Picture a bridge built from both banks at once. The engineering is sound and the steel is good. But the two halves were never designed to meet, so the span ends in mid-air.
That is a lot of enterprise AI in 2026. The technology works and employees say it helps. But the value never reaches the other bank, where the P&L sits.
The real cost is not one stalled pilot. It is a growing AI budget where leadership cannot tell productive investment from expensive experimentation.
The gap is real. Read the numbers carefully.
McKinsey’s August 2026 State of AI survey of 1,719 respondents shows the split. Eighty percent report that AI has improved their individual productivity. Only 37% attribute any EBIT impact to AI, about the same as in 2025. AI high performers, who attribute at least 5% of EBIT to AI and call its value “significant,” are still about 6%.
These are respondent-reported figures, not audited financials. Still, the direction is hard to ignore.
MIT NANDA’s The GenAI Divide (July 2025) went further, reporting that “95% of organizations are getting zero return” on generative AI. Treat that number as directional. The authors call their findings preliminary, drawn from 153 surveyed leaders, 52 organizational interviews and 300+ public initiatives, and based on interviews rather than official company reporting. It is not proof that 95% of AI spend was audited and found worthless. It is a signal that pilots rarely reach measurable P&L impact.
Time saved is not money saved
Say AI cuts contract review from 60 minutes to 30. The business case claims a 50% productivity gain. The CFO asks: did external legal spend fall? Did we avoid a hire? Did we close more contracts, or reduce exposure?
If none of that happened, you created capacity, not yet financial value. Both matter. They are not the same thing.
| Value type | Example | How to prove it |
|---|---|---|
| Realized financial | Lower external spend, avoided hiring, better margin | Finance-validated results |
| Operational capacity | More cases with the same team, smaller backlog | Throughput, quality and cycle time |
| Strategic | Better customer experience, reduced risk, faster innovation | Agreed outcome indicators and a realization horizon |
AI does not create financial value because people finish tasks faster. Value is realized when that capacity changes the economics of the business.
Five pillars, one foundation
When we diagnose a stalled pilot, model capability is rarely the whole story. Accuracy, reliability and cost can block production. More often, though, the design around the model decides whether its capability becomes useful. We look for five pillars.
- Data the workflow can use. Pilots often run on a clean, hand-picked extract. Test on representative, permissioned, production-like data before trusting the result, without handing the pilot unrestricted production access.
- A documented process. You cannot automate a process that lives in three people’s heads. As Microsoft puts it, “Adding agents to a broken process still leaves a broken process.”
- An accountable owner. Pilots get sponsored. Production systems get owned: outcome, budget, risk and the decision to switch it off.
- Adoption, not rollout. Microsoft found that “a tool licensed to over 200,000 people does not change how the work gets done.” Redesigning the work did.
- Measurement. McKinsey’s high performers are twice as likely to report defined processes for measuring AI impact. Most programs skip this one.
The foundation: production architecture
A production AI system is not a pilot with better data. It is an operational system that carries business-critical work. Six questions tell a CIO or CTO whether it is ready:
| Area | Executive question |
|---|---|
| Data and context | Can it reach the right information, under the right permissions? |
| Integration | Can it act inside existing applications, transactions and approvals? |
| Security and governance | Can we control what agents see, decide and do? |
| Evaluation and reliability | How do we know results stay acceptable? |
| Operational economics | What does one successfully completed transaction cost? |
| Lifecycle | Who monitors it, handles failures and manages model changes? |
Economics is now a live issue. In the same McKinsey survey, about one in five respondents say AI operating costs, including tokens, constrained their AI use, and 32% say they skipped at least one software purchase because they could build it with agentic coding tools.
Not every workload needs the largest frontier model. Smaller models, rules engines and deterministic automation often do the job. The right architecture is the simplest one that reliably delivers the outcome at acceptable total cost and risk.
Measurement: from capability to the P&L
Most AI programs report what is easy to count: licenses, active users, prompts. That shows activity, not value. Value has to be traced across four stages, which together form the span between the two banks:
| Stage | What to measure |
|---|---|
| 1. Technical performance | Quality, reliability, latency, cost per AI operation |
| 2. Workflow performance | Cycle time, throughput, exception rate, cost per completed case |
| 3. Business outcome | Capacity, revenue opportunity, quality, reduced risk |
| 4. Financial realization | Finance-validated savings and profitable growth |
Microsoft shows what stage 2 looks like. Based on its internal analysis, average cycle time in selected cloud supply-chain workflows fell from about ten business days to under 2.5 across five monthly planning cycles. That is strong operational evidence. It is not, on its own, proof of EBIT impact.
Three disciplines get you from stage 2 to stage 4:
- Baseline or control. Measure the process before AI touches it, or against a comparable team that does not use it.
- Fully loaded cost. Include integration, inference, licenses, human review and ongoing operations.
- Realizable benefit. Agree with Finance up front how benefits are counted, so the result is accepted later.
A hypothetical: same pilot, two scorecards
Hypothetical composite, for illustration only. Two business units pilot the same AI tool for vendor contract summaries.
Unit A reports 400 active users and 12,000 summaries in a quarter. The steering committee is pleased, but nobody can say what changed for the business.
Unit B agreed a baseline with Finance first: review time, share escalated to legal, and cost per review. At quarter end it can show what moved, by how much, and what it cost.
Unit A can demonstrate adoption. Unit B can determine whether the investment deserves to scale. That is the difference between reporting activity and making an informed capital allocation decision.
The funding gate: an AI value-realization contract
Before any initiative gets production funding, business ownership, technology delivery and Finance should agree on seven points:
| Requirement | Decision criterion |
|---|---|
| Business outcome | What business condition are we changing? |
| Economic hypothesis | How will that change create value? |
| Baseline and target | How much improvement, against what comparison? |
| Accountable owner | Who achieves and sustains the outcome? |
| Production readiness | Are data, integration, security, reliability and support sufficient? |
| Financial validation | Has Finance agreed how benefits and costs are calculated? |
| Decision gate | What evidence triggers scale, redesign or stop? |
This quarter, run every live pilot through that table. Gaps are your first finding. Re-scope or stop anything without an owner or an agreed metric. Fund the foundations of the few that pass.
The question that finishes the bridge
Every executive, from CEO and CFO to COO, CIO and CTO, can ask the same thing of any AI investment:
Can we trace this investment from its technical capability, through a measurable workflow improvement, to a business outcome leadership can substantiate?
Business and technology leaders cannot answer that alone. The bridge only meets in the middle when both banks are designed together.
Redrock Systems architects have designed enterprise systems since 2008. Our Baseline, the AI Cost and Architecture Diagnostic, is a fixed-scope assessment over three to four weeks. We map where AI is used, what it costs, what it returns and what to do first, and your leadership team leaves with a prioritized, board-ready 90-day plan.
Want to know which of your AI pilots can carry a business case?
Sources
- McKinsey, “The state of AI in 2026: On the road to ROI” (25 Aug 2026; online survey, 1,719 participants, fielded May–June 2026): https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- MIT NANDA, The GenAI Divide: State of AI in Business 2025 (PDF, Jul 2025; preliminary findings): https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
- Microsoft, “What we’ve learned from Microsoft’s own AI transformation” (17 Sep 2026): https://blogs.microsoft.com/blog/2026/09/17/what-weve-learned-from-microsofts-own-ai-transformation/