← AI Adoption Cycle 🕐 12 min read
AI Adoption Cycle

AI Use Case Selection: A Practical Guide for Fortune 500 Decision-Makers

Brandon Sneider · April 2026

Most enterprise AI selection processes begin with a technology: "We're evaluating Copilot / Gemini Enterprise / Claude for Work — which workflows should we apply it to?" This inverts the sequence that

See also (wiki): wiki/ai-use-case-selection.md, wiki/ai-delivery-pods.md, wiki/workflow-redesign.md, wiki/data-readiness.md, wiki/ai-first-use-case-selection.md

Executive Summary

  • The median Fortune 500 AI program produces zero measurable business impact. NBER (Feb 2026, n=5,867 executives across 4 countries) finds 89% of firms report zero productivity impact from AI over three years. The differentiating factor is not technology access, budget size, or model quality — it is use case selection discipline and the domain concentration that follows.
  • The “use-case implementation race” is the primary organizational anti-pattern. McKinsey Rewired (Lamarre, Smaje, Levin, 2nd ed., 2024, Ch. 3, p.49) documents the pattern precisely: enterprises launch 50–200 individual AI pilots, each too small to move a P&L line item, each owned by a different team without shared infrastructure. The result is high AI activity and flat financial performance.
  • The domain concentration remedy is empirically supported. BCG AI at Work 2025 (n=10,600, 32 industries) finds organizations concentrating AI investment on 1–3 business domains produce measurable results; those spreading across 100 use cases rarely do. McKinsey State of AI 2025 (n=1,993) finds only 6% of companies generate meaningful EBIT impact from AI — the differentiator is rigorous use case selection for the specific operational context.
  • The agentic era adds a modality selection layer. Rewired Ch. 5’s four-level automation framework (individual augmentation → task automation → agentic workflows → agentic systems) provides a defensible, vendor-neutral basis for choosing the right automation approach for each workflow step. Not every step needs an LLM agent; selecting the wrong modality adds cost and hallucination risk without adding value.
  • Portfolio governance is as important as initial selection. The selection decision is not a launch-time judgment — it is a recurring quarterly governance process that requires sunset authority as well as approval authority.

Part 1 — Why Most Selection Processes Fail

The technology-first inversion

Most enterprise AI selection processes begin with a technology: “We’re evaluating Copilot / Gemini Enterprise / Claude for Work — which workflows should we apply it to?” This inverts the sequence that produces results.

The correct sequence is:

  1. Identify which business domains have the largest gap between current performance and the economic ceiling
  2. Map the workflows within those domains where that gap is most concentrated
  3. Determine what change in those workflows would move the measurable metric
  4. Select the AI modality (and then the vendor) that can deliver that specific workflow change at production quality

Organizations that invert this sequence — technology-first — consistently build AI for workflows where the baseline was never measured, where the data requires significant cleanup before the AI can run, and where success criteria are defined post-hoc. The result is the “interesting demo, no ROI” pattern.

McKinsey State of AI 2025 (n=1,993) documents the outcome: companies that matched AI tools to pre-defined business workflow gaps were 4.5x more likely to generate meaningful EBIT impact than companies that identified workflows after selecting tools. The sequencing is causal.

The use case vs. domain distinction

Rewired (Ch. 3, p.49) makes the definitional distinction that most Fortune 500 selection frameworks miss:

  • A use case is a specific AI-enabled task: contract summarization, invoice classification, email draft assistance, meeting transcription. Each use case, working correctly, saves hours per week per user. Summed across an enterprise, these savings may never appear in a P&L line item because the freed hours recycle into other activity rather than producing a measurable output change.

  • A domain is an end-to-end business process: revenue cycle, supply planning, claims processing, customer acquisition. A domain is large enough that reimagining it end-to-end moves a named line item. A domain can contain 5–20 use cases; a single use case cannot become a domain.

The test for domain scope: “If every AI deployment in this program executes exactly as designed, which line on the P&L changes, and by how much, within 12 months?” If the answer is vague (“it frees up time”), the scope is use-case-level. If the answer is specific (“we reduce claims cycle time from 22 to 11 days, releasing $14M in working capital per quarter”), the scope is domain-level.

Part 2 — The Selection Framework

The three-gate screening criteria

Before any use case enters active development, it should clear three gates sequentially. These gates are sequential — failing Gate 1 disqualifies without proceeding to Gate 2.

Gate 1 — Business economics gate

Is there a measurable current-state cost or performance gap that this deployment addresses?

Question Disqualifying answer
What does the current process cost per unit/cycle/transaction? “We don’t have that number.”
What would a 30% improvement in that metric be worth annually? Less than $500K at mid-market / $5M at Fortune 500 scale
Which P&L line or operational KPI changes when this works? “It’s hard to attribute directly.”
Is there an existing owner who is accountable for that metric? “IT owns it” / no named business leader

If the answers to these four questions are not satisfactory, the use case does not advance. The economics do not exist, or the organization does not know what they are — both disqualify.

Gate 2 — Data readiness gate

Does the required data exist in usable form today, without a multi-month data infrastructure project?

