DigiUsher Briefing May 21, 2026 Lawrence Cole 11 min read

Board-Level AI ROI: Why $600B in Investment Is Delivering Single-Digit Returns — and What Fixes It

Only 7% of CFOs see high ROI from AI despite $270B in enterprise spend forecast for 2026. Here's the governance framework boards are demanding — and why cost attribution is the missing layer.

Enterprise AI investment is delivering single-digit ROI because organizations are measuring AI spend as IT cost rather than as capital allocation tied to P&L outcomes. Only 7% of CFOs report high AI ROI despite $270 billion in forecast enterprise AI spend for 2026. The root cause is missing cost attribution — AI infrastructure spend cannot be traced to revenue, margin, or productivity outcomes without a governance layer above the cloud billing layer.

AI capital allocation AI board reporting AI cost forecasting

Enterprises have deployed an estimated $600 billion more in AI capital than they’ve generated in measurable enterprise revenue. That’s not a technology problem. It’s a board-level governance failure, and it’s accelerating as AI budgets grow and the questions boards are prepared to ask get sharper.

Only 7% of CFOs report seeing high ROI from AI in their finance functions, and a December 2025 survey of 200 US finance chiefs found just 14% have seen a clear, measurable financial impact from AI investment so far — despite two-thirds expecting impact within two years. Three-quarters of CFOs are raising their technology budget for 2026, with nearly half increasing it by 10% or more. That acceleration makes the measurement problem more urgent, not less. Boards that accepted “build and learn” rationales for AI pilots in 2024 are applying real capital allocation discipline in 2026, and CFOs who can’t connect AI infrastructure spend to a P&L line in financial rather than anecdotal terms are losing credibility in the room.

The accountability gap, and why boards now own it

The maths behind enterprise AI investment in 2026 is stark. Worldwide AI spending is projected to top $2 trillion this year. Enterprise AI application software spend alone is projected to reach $270 billion, nearly triple 2024 levels, while the largest tech companies are projecting over $500 billion in AI infrastructure investment. The aggregate return on that capital, expressed in terms boards actually recognise, doesn’t match the investment thesis — the gap between capital deployed and revenue generated has grown to roughly $600 billion. Boards have stopped counting pilots and started counting dollars, and 61% of senior business leaders now feel more pressure to prove AI ROI than they did a year ago.

That pressure isn’t evenly distributed. A May 2025 Gartner survey of 506 CIOs found 72% of organisations are breaking even or losing money on their AI investments. As a result, the CFO’s role in AI decisions has expanded — Deloitte’s Finance Trends 2026 survey found 57% of finance executives now count themselves among the leaders actively driving AI strategy, a real shift from the traditional CFO role as budget gatekeeper. That shift has consequences for governance: when the CFO is a co-owner of AI strategy rather than just a gatekeeper, they need financial infrastructure giving them a defensible view of what AI is actually producing. A cloud cost dashboard showing aggregate GPU spend doesn’t provide that. A board-level AI ROI framework needs attribution at the unit of production.

The board question itself has changed structurally. Eighteen months ago it was “are we investing enough in AI.” Today it’s “what is each AI investment returning, and how does that compare to what the same capital could generate elsewhere.” That second question needs an AI-Specific P&L — a discrete financial ledger within an organisation’s management accounts attributing AI infrastructure cost (model inference tokens, GPU compute, orchestration overhead, data platform charges) to the revenue lines, products, or customer segments those AI workloads actually serve. Unlike general cloud cost allocation, it captures the full token economics of each AI application and connects gross margin impact to a measurable business outcome. Google Cloud’s research on top-performing AI enterprises found they generate $10.30 in value for every dollar of AI investment, against a broader average of $3.70 — a 2.8x gap explained not by model sophistication, but by attribution discipline: the capacity to measure, at workload and application level, what each capability actually costs and returns.

