Make the Automation-vs-Augmentation Call Before Anyone Picks a Tool

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

The most expensive sentence in enterprise AI is "let's just add it to the current process." McKinsey's finding is blunt: workflow redesign is the strongest predictor of AI impact, yet only about a fifth of companies redesign end to end — most bolt a capable model onto one step of a sequential-handoff process and then wonder why nothing moved. The analyst who can say, before tool selection, "this one augments, this one automates, and this one is worthless without redesigning the queue around it" is doing the work that separates funded impact from pilot theater.

When to use — and when not to

• Use it on every candidate that survives an opportunity scanthe treatment call from the scan is provisional; this is where it gets argued properly.
• Use it when a sponsor arrives with a tool already in mindrunning the triage first reframes "should we buy X?" into "what is the right treatment for this step?", which is a question you can actually answer.
• Use it when the value case looks great but the process is a relay racesequential handoffs are exactly where the layering anti-pattern hides.
• Skip it for steps already triaged KEEP-HUMANrelitigating settled judgment calls burns credibility; revisit only when the step's nature changes.
• Skip it as a headcount exercisethe output is a work-design decision, not a reduction plan; letting it read as one poisons every SME conversation that follows.
• Skip tool comparisons entirely at this stagethe moment a vendor name enters the analysis, the analysis starts justifying the vendor.

The recipe

1. 1.

   Write the step's decision profile. For the candidate step, document: what is actually being decided (not what is being typed), volume and variance, how verifiable the output is (can a checker confirm correctness cheaply?), the cost of a wrong outcome, tolerance for review latency, and who is accountable when it goes wrong. Interview the people doing the work — the silent judgment they apply is the single most common thing this profile misses.

2. 2.

   Apply the heuristics explicitly. Run the profile through the taught pattern:

   Copy

   You are helping a business analyst make an automation-vs-
   augmentation call. Here is the step's decision profile:
   [PASTE PROFILE]
   
   Apply these heuristics and show your reasoning per line:
   - AUTOMATE leans: high volume + low variance + output cheap
     to verify + wrong answers detectable and tolerable +
     accountability can rest on the process, not the person.
   - AUGMENT leans: judgment or negotiation in the loop + high
     stakes per case + verification requires the expert anyway
     + accountability must stay with a named human + trust or
     relationship IS part of the output.
   
   Output: 1) the lean per heuristic line with a one-line
   justification from the profile; 2) an overall call —
   AUTOMATE / AUGMENT / SPLIT (name which sub-tasks go which
   way); 3) the two strongest arguments against your call;
   4) what evidence would flip it.
   Do not consider any specific product or vendor.
   

   SPLIT is the most common honest answer: drafting automates, the judgment stays augmented, and naming the boundary is the deliverable.

3. 3.

   Hunt the layering anti-pattern. Now examine the surrounding process, not the step: map the handoffs upstream and downstream and ask the defining question — if this step becomes instant, what actually improves end to end? Sample finding:

   Copy

   Step: draft claim-assessment letters (AUTOMATE lean).
   Current flow: assess → queue (1.8d avg) → draft → queue
   (0.9d) → review → dispatch batch (daily, 17:00).
   Layering check: automating drafting removes 40 min of an
   8.2-day cycle. End-to-end delta: ~0.5%. ANTI-PATTERN —
   value case requires collapsing queues (draft at assessment
   time, continuous dispatch), i.e. redesign, not insertion.
   

   If the answer is "the step gets faster but the process doesn't," name it: the McKinsey layering anti-pattern, in writing, in the memo.

4. 4.

   Decide redesign scope before tool selection. Three scopes, chosen deliberately: INSERT (AI into the existing flow — legitimate when the step itself is the bottleneck), RESHAPE (collapse adjacent queues and handoffs around the AI step), REDESIGN (rebuild the flow around what the AI makes possible — often turning a sequential relay into a review-the-exceptions pattern). Sketch the to-be flow for the chosen scope at BPMN level with the AI's help, and cost the change-management delta honestly — RESHAPE and REDESIGN change people's jobs, and that cost belongs in the same memo.

5. 5.

   Run the scenario comparison. Have the AI draft the comparison across scopes: end-to-end cycle time, touch time, exception handling, review load, and failure behavior per scenario, each figure tied to the step data or marked ASSUMPTION. You adjudicate every number. This table is what makes "insert now, redesign later" an informed choice rather than a default.

6. 6.

   Write the one-page triage memo. The call (automate/augment/split), the boundary if split, the layering finding with the end-to-end math, the chosen redesign scope with its to-be sketch, the arguments against and what would flip the call, and the recommended next step (feasibility spike or business case — with the scope decision attached so tool selection inherits it). Signed by you and the process owner.

Make it yours

• Redesign canvases for workshop settings: run steps 2–4 as a structured canvas session with the process owner and SMEs — the heuristic lines become sticky-note prompts, and the layering check becomes a wall exercise on the printed map.
• BPMN to-be modeling for the chosen scope: generate the to-be flow as Mermaid/BPMN from the scope decision and run it through the same SME walkthrough discipline as any process model.
• Claude scenario analysis for the comparison table: attach the step data and let the model build the INSERT/RESHAPE/REDESIGN comparison — then audit every arithmetic line before it reaches the memo.
• Portfolio mode: keep every triage memo's call, scope, and flip-conditions in one register — patterns emerge fast (the org that discovers every AUTOMATE call is blocked by the same shared queue has found its real project).

Quality check

• The decision profile came from the people doing the worksilent judgment surfaced by interview, not assumed from the process map.
• Heuristics were applied line by line, in writingthe call is traceable to the profile, and the arguments against it are recorded.
• The layering check did end-to-end mathstep-level savings versus process-level delta, with queues and handoffs counted.
• Redesign scope was decided before any tool talkand the memo states it explicitly so vendor selection inherits the constraint.
• Change cost is in the same memo as the benefitRESHAPE/REDESIGN scopes carry their people-impact honestly.
• The flip conditions are namedwhat evidence would change the call, so the decision is revisitable instead of re-arguable.