The minimal viable data condition: the data exists in structured digital form, is accessible from a single system or a small number of integrations, and the quality is sufficient to train/test an evaluation set. Rewired (Ch. 25, p.391) states the principle at full scale: “No data architecture, no AI advantage.” The Gate 2 version is narrower: does enough data exist to start, even if the full data platform build is still in progress?

Use cases that fail Gate 2 are not rejected — they are deferred to a later sequencing slot after the data readiness work is complete. See wiki/data-readiness.md for the full data assessment framework.

Gate 3 — Output quality verification gate

Can a human in the workflow verify the AI output before it takes a consequential action?

This is not a permanent constraint — it is a first-deployment sequencing principle. AI output that replaces human judgment in an unverifiable step creates governance, legal, and political exposure simultaneously. The verification question has two sub-parts: (a) does the output quality bar matter (are errors costly or low-stakes?), and (b) can a subject matter expert verify an output sample in under 60 seconds per item?

Use cases that fail Gate 3 — where errors are costly and verification requires deep expertise or significant time — are not first deployments. They may be third or fourth deployments, when the organization has built evaluation infrastructure and demonstrated system reliability in adjacent workflows.

The four-level modality selection filter (Rewired Ch. 5)

After passing the three gates, a use case requires a modality decision: which automation approach is appropriate for this specific workflow step?

Rewired (Ch. 5, p.82, Exhibit 5.1) provides the framework:

Level Description Best for Wrong for
1 — Individual augmentation AI assists; human decides and acts High-judgment, unstructured, variable-quality tasks Structured, high-volume, deterministic tasks
2 — Task automation Specific bounded steps automated end-to-end Structured, repetitive, well-defined steps within a human-supervised workflow Tasks requiring dynamic exception handling
3 — Agentic workflows Multi-step sequences with tool use, routing, exception handling Dynamic, unstructured, edge-case-heavy workflows with clear success criteria Workflows where errors are hard to detect
4 — Agentic systems Networks of specialized agents coordinating at domain level Complex, multi-party, multi-system domain operations at scale Early-stage deployments without mature evaluation infrastructure

The selection principle Rewired makes explicit: “Choosing the right tool for the job is not a sign of technological conservatism; it’s a mark of maturity” (Ch. 5, p.92). A workflow that benefits from Level 2 automation is worse, not better, if it is implemented at Level 3 — the added complexity adds latency, cost, and failure modes without returning value.

For Fortune 500 program governance, the modality selection decision should require explicit sign-off from the domain owner alongside engineering — it is a business architecture decision, not a technical preference.

Part 3 — Domain Portfolio Scoring Matrix

Use this matrix to rank candidate domains before committing development resources. Assign 1–5 for each dimension; multiply by the weighting; sum for total score.

Dimension Weight 5 3 1
Economic gap size 25% Annual value >$10M clearly quantified $2–10M estimated <$2M or unquantified
Data readiness 20% Structured data in one system today 2–3 systems, known quality issues Major cleanup required
Business ownership clarity 20% Named executive, >25% time allocated Senior manager, intent to allocate IT-owned or unclear
Modality fit 15% Clear fit to Level 1–2 automation; low hallucination risk Level 3 with strong evaluation plan Level 4 or unresolved modality
Reuse potential 10% Data and eval artifacts reusable across 3+ subsequent use cases Reusable for 1–2 additional One-off, minimal reuse
Change readiness 10% Department has active AI champion; manager training complete Champion nominated, training planned No champion; change plan unclear

Threshold decisions:

  • Score ≥ 4.0: Priority 1 domain — launch next pod cycle
  • Score 3.0–3.9: Priority 2 domain — invest in prerequisites (data readiness, change capacity) before launching
  • Score < 3.0: Defer — the prerequisites are too far from ready to justify development investment now

Part 4 — Portfolio Governance

The approval function is a concentration function

An AI steering committee that approves every request that comes in is not performing governance — it is performing administration. The approval function exists to enforce the domain concentration principle: at any given time, the enterprise’s AI development capacity should be concentrated in a small number of domains with sufficient investment to reach production, not spread across a large portfolio of underfunded pilots.

The practical governance rule: the number of active AI development pods should not exceed the number of qualified domain owners who can commit 20%+ of their time. A Fortune 500 with 100 business units and 50 approved AI pilots has approved 50 pilots and committed to zero of them at the resource level required to succeed.

Portfolio ceilings by organizational capacity:

Enterprise AI maturity (MIT CISR Stage) Maximum concurrent active domains Rationale
Stage 1 — Experimenting 1–2 Change capacity and data infrastructure do not yet support more
Stage 2 — Piloting 3–5 Infrastructure from Domain 1 begins to reduce Domain 2–3 setup cost
Stage 3 — Scaled AI ways of working 6–12 Reuse infrastructure, established pods, and governance cadence can sustain higher parallelism

The sunset authority

Portfolio governance that only has approval authority is incomplete. Every domain in the portfolio should have an annual review date at which the domain owner and program sponsor answer: is this deployment still producing Tier 3 business impact? If not, is there a specific, time-bound remediation plan?

