Roughly 95% of enterprise GenAI pilots show no P&L impact, and the strongest predictor of the exceptions is workflow redesign — analysis work. This module teaches the differentiator half: per-step opportunity scans, the automation-versus-augmentation call, business cases with measured baselines and full cost stacks, data-readiness assessments, feasibility spikes, and pilots with kill criteria.
Raj is a Principal Consultant at a large services firm, fourteen years in, recently retitled "AI Transformation Consultant." He has watched three client AI pilots die, and the third one hurts because it was his. The client's CFO wanted "agents in accounts payable." A vendor demo dazzled. A pilot launched: an AI agent drafting responses to invoice-exception emails, bolted onto step four of a nine-step process. Six months later the steering committee killed it, and the post-mortem found no villain — just five absences.
No measured baseline: nobody had timed exception handling before the pilot, so "faster" was unprovable. No redesign: the agent drafted emails that still waited in the same queues, behind the same approvals — the process ran at handoff speed with one shinier step. No data-readiness check: the agent needed PO history that lived in a system the pilot could not legally access under the client's data-residency terms; discovered in week seven. No honest cost stack: tokens, integration and the change-management effort nobody budgeted quietly tripled the projection. No kill criteria: so the pilot zombied along for three extra months because nobody could say what failure looked like.
Raj's conclusion was not "AI doesn't work." It was that nobody had done the analysis — and that every absence on the list is a standard BA artifact pointed at a new subject. That reframing is this module. McKinsey's finding that only ~21% of companies redesign workflows end-to-end, MIT's ~95% no-impact figure, Gartner's prediction that 60% of AI projects unsupported by AI-ready data will be abandoned — read together, they say the market is drowning in AI enthusiasm and starving for AI analysis. Scarcity is a career strategy.
Every failed-pilot statistic in this module is an argument for analysts, not against AI. Pilots rarely die because the model was too weak; they die of unmeasured baselines, unredesigned processes, unready data and unkillable scope — all preventable by artifacts you already know how to build.
The scan is the differentiator's opening move, and it starts from an asset you built in Module 4: a validated process map with volumes and error rates. (Unvalidated maps produce opportunity scans of a process that doesn't exist — Ananya's merged exception paths would have poisoned every downstream call.) Per process step, classify the candidate treatment:
Before automating anything, ask whether the step should survive redesign at all. Waste identified in Module 4 — the double-approval nobody remembers instituting — gets deleted, not accelerated. Automating waste is the most expensive way to keep it.
Fixed logic, structured inputs, exact outcomes — validation checks, routing on clear fields, calculations. Cheaper, faster and more auditable than any model. Using an LLM where a rule suffices is buying non-determinism you don't need.
Judgment tasks with verifiable output: drafting responses, summarizing case files, extracting fields from documents, proposing classifications a human confirms. Human review is the containment; most near-term value lives here.
Multi-step work executed autonomously within bounds, with exception queues and escalation designed in (Module 8). High volume, tolerable error cost, measurable outcomes, reversible actions — and only after assist-tier evidence.
Steps that are judgment-heavy, high-stakes, low-volume or relationship-bearing stay human — classified deliberately, with a rationale, so "keep human" is a decision rather than a leftover. Score each candidate on value × feasibility × risk, and the output is a prioritized opportunity backlog with rationale per row — an artifact a steering committee can interrogate, not a slide of vibes.
AI helps run the scan itself: given the map, volumes and error data, a model proposes per-step classifications with reasoning. Treat its proposals like any Module 2 draft — useful, citable, and validated by humans who know which "obvious automation candidate" is actually a regulatory control wearing a boring costume.
For every AI-tier candidate, the triage heuristics: high-volume + verifiable-output + review-tolerant → automate (the cost of checking is low and the volume pays for the plumbing); judgment-heavy + high-stakes + irreversible → augment (AI drafts and enumerates, the human decides, because being wrong is expensive and accountability is non-delegable).
Then the decision that outranks both: redesign scope, decided before tool selection. The McKinsey anti-pattern — taught here by name because Raj's pilot died of it — is bolting an agent onto one step of a sequential-handoff process. The agent accelerates its step; the process still moves at the speed of its queues and handoffs; measured end-to-end improvement rounds to zero; the pilot is declared a failure and the tool takes the blame. High performers are ~3x more likely to fundamentally redesign the workflow — collapsing handoffs, reordering steps around what AI makes cheap, redrawing human roles around verification and exception judgment. The redesign question ("if drafting were free and instant, would this process still have nine steps?") is analysis, it is your job, and it must be answered before anyone signs a vendor demo's guest book.
