Scan a Process for AI Opportunities Step by Step — and Score Them Honestly

Workflow area: AI Opportunity & Feasibility · Tier: Advanced · Hands-on time: ~45 minutes

Most "AI opportunity" work starts from the technology and goes hunting for a use — which is how organizations end up in the 95% of GenAI pilots that never touch the P&L. The analyst's edge is starting from a process that is actually understood: a validated map with volumes, cycle times, and error rates per step. Against that substrate, the treatment scan is systematic rather than inspirational, and the scoring matrix converts enthusiasm into a ranked backlog where the boring rules-automation candidate is allowed to beat the glamorous agent.

When to use — and when not to

• Use it when leadership asks "where should we use AI?"the per-step scan converts an unanswerable strategy question into a process-analysis exercise you already know how to run.
• Use it on a validated process map with operational datavolumes, cycle times, error rates per step; that evidence is what the scoring feeds on.
• Use it to triage a pile of AI ideas from the businessplacing each idea on the map and scoring it alongside systematically generated candidates kills pet projects gently and with evidence.
• Skip it on an unvalidated mapscanning a process that does not reflect reality produces confidently misplaced opportunities; run the SME walkthrough first.
• Skip it without step-level datatreatment calls without volumes and error rates are taste; gather at least rough per-step figures or mark the scan provisional.
• Skip promising treatments as decisionsthe backlog ranks candidates for feasibility work (spikes, data-readiness checks, business cases); it does not commit to building anything.

The recipe

1. 1.

   Stage the substrate. Take the validated process map and attach per-step data: volume per period, cycle time, error/rework rate, headcount touching it, and systems involved. Where a number is missing, enter ESTIMATE with its basis. The scan's credibility is exactly the credibility of this table.

2. 2.

   Run the per-step treatment scan. Attach map and data:

   Copy

   You are supporting a business analyst scanning a process for
   AI opportunity. Attached: validated process map and per-step
   operational data.
   
   For every step, output one row:
   STEP | what it does | volume | cycle time | error rate |
   TREATMENT: ELIMINATE (step exists only to compensate for an
   upstream problem) / RULES (deterministic logic, no judgment)
   / AI-ASSIST (human decides, AI drafts or retrieves) /
   AI-AGENT (AI acts within bounds, human handles exceptions) /
   KEEP-HUMAN (judgment, negotiation, accountability, or trust
   IS the work) | rationale in one sentence tied to the step's
   own data | what the AI would need (inputs, data, integration)
   | primary risk if the treatment fails.
   
   Rules: consider ELIMINATE before any automation — do not
   optimize a step that should not exist. RULES beats AI where
   logic is expressible: cheaper, deterministic, auditable.
   Steps whose output feeds a step you marked KEEP-HUMAN must
   note the handoff. Do not assign AI-AGENT to any step lacking
   a defined verification path. Flag any step where treating it
   alone (leaving handoffs intact) would trap the process at
   handoff speed — mark those REDESIGN-SCOPE.
   

3. 3.

   Challenge the treatment calls — this is your analysis, not the model's. Three challenges per row: Is the ELIMINATE really removable, or is it a control wearing a compensation costume? Is the RULES call hiding judgment the SME applies silently? Does the AI-AGENT call survive the question "how would anyone know it acted wrongly?" Take contested rows to the process owner; their corrections are findings.

4. 4.

   Score the surviving candidates. For each non-KEEP-HUMAN row, score three axes 1–5 with written anchors: VALUE (hours × volume × error cost actually addressable), FEASIBILITY (data exists and is usable; output is verifiable; integration is plausible), RISK inverted (customer impact, compliance exposure, failure detectability). Have the AI compute value × feasibility × (6 − risk) and sort — but write the axis scores yourself or adjudicate the model's proposals row by row. Sample backlog:

   Copy

   1. Claims triage (AI-ASSIST) — V5 F4 R2 → 80. 41k/yr, 22%
      misroute rate; suggestion+review preserves audit path.
   2. Document completeness check (RULES) — V4 F5 R1 → 100 raw,
      capped: candidate for existing workflow engine, not AI.
   3. Status-inquiry responses (AI-AGENT) — V4 F3 R3 → 36.
      REDESIGN-SCOPE: bolting onto current queue keeps 3-day
      SLA; value case assumes queue redesign.
   

5. 5.

   Write the redesign-scope note. For every REDESIGN-SCOPE flag, state plainly: this opportunity's value case depends on redesigning the surrounding handoffs, not inserting AI into the existing sequence — the McKinsey finding that workflow redesign, not model choice, predicts impact. This one paragraph is what separates your backlog from a tool-shopping list, and it must be visible to the sponsor before tool conversations start.

6. 6.

   Package the backlog with its exits. Deliverable: the treatment table (KEEP-HUMAN rows included — showing what you chose not to automate is half the credibility), the scored backlog, the redesign-scope notes, and a recommended next action per top candidate — feasibility spike, data-readiness assessment, or business case. Each next action has an owner and a cheap kill condition.

Make it yours

• Celonis/Signavio insights as the data layer: where process mining runs, pull per-step volumes, durations, and rework rates from the tool instead of estimates — and let variant data reveal steps the official map forgot.
• Opportunity-matrix templates for portfolio rollups: keep the scoring anchors and matrix format standard across processes so backlogs from different departments merge into one investment view.
• Workshop mode: run steps 2–4 live with the process owner, projecting the treatment table and arguing the calls — slower per step, but the validation happens in-line and adoption starts in the room.
• Lightweight mode for a small process: for a sub-ten-step process, collapse to the treatment table plus a top-three ranking — the discipline of ELIMINATE-first and REDESIGN-SCOPE flags matters more than the full matrix.

Quality check

• The map was validated firstthe scan ran against walkthrough-corrected reality, not the official diagram.
• Every treatment ties to the step's own datarationales cite volume, error rate, or cycle time; ESTIMATE-based rows are marked provisional.
• ELIMINATE was considered before automationand suspected controls were confirmed with the owner before being scheduled for removal.
• No AI-AGENT without a verification pathevery agent candidate states how wrong actions get detected and by whom.
• REDESIGN-SCOPE flags are visible in the summarythe sponsor sees which value cases depend on redesign before any vendor does.
• Scores are anchored and ownedaxis definitions written down, scores adjudicated by you and the process owner, arithmetic auditable.