Stop spreading AI thin: pick one or two domains and go end to end
Most leadership teams say they're "all in" on AI, yet many are stuck in pilot purgatory. A widely cited study this summer reported that 95% of organizations saw no measurable return from generative AI. The gap is clear: experimentation is easy; scaling value is not.
NTT Data CEO Abhijit Dubey put it bluntly at the 2025 Fortune Global Forum: companies are trying to do too much. The better bet is to pick one or two domains with outsized economics and take them end to end. Think underwriting in insurance or supply chains in manufacturing-areas where a single percentage point shift has real P&L consequences.
What "end to end" actually looks like
Focus isn't just narrowing scope. It means owning the full journey-from data plumbing and model deployment to workflow change and governance-until results show up in financials. That's where most pilots fall apart.
- Choose domains tied directly to revenue, margin, cost-to-serve, or working capital-not vanity use cases.
- Map the full process (customer and operator journeys), then redesign steps where AI augments or automates decisions.
- Build the data foundation first: reliable sources, clear ownership, and a unified model of the entities you operate on.
- Ship as products, not projects: APIs, MLOps, monitoring, SLAs, and a clear release cadence.
- Set measurable targets (e.g., loss ratio, forecast accuracy, cycle time) and instrument everything from day one.
- Design for adoption: training, new roles, incentives, and fast support loops for frontline teams.
How operators are actually scaling
FedEx has focused AI on three lanes: internal operations, customer experience, and new value levers for customers like better demand forecasting and fewer returns, said regional president Kami Viswanathan. The lesson: prioritization drives scale; without it, AI stays a slide, not a system.
Vortexa's CEO, Fabio Kuhn, emphasized human supervision to keep hallucinations from steering decisions. Explainability matters just as much as speed and quality-leaders need to know why a model recommends a move before they sign off. If you need a reference framework for risk, the NIST AI Risk Management Framework is a solid starting point.
Noosheen Hashemi of January AI underscored the point: large models can be confident and wrong. Their team keeps a physician in the loop and tracks hallucination to under 1%. In U.S. health care, the bigger blocker is data silos across payers, providers, labs, and patients-interoperability standards like HL7 FHIR help, but leadership will is what moves the needle.
Governance that prevents costly surprises
- Human-in-the-loop for high-impact decisions with clear approval thresholds.
- Use retrieval and citations so outputs are grounded in your source of truth.
- Enforce prompt, data, and access controls; log every decision and dataset version.
- Explainability by default: rationale, features used, confidence, and known gaps.
- Production monitoring: drift, bias, hallucination, latency, and incident playbooks.
From pilot to P&L impact
- Baseline with control groups before rollout; publish the measurement plan upfront.
- Bundle small wins into one productized platform per domain to cut integration tax.
- Define "kill rules" for pilots that don't hit targets by a set date-free up capacity fast.
- Lock unit economics: track cost per decision, per query, or per workflow minute saved.
- Simplify the stack: fewer vendors, clearer ownership, cleaner security posture.
Executive checklist
- Pick 1-2 domains with disproportionate value and commit for 12-18 months.
- Appoint a single owner with budget across data, product, and change management.
- Fund the data layer first; no data, no scale.
- Set governance and human oversight before go-live, not after an incident.
- Tie incentives to the domain KPI that matters (e.g., claims leakage, inventory turns).
- Publish a simple scoreboard weekly: adoption, quality, ROI, incidents.
The pattern is consistent: pick a lane, build the plumbing, wire in the guardrails, and keep humans where the stakes are high. Do that twice, prove value, then reuse the patterns elsewhere. That's how AI moves from deck to P&L.
If your team needs a practical path to upskill by role, explore curated options here: AI courses by job.
Your membership also unlocks:
