused
   judgment (skilled people doing mechanical work).
   
   Output a WASTE REGISTER table with columns:
   - ID (W-01...)
   - Step(s) affected
   - Category
   - Observation — what the map/data shows, citing the step number
     and any data point used
   - Estimated impact — only from provided data; write "no data" if
     none supports a number
   - Possible legitimate reason — your best guess at why this might
     exist deliberately (control, regulation, risk)
   - Question for the step owner — the neutral question that would
     confirm or kill this candidate
   
   Rules: every row cites step numbers. Never present impact numbers
   not derivable from the data given. The "legitimate reason" column
   is mandatory — assume the process has reasons you cannot see.
   

3. 3.

   Read a sample of good output.

   Copy

   W-03 | Steps 4→5 | Waiting | Supervisor approval sits in a shared
   mailbox; median 4.2 days of the 9-day cycle (from cycle data) |
   ~47% of total cycle time | May be a segregation-of-duties control
   | "Is this approval required by policy, and must it be this
   person, or this role?"
   
   W-07 | Steps 6, 9 | Rework | Vendor data re-keyed into ERP after
   entry in CRM (steps cite two systems) | no data | Integration may
   have been descoped for cost | "What happens when the two entries
   disagree, and how often?"
   

4. 4.

   Triage before you validate. Sort the register: candidates with data support and large impact first. Kill obvious false positives yourself — you know some history the model does not. Mark each survivor with what it needs: data confirmation, owner confirmation, or both.

5. 5.

   Validate each candidate with its owner — the non-negotiable step. Take the questions (not the conclusions) to the people who run the steps. The framing matters: "the map shows the approval adds four days — help me understand what it protects" invites information; "we found waste in your step" invites war. Three outcomes per candidate: confirmed waste (owner agrees it serves no purpose), justified control (there is a reason — record it on the map so the next analyst doesn't rediscover it), or partial (the purpose is real but the implementation is wasteful — often the richest category).

6. 6.

   Publish only the confirmed register. The circulating artifact contains confirmed items with their data, owner acknowledgments, and the justified-control notes. Candidates that died in validation stay in your working file as institutional learning.

Make it yours

• Claude/ChatGPT with a value-stream frame: for end-to-end flows, ask additionally for value-add / non-value-add / non-value-add-but-required classification per step and a rough value-add ratio — the classic value-stream lens, computed honestly from your cycle data.
• Miro value-stream templates: rebuild the confirmed register visually on a Miro value-stream map for the improvement workshop; the AI critique becomes the pre-work that makes the workshop start at "what do we fix" instead of "what is wrong".
• Process-mining data as input: if you have Celonis/Signavio metrics (pb-25 territory), feed variant frequencies and waiting-time distributions into the same prompt — the register's impact column stops saying "no data".
• Team scale: run the same prompt structure across several processes with identical categories and columns, and the registers roll up into a portfolio-level waste picture leadership can prioritize.

Quality check

• Every observation cites step numbers and real dataan impact figure that cannot be traced to the provided data was invented; strike it.
• No invented process behaviorthe model may only critique steps that are on the map; if a row references a behavior the map does not show, it hallucinated a step.
• The legitimate-reason column is filled everywherea register that assumes malice or stupidity everywhere is a polemic, not analysis.
• Every published item has owner validationconfirmed, justified, or partial, with a name and date; unvalidated candidates never appear in circulated documents.
• Justified controls got documentedthe "why this exists" answers are annotated onto the process map itself, permanently.
• Tone survivedread the register as the process owner would; if any row reads as blame, rewrite it as a question.