Deployments without measurable Tier 3 impact after 12 months should be sunset unless there is a specific technical explanation (e.g., data pipeline was delayed and is now complete) with a defined 90-day test window. The budget and pod resources freed by sunsetting underperforming deployments fund the concentration that produces results.

Part 5 — Industry-Specific Domain Priorities

Evidence-backed high-priority domains by industry. These are starting points for Gate 1 scoring — the organization’s specific economic gap data overrides generic industry patterns.

Financial services

  • Claims and underwriting automation — Capgemini World Cloud Report for Financial Services 2026 (n=1,100): AI-enabled underwriting reduces decision cycle from 5–7 days to same-day for standard-risk policies; 30–40% reduction in adjudication cost per claim
  • Fraud detection and prevention — Mastercard deployment evidence (Tech Unpacked, April 2026): real-time transaction pattern analysis; 40% improvement in fraud detection rate with no increase in false positive rate
  • Revenue cycle optimization — documented as the highest-ROI domain in mid-market financial services across BCG and Bain deployment evidence

Healthcare

  • Ambient clinical documentation — Menlo Ventures State of AI in Healthcare 2025: 60% reduction in documentation time for physicians using ambient documentation; 2–4 week time to measurable impact; the only healthcare AI domain with consistent, short-cycle ROI evidence
  • Prior authorization processing — structural workflow redesign candidate; approval cycle time reduction of 40–60% documented in multiple health system deployments

Manufacturing and supply chain

  • Predictive maintenance — Tech-Stack 2025 manufacturing AI benchmarks: unplanned downtime reduction of 20–35% in mature deployments; 6–18 month time to full ROI
  • Demand planning and inventory optimization — Carrefour/SymphonyAI deployment: documented reduction in excess inventory holding costs; 6–12 month time to measurable P&L impact

Professional services

  • Contract review and analysis — Thomson Reuters Future of Professionals 2025 (n=2,275): 62% of legal professionals report time savings on contract review; 30–60 day time to measurable throughput improvement
  • Research and due diligence synthesis — consistent first-win across law, accounting, and consulting; lowest data complexity, highest output-verification feasibility

Part 6 — Reuse Architecture: Building the Compound Advantage

The selection decision for Use Case 2 onward is different from the selection decision for Use Case 1. By the time the enterprise reaches Domain 2, it should have:

  • A production-grade data pipeline for the Domain 1 data sources
  • An evaluation framework with ground-truth test sets for Domain 1 workflow outputs
  • A change management template library from Domain 1’s adoption program
  • A governance artifact library (acceptable use policy, risk classification, escalation paths)

The Domain 2 selection criterion that most program governance frameworks miss: degree of infrastructure overlap with Domain 1. A Domain 2 that draws on the same data sources as Domain 1 costs 30–50% less to stand up. A Domain 2 that requires entirely different data infrastructure, evaluation criteria, and change management approaches provides no reuse benefit and should be treated as a first deployment economically.

Rewired (Ch. 31, p.469) states the principle: “The best use case is the reuse case.” BCG (2025) puts numbers on the outcome: top-quartile AI performers redeploy the outputs of high-performing workflows as inputs to adjacent ones, and the median gap between first and fifth deployment cost (as a percentage of Domain 1 build cost) is 62% for organizations with reuse infrastructure versus 95% for those without.

Actionable Next Steps

For the CIO before the next portfolio review:

  1. Run the Domain Portfolio Scoring Matrix against the current active pilot list. Any pilot scoring below 3.0 should be either remediated or sunset before new pilots are approved.
  2. Confirm named domain owner with 20%+ time commitment for every Priority 1 domain. If the domain owner is IT or a technology leader, reassign to the business unit head.
  3. Audit the current portfolio for the “use-case implementation race” pattern: more than 20 active pilots without a shared data pipeline across at least 3 of them is a concentration failure.

For the program sponsor presenting to the CFO:

  • Lead with the Gate 1 economics for every active domain — current-state cost per unit, projected improvement, named P&L line item, 12-month target.
  • Do not present activity metrics (pilots launched, tools deployed, users onboarded) as evidence of value. CFOs will correctly discount them.

For the domain owner standing up a new pod:

  • The first 30 days before any engineering investment: validate the Gate 1 economics with the CFO or their team. Get a signed-off baseline. The signed-off baseline is the only evidence that will stand up at the 12-month ROI review.
  • wiki/ai-use-case-selection.md — concept page; domain vs. use case; four-level modality selection; reuse model; common failure modes
  • wiki/ai-delivery-pods.md — the operating unit that owns domain-level deployment after selection
  • wiki/workflow-redesign.md — the workflow architecture work that determines whether a selected use case reaches the 71% productivity gain or the 30% tool-augmentation baseline
  • wiki/data-readiness.md — Gate 2 data assessment; the full data architecture framework for domain-level deployment
  • wiki/ai-roi-evidence.md — the evidence base for why 89% of programs produce zero impact and the top quartile does not; NBER, BCG, MIT CISR