Three years into the generative AI boom, the question in most boardrooms has shifted from “what can AI do?” to “where is the money?” Adoption is no longer the problem. McKinsey’s State of AI in 2026 survey, fielded in May and June 2026 with 1,719 participants, found that roughly 90% of respondents regularly use AI in at least one business function. Yet only 37% attribute any EBIT impact to AI, a share McKinsey describes as essentially unchanged from the year before. Usage is rising; measurable value is not keeping pace.
This gap is not a technology failure. It is a strategy, measurement and operating-model failure, and it is fixable. This article explains why AI value stalls between pilot and profit and loss, and lays out a practical playbook that business and IT leaders can use in 2026 to manage AI as a portfolio of investments with owners, baselines and exit criteria rather than as a collection of experiments.
The Value Gap in Numbers: Why 2026 Is the Reckoning Year
Several independent sources point in the same direction. McKinsey’s 2026 survey reports that about 44% of respondents say AI is scaling across their enterprise, up from 38% a year earlier, and that about 40% of large organizations (over $1 billion in revenue) are scaling AI agents. But the share of “AI high performers”, which McKinsey defines as companies attributing at least 5% of EBIT to AI, stayed at about 6%, the same as in 2025.
The MIT NANDA initiative’s 2025 report, The GenAI Divide, was blunter. As reported by The Register, the authors concluded that 95% of the organizations they studied had seen no measurable return from their generative AI efforts, based on 52 structured interviews, an analysis of more than 300 public AI initiatives and a survey of 153 business professionals. The report’s methodology has been debated, but its central diagnosis resonated with many executives: tools that cannot remember context, adapt to a workflow or learn from feedback rarely survive contact with real operations.
Gartner, meanwhile, predicted in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. Cost is now a visible constraint too: in the McKinsey 2026 survey, about one in five respondents said AI operating costs, including token consumption, had limited their use of AI.
Put together, the message for 2026 budgets is clear. Finance leaders are no longer willing to fund AI on faith, and the organizations that can show a credible line from model to margin will win the next round of investment.
Why AI Value Stalls Between Pilot and P&L
When we review stalled AI programs, the root causes are remarkably consistent. Very few are about model quality. Most fall into one of five patterns.
1. Use cases chosen for demo appeal, not economics
Pilots are often selected because they are visible and easy to show: a chatbot on the intranet, a content generator for marketing. The MIT NANDA authors noted that roughly half of the AI budgets they examined flowed to sales and marketing, while some of the clearest returns came from less glamorous back-office work such as eliminating outsourced processing, reducing agency spend and automating risk checks.
2. No baseline, so no proof
A surprising number of initiatives start without a measured “before” state. If nobody recorded the cycle time, error rate or cost per transaction before the pilot, any improvement is anecdotal and easy for a skeptical CFO to dismiss.
3. Productivity that never becomes capacity or cash
Individual productivity gains are real. In McKinsey’s survey, 80% of respondents said AI had improved their personal productivity. But saved minutes scattered across thousands of employees do not appear on the income statement unless someone deliberately redeploys that capacity, reduces external spend or grows throughput with the same headcount.
4. Bolting AI onto unchanged workflows
This is the strongest differentiator in the data. Nearly three-quarters of McKinsey’s 2026 high performers say they have fundamentally redesigned workflows around AI, compared with about one-quarter of other respondents. Gartner gives the same advice for agentic AI: redesign workflows from the ground up rather than layering agents onto legacy processes.
5. Invisible run costs
A pilot that costs a few thousand dollars a month can become a six- or seven-figure line item when rolled out to every user, especially with agentic patterns that make many model calls per task. Without unit economics, value cases quietly turn negative after scale-up.
A Value Realization Framework: From Idea to Audited Benefit
The fix is to treat every AI initiative as an investment with a lifecycle, the same way a well-run organization treats capital projects. We use a five-stage framework that business and IT leaders can adopt without new tooling.
- Frame the value hypothesis. State the business outcome in one sentence, the metric that proves it, the owner who will be accountable for it, and the mechanism by which the benefit reaches the P&L (lower cost, higher revenue, lower risk or working capital).
- Baseline before you build. Measure the current process for at least four to six weeks: volumes, cycle times, rework, error rates, cost per unit and customer outcomes. Agree the measurement method with finance up front.
