AI Investments Are Growing 35% a Year. Boards Want ROI. The Measurement Frameworks Barely Exist.

There is a question being asked in boardrooms right now that almost nobody has a clean answer to: what is the return on investment for generative AI? I have heard versions of it from practitioners, read it in analyst reports, and it is right at the center of what Gartner identifies as the major strategic shift CIOs are navigating in 2026: moving from running pilots to demonstrating ROI (https://www.gartner.com/en/newsroom/press-releases/2026-04-07-gartner-forecasts-worldwide-it-spending-to-grow-9-8-percent-in-2026). AI investments are growing at more than 35% year over year. The spending is accelerating. The ability to measure what the spending produces is not keeping pace.

McKinsey's State of AI 2025 fills in the picture (https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai). 88% of organizations use AI in some form. Only 7% have fully scaled it across their operations. The technology has spread widely but shallowly. Most organizations are using GenAI somewhere. Almost none have embedded it deeply enough to produce measurable organizational-level outcomes. Boards are noticing this, which is why the pressure for ROI is intensifying at exactly the moment adoption metrics are being celebrated. "We deployed to 80% of our employees" is not an ROI figure. It is an adoption figure. Boards with growing AI line items in the budget are starting to ask for something harder.

The measurement problem has a few distinct layers. The first is diffuseness. GenAI productivity gains tend to be distributed across many individual tasks rather than concentrated in one measurable output. A developer uses an AI assistant to complete routine code faster. An analyst uses a chatbot to draft a first-cut memo. A customer service representative uses a suggestion engine to pull relevant precedents more quickly. Each of these is a real gain. None of them shows up cleanly in a revenue or cost figure at the organizational level. To aggregate them, you would need to measure baseline task completion times, compare them to post-deployment times, control for the fact that motivated early adopters are not a random sample, and then somehow aggregate across thousands of heterogeneous tasks. Most organizations do not have that measurement infrastructure. Building it retroactively is expensive. So instead, organizations report adoption rates, which are easy to measure, and treat them as proxies for value, which they are not.

The second layer is attribution. Even where productivity gains are measurable, isolating the AI contribution is hard. If a sales team using AI-assisted CRM tools closes more deals, is that because of the AI, the additional training that came with the rollout, a change in the incentive structure, or an improvement in market conditions? Isolating one variable in a live organizational environment requires an experimental or quasi-experimental design that organizations rarely build into technology rollouts. Without that design, you end up with correlational evidence at best, which boards rightly treat with skepticism when the question is whether to approve the next round of AI spending.

The third layer is the tension between cost reduction and value creation as ROI framings. Cost reduction is easier to measure: if AI replaces a vendor contract, or reduces headcount through attrition, or cuts the cost of a specific workflow by a calculable amount, the number is defensible. This is why AI ROI cases most often lead with cost reduction. But cost reduction as the primary framing produces a narrow deployment strategy. It means deploying AI where headcount can be reduced, not necessarily where value can be created. Value creation is harder to measure but often larger: AI that improves the quality of decisions, expands service capacity, or enables a genuinely new product produces value that does not show up in a cost comparison. The organizations that measure only cost reduction will optimize for a smaller version of their current operations, not a genuinely different one.

This is where the IS theory most relevant to this moment is not a new framework. It is Melville, Kraemer, and Gurbaxani's (2004) integrative IT value model. They established that IT creates value through a chain: IT capabilities enable improved business processes, improved processes drive organizational performance, and whether the firm captures that performance improvement depends on the competitive environment and its complementary organizational resources. GenAI is not different from prior technology waves in this fundamental respect. The challenge is that the intermediate step, process improvement, is exactly what gets skipped when organizations layer GenAI on top of existing workflows without redesigning those workflows. They get efficiency gains within an unchanged process, not the process transformation that produces larger performance gains. And efficiency gains within unchanged processes are both harder to measure and smaller than the gains from genuine process redesign.

What also concerns me here is the DeLone and McLean IS Success Model (2003). They argued that system quality, information quality, and service quality combine to drive use, which then produces net benefits. The ROI measurement problem with GenAI maps onto their model in an uncomfortable way: organizations are measuring use (adoption rates) and calling it net benefits. The actual causal chain requires measuring whether the system is producing high-quality outputs, whether those outputs improve decisions or processes, and whether those improved decisions and processes produce organizational-level net benefits. That full chain is almost never measured. The field stops at adoption.

As an IS researcher, what raises a genuine research question for me is why the IT value measurement literature, which is several decades deep and quite well developed, is not being applied to AI investments. We know from Brynjolfsson and Hitt (1996) that IT investment produces value through process change, not through the investment itself. We know from the DeLone and McLean model that use is not the same as net benefit. We know from Melville et al. that complementary organizational resources are what convert technology capability into performance. All of this was established before large language models existed. None of it is being systematically applied to how organizations are measuring their GenAI spending.

A CIO who wanted to actually answer the board's ROI question would start with measurement design before deployment, not after. Before launching a GenAI initiative, define what a successful outcome looks like at the process level. Not "X% of employees will use the tool" but "this specific process will produce this specific outcome faster, at higher quality, or at lower cost, and we will measure it against this baseline." Build the measurement infrastructure before deployment. Run a controlled rollout if possible. Specify in advance what you will do with the data.

That is a harder project to sell than "we are deploying AI to the whole organization." It requires knowing what success looks like before the technology arrives. But it is the only way to answer the board's question with something other than adoption metrics and optimism.

---
claims_checked:
- "AI investments growing 35%+ YoY for CIOs": "https://www.gartner.com/en/newsroom/press-releases/2026-04-07-gartner-forecasts-worldwide-it-spending-to-grow-9-8-percent-in-2026"
- "CIO shift from pilots to ROI as major 2026 strategic change": "https://www.gartner.com/en/newsroom/press-releases/2026-04-07-gartner-forecasts-worldwide-it-spending-to-grow-9-8-percent-in-2026"
- "88% of organizations use AI; only 7% fully scaled": "https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai"
claims_unverified:
- "GenAI use 79% in 2025 (up from 33% in 2023): referenced in prior version from McKinsey; exact figures not re-verified from a freshly fetched URL for this rewrite"
- "Analysis of diffuseness, attribution, and cost vs. value framing: my own theoretical framing, grounded in IT value literature but not directly cited from a single paper"
- "Prediction that organizations measuring only adoption will undercount GenAI ROI: opinion, clearly framed as my read"
sources_used:
- "https://www.gartner.com/en/newsroom/press-releases/2026-04-07-gartner-forecasts-worldwide-it-spending-to-grow-9-8-percent-in-2026"
- "https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai"
word_count: 1045