Question 1

Is enterprise AI delivering ROI in 2026?

Accepted Answer

Locally yes, enterprise-wide rarely. McKinsey's State of AI 2025 (November 2025, n=1,993) found that only 39% of respondents attribute any EBIT impact to AI and only ~6% qualify as AI high performers attributing more than 5% of EBIT and significant value. MIT NANDA's GenAI Divide (July 2025) estimates that 95% of generative-AI pilots are failing to deliver measurable returns despite an estimated $30-40 billion of enterprise spend. Business-unit wins are real (Morgan Stanley adviser adoption, GitHub/Accenture developer telemetry, focused customer-service deployments), but firm-wide P&L impact still requires workflow redesign, measurement and governance most organizations have not yet built.

Question 2

What does an enterprise LLM workload actually cost?

Accepted Answer

Less than executives assume on the meter, more than they assume on total cost of ownership. For an internal knowledge assistant doing 55,000 Q&A turns per month (~110M input + 33M output tokens), public list prices captured in May 2026 range from roughly $231 on GPT-5.4 mini to $1,650 on Claude 3.5 Sonnet via Bedrock. A 300-seat Microsoft 365 Copilot Business deployment at the annual list price already runs about $5,400 per month, before implementation. The expensive parts usually sit outside the base token price: data connectors, permissions, evaluation harnesses, security reviews, observability, change management and human review.

Question 3

Why do most generative-AI pilots stall?

Accepted Answer

Gartner forecasts that at least 30% of generative-AI projects will be abandoned after proof of concept by end-2025; MIT NANDA puts the share returning zero P&L impact at around 95%. The reasons are operational, not technical: poor data quality, inadequate risk controls, escalating costs, unclear business value, missing baselines, and no plan for redeploying time saved. McKinsey 2025 confirms it: workflow redesign is the strongest leadership predictor of EBIT impact, and AI high performers are nearly 3x more likely than peers to have fundamentally redesigned a workflow. Pilot success says almost nothing about production economics.

Question 4

What is the right operating model for enterprise AI?

Accepted Answer

A central control plane (data governance, model access, observability, policy, evaluation) combined with partly decentralised business-unit experimentation. Use a portfolio approach to use-case selection: start with high-frequency tasks where the workflow is measurable and the error surface is controlled. Run AI FinOps from day one: output caps, prompt caching, batch processing, routing simple requests to smaller models, separate budgets for model spend, search and human review. Tier human oversight by risk. Klarna's May 2025 reversal is the cautionary case study on going too far on automation alone.

Question 5

Should we build, buy, or partner for enterprise AI?

Accepted Answer

All three, sequenced by use-case maturity and switching cost. Buy seat-based copilots for productivity work where vendor-managed grounding is acceptable. Build only where you need proprietary workflow logic, data residency, or model-routing economics that off-the-shelf platforms cannot give you. Partner for edge capabilities (vertical models, evaluation, governance tooling) where the market is moving faster than internal cycles. Be aware that vendor lock-in has moved from prompt portability to operational portability: PTUs, GSUs, cache semantics and routing tiers compound over time. MIT NANDA's 2025 data suggests purchased copilots currently outperform custom internal builds on adoption and ROI.

Question 6

How do I measure AI impact credibly to my board?

Accepted Answer

Measure the business process, not the activity. Resolution time, merge rate, cycle time, conversion, gross margin, claims accuracy and working-capital velocity are the right units of account. Use holdouts, pre/post baselines or randomised roll-outs where possible. The GitHub/Accenture study is the strongest public template: it combines telemetry, user surveys and an explicit experimental design. Treat vendor-sponsored ROI multiples as ambition-setting, not as audited returns; treat MIT NANDA's 95%-fail and McKinsey's 6%-high-performer figures as the realistic outer bounds of where most programmes currently sit.

Enterprise Impact of the AI Fever: What Almost Four Years of GenAI Have Actually Bought

Key takeaways