When a sponsor's request arrives pre-shaped as "add an agent to step four," the analysis has been skipped, not done. Your first deliverable is the scan and the redesign-scope recommendation — delivered before tool selection, because after tool selection every finding will be bent to justify the purchase.
The business case is where AI initiatives earn funding, and where MIT's 95% is manufactured. The honest version has five load-bearing parts:
AI assists throughout — drafting the model, stress-testing assumptions, red-teaming the benefits logic — under Module 5's law: shown math, audited by a human, every assumption labeled.
Gartner's 60% figure has a corollary: the data-readiness assessment is the highest-ROI artifact in the AI-BA toolkit, because it costs a week and prevents a write-off. Per use case, five questions:
| Dimension | The question | What kills projects here |
|---|---|---|
| Existence | Does the data the AI needs actually exist, with history? | The "customer intent" field nobody ever captured |
| Quality | At what completeness and accuracy? (Module 5 profiling evidence attached) | 34% nulls in the critical field; three date formats; duplicate entities |
| Ownership | Who owns it, and will they provision access on your timeline? | Six-week access negotiations discovered in week seven |
| Legal usability | Is the intended purpose covered — consent basis, purpose limitation, DPDP/GDPR, client MSA, residency? | Data collected for billing, repurposed for AI training without a basis |
| Retrievability | Can it be fetched at runtime, at latency and cost the design tolerates? | The nightly-batch warehouse behind a real-time use case |
Findings become a readiness scorecard, and gaps become prerequisites with named owners — not footnotes. "Blocked pending consent-basis remediation, owner: DPO, est. 8 weeks" is an honest project plan; discovering the same fact in week seven of a funded pilot is Raj's scenario. Legal usability deserves special weight for this track's audience: under DPDP and GDPR, purpose limitation means data lawfully held for one purpose is not automatically usable for an AI feature — a question for the data-protection officer during analysis, not after procurement.
Three artifacts turn "should we do this?" from a meeting series into evidence.
The feasibility spike. Before any business case: one day, 20–30 real sampled cases (drawn from the tickets and exceptions you mined in Module 2 — including the ugly ones), a drafted prompt, a simple scoring rubric, run in a model workbench. No engineering dependency; analyst-runnable end to end. If quality isn't plausibly reachable on a favorable sample, the initiative dies at the cost of one day — the cheapest kill available, and the direct antidote to Gartner's abandoned-after-PoC statistic. If it looks promising, the spike's cases and rubric become the seed of the eval set.
Vendor evaluation with an eval harness. Module 6's claims-table discipline, upgraded for AI products where demos are maximally misleading — every demo shows the happy path a demo engineer rehearsed. Build a scenario set from your process's real cases (edge and exception cases included, weighted toward what actually hurts), run the identical set across candidates via promptfoo or a disciplined spreadsheet, and score quality, cost and latency against a pre-agreed rubric. The decision memo cites scores, not impressions. Demo-friendly candidates that fail four exception cases and hallucinate a policy clause once are found before the contract — which is the entire point.
The pilot charter with kill criteria. Before launch, in writing, sponsor-signed: the hypothesis ("agent-assisted triage reduces median handle time ≥25% at ≤2% critical-error rate"), the measurement method and duration, the counterfactual (what would the team have done anyway — the improvement a process gets from mere attention is real and must not be billed to the AI), and explicit kill/scale criteria. Kill criteria agreed before launch are a mercy: they let a failing pilot end with organizational dignity — "it did not meet the bar we all set" — instead of zombieing on because ending it would embarrass a sponsor.
Sequence the artifacts as a funnel and let each one earn the next: scan → spike → readiness → business case → charter. Most candidates should die in the first three stages, cheaply. An analyst whose funnel kills eight ideas for every one it advances is not being negative — they are the reason the advanced one works.
The AI-transformation market is drowning in demos and starving for analysis. Scan before selecting, decide redesign scope before tools, measure baselines before launch, check the data before the contract, spike before the business case, and set kill criteria before the kickoff. The analyst who makes AI initiatives die cheap or succeed measurably is the scarcest person in the room — and this module is that job description.