Why activity metrics fail boards

The dominant AI measurement architecture inside most large enterprises was built for a different audience. Queries processed per day, tokens consumed per month, model uptime, latency percentiles, adoption rate — these are operational metrics, telling platform engineering whether the infrastructure is functioning. They tell a CFO or a board nothing about capital efficiency. A Gartner survey of 782 infrastructure and operations leaders found only 28% of AI use cases fully succeed and meet ROI expectations, with 20% failing outright — and the primary cause of failure wasn’t technical deficiency, it was the absence of business outcome integration. The leaders whose AI initiatives succeeded attributed that primarily to integrating AI into existing workflows and securing genuine business executive support.

The problem runs deeper than metric choice. Research from early 2026 found enterprises typically discover more than 150 AI applications in active use once they conduct a proper inventory — against an expectation of roughly 30. The ungoverned majority of that estate has no attribution infrastructure at all; spend that proliferated organically through developer tooling, SaaS subscriptions, and marketplace purchases is generating token cost with no measurement of what those costs are producing. That spend can’t demonstrate its value to a board regardless of what it’s actually returning, because the financial infrastructure to prove it simply doesn’t exist.

Three measurement failures show up consistently. Cost without attribution: AI infrastructure cost lands in cloud bills broken out by account and service, but not by the product, customer segment, or process it serves — a CFO presenting a $4M quarterly AI infrastructure cost can’t say what fraction is generating customer-facing revenue and what fraction is idle GPU capacity or an experimental workload with no production ROI case. Activity metrics instead of outcome metrics: token consumption and API call volume get reported as evidence of programme momentum, but none of it connects to gross margin, so a board member asking “what is this generating in revenue” gets no meaningful answer from an activity report. And no baseline before deployment: without a defined performance baseline set before an AI capability goes live, calculating ROI afterward is close to mathematically impossible — you can’t measure progress from an undefined starting point.

The AI-Specific P&L imperative

The term “AI-Specific P&L” is appearing in board papers and audit committee agendas that didn’t contain it twelve months ago, and the reason is structural: AI cost behaves fundamentally differently from the technology cost boards have governed before, and existing financial frameworks don’t contain it well. Traditional infrastructure cost is provisioned-resource cost — a server runs at a known rate, a licence costs a fixed annual fee, the depreciation schedule is predictable, and a budget variance is explainable in a normal quarterly review. AI cost is token-consumption cost, determined by design decisions, model versions, prompt architectures, and agentic workflows that change continuously, often with no capital approval gate at all — even a minor change in a prompt, model version, or agent workflow can spike GPU-hours or token consumption 100x overnight. Only 15% of enterprises can forecast AI cost within ±10% accuracy, and nearly one in four miss their AI cost forecast by more than 50%. A 50% variance on any other material cost category would trigger an audit committee response immediately; the reason it hasn’t for AI spend is that boards don’t yet have the reporting infrastructure to see it happening in real time.

What board-ready AI financial reporting actually contains: revenue attribution broken into AI-enabled revenue by category (customer-facing AI products, AI-augmented sales and service, quantified internal productivity gains); the fully-loaded AI cost of production (inference cost by provider — Azure OpenAI, Bedrock, Vertex/Gemini — plus GPU compute for training and idle time, orchestration and data platform cost, and governance/observability overhead); the resulting AI gross margin contribution, tracked against both the prior quarter and the board’s target; and attribution coverage itself — the percentage of AI spend that’s genuinely attributed versus unattributed or Shadow AI spend, with the unattributed portion flagged and quantified as an explicit governance risk rather than left invisible. That structure is what actually answers a board’s questions. It isn’t produced by a cloud billing dashboard — it needs governance architecture operating at the token and inference level, attributing cost to the business units and products consuming it, and surfacing the unattributed estate as a quantified risk rather than a blind spot.

What the 20% who prove AI revenue impact do differently

Deloitte’s 2026 State of AI survey, across 3,235 leaders in 24 countries, found 74% of enterprises want AI to drive revenue growth; only 20% have achieved it. That 20% aren’t using more sophisticated models — they’re using more sophisticated measurement, across three disciplines.

Attribution before deployment, not after. The enterprises capturing demonstrable AI ROI built their attribution architecture before any capability went into production, rather than retrospectively assigning cost to spend already incurred — attribution after the fact is estimation; attribution