- Prove in production conditions. Run the pilot on real data with real users, ideally with a control group. Track adoption and quality alongside efficiency, because a fast answer that users do not trust has no value.
- Scale with redesign. Change the workflow, roles, controls and incentives around the AI capability, not just the interface. This is where most value is created and where most programs underinvest.
- Harvest and audit. Convert gains into budget decisions: reduced contractor spend, avoided hires, faster cash collection or measurable revenue uplift. Have finance validate the benefit at 6 and 12 months.
Each stage should have an explicit gate. If a use case cannot name an owner and a P&L mechanism, it does not get to build. If it cannot beat its baseline in production conditions, it does not get to scale. This sounds bureaucratic, but in practice it frees resources: killing weak ideas early is how strong ones get funded.
AI value is not discovered at the end of a project; it is designed at the start. If a use case has no owner, no baseline and no route to the P&L, it is an experiment, not an investment.
Building the AI Portfolio: Balance, Not a Backlog
Most organizations have more AI ideas than they can fund. The goal is not a long backlog but a balanced portfolio, managed quarterly, that mixes quick returns with strategic bets.
Three horizons for AI investment
- Efficiency plays (0–9 months): document processing, service desk triage, reconciliation, code assistance and knowledge retrieval for well-defined tasks. Lower risk, measurable quickly, often self-funding.
- Workflow transformation (6–18 months): end-to-end redesign of a process such as order-to-cash, claims handling, procurement or credit underwriting, often combining predictive models, generative AI and automation with new roles and controls.
- Growth and new offerings (12–36 months): AI-enabled products, new service models or pricing approaches. McKinsey notes that high performers pursue growth and innovation objectives alongside efficiency rather than treating AI purely as a cost lever.
Scoring use cases
A simple scoring model is enough to bring discipline to prioritization. Score each candidate from 1 to 5 on the following criteria and weight them to reflect strategy:
- Value at stake, expressed in annual P&L terms and validated by finance
- Data readiness: availability, quality, access rights and lineage of the data required
- Workflow fit: how clearly the task is defined, how measurable the output is, and how much redesign is needed
- Risk profile: regulatory exposure, customer impact, reversibility of errors and need for human oversight
- Run cost: expected model, infrastructure and support costs at full scale
- Reusability: whether components such as data products, retrieval pipelines or guardrails can serve other use cases
Reusability is often the most underrated criterion. The second and third use cases on a shared data and platform foundation are dramatically cheaper than the first, which changes the economics of the whole portfolio.
Metrics That Finance Will Accept
The language of AI teams (accuracy, latency, tokens) is not the language of the CFO. A credible value case connects three layers of metrics.
- Model and system metrics: accuracy, groundedness, latency, uptime and cost per call. Necessary for engineering but insufficient for investment decisions.
- Process metrics: cycle time, first-time-right rate, straight-through processing, backlog, escalations and adoption. These show whether the workflow actually changed.
- Financial metrics: cost per transaction, external spend avoided, revenue per employee, days sales outstanding, loss rates and margin. These are what appear in the P&L.
Two practices make these metrics trustworthy. First, define unit economics early: the fully loaded cost to process one invoice, resolve one ticket or underwrite one loan, both before and after AI. Unit economics expose hidden run costs and make scale decisions straightforward. Second, separate “hard” benefits that change budgets from “soft” benefits such as time saved or satisfaction. Both matter, but only hard benefits should be counted toward payback.
A note on token economics
With agentic systems, a single business task can trigger dozens of model calls, tool invocations and retries. Track cost per completed task, not cost per call, and set budgets per workflow. Model routing (using smaller, cheaper models for simple steps and reserving frontier models for complex reasoning), caching and retrieval tuning can change unit costs materially without hurting quality, so they belong in the value plan rather than being left as an afterthought.
The Operating Model Behind Scaled AI Value
Frameworks and metrics only work if the organization is set up to act on them. Four operating-model elements consistently separate scaled programs from stalled ones.
Business ownership with a shared platform
Each AI product should have a business owner accountable for the outcome, supported by a central platform team that provides data products, model access, evaluation tooling, security and guardrails. Fully centralized models create bottlenecks; fully decentralized ones duplicate spend and risk.
A value office, not just a center of excellence
Many organizations created AI centers of excellence focused on enablement. In 2026 the more useful structure is a small value office, often jointly run by the CIO or CDAO and finance, that owns the portfolio, enforces stage gates, tracks benefits and reports to the executive committee.
Partners held to outcomes
The MIT NANDA authors observed that organizations getting results tended to treat AI vendors more like business process outsourcing partners than software suppliers: demanding deep customization and holding them to business metrics. The same logic applies to delivery partners. Contracts that pay for outcomes, or at least tie milestones to measured process improvements, align incentives far better than time-and-materials pilots.
Governance proportional to risk
Governance should accelerate low-risk use cases and apply deeper scrutiny where customers, regulators or safety are involved. A tiered approach, with lightweight review for internal productivity tools and full model risk and compliance review for high-impact decisions, prevents governance from becoming the reason nothing ships.
A Worked Example: Accounts Payable Automation and the Pitfalls Around It
Consider a generic mid-sized manufacturer processing about 40,000 supplier invoices a month across several ERP instances. Today, a shared-services team plus an outsourced provider handles matching, exceptions and supplier queries. Leadership wants “AI in finance” and has three proposals on the table: a finance chatbot, a forecasting model and an AI-assisted invoice exception workflow.
Applying the framework changes the conversation. The chatbot has no clear P&L mechanism beyond time saved, so it is parked as a soft-benefit tool. The forecasting model is promising but depends on data that is not yet reliable, so it moves to the second horizon behind a data quality workstream. The invoice exception workflow has a clear owner (the head of shared services), a measurable baseline and a direct route to the P&L through reduced outsourced processing cost and fewer late-payment penalties.
The team baselines six weeks of data: cost per invoice, touchless processing rate, exception cycle time and penalty fees. The pilot combines document extraction, a retrieval-based assistant that proposes resolutions using contract terms and past cases, and rules that route only uncertain items to humans. Crucially, the scale phase redesigns the workflow: exception queues are restructured by root cause, supplier master data issues are fixed at source, and the outsourced contract is renegotiated in stages as volumes fall.
At the 12-month audit, finance validates benefits in three categories: reduced outsourced processing spend, lower penalty fees and a modest working-capital gain from capturing more early-payment discounts. The run cost, including model usage and support, is netted off. The result is not spectacular, but it is real, audited and repeatable, and the shared extraction and retrieval components cut the cost of the next finance use case.
Common Pitfalls to Avoid
- Counting time saved as cash without a plan to redeploy capacity or reduce spend
- Scaling before the workflow is redesigned, which locks in old inefficiencies at higher cost
- Ignoring adoption metrics; low usage is the earliest warning that value will not materialize
- Letting each business unit build its own stack, multiplying vendor contracts, security reviews and run costs
- Treating governance as a final approval step rather than designing controls into the workflow from the start
- Failing to retire pilots; zombie projects consume budget, attention and goodwill
Outlook: What Changes Through 2027
Three shifts will shape AI value realization over the next 18 months. First, agentic AI will push organizations from task-level assistance to process-level automation, which raises both the potential value and the importance of workflow design and controls. Gartner expects that by 2028, 15% of day-to-day work decisions will be made autonomously through agentic AI.
Second, AI spending will increasingly be governed like any other investment class, with portfolio reviews, benefit tracking and post-implementation audits becoming standard practice. Organizations that build this discipline now will be able to fund the next wave from the returns of the current one.
Third, cost optimization will become a core competency. As usage grows, the gap between organizations that manage model routing, caching, data efficiency and vendor terms and those that do not will show up directly in margins.
Conclusion
The AI value gap is real, but it is not mysterious. Organizations that capture value choose use cases for their economics, baseline before they build, redesign workflows rather than decorate them, measure unit costs and hold owners and partners to outcomes. None of this requires a breakthrough model; it requires management discipline applied consistently. For business and IT leaders heading into 2027 planning, the most valuable AI investment may be the one that makes every other AI investment measurable.
How Hanoi Technology Can Help
Hanoi Technology helps business and IT leaders build AI portfolios, value cases and delivery roadmaps that connect models to measurable P&L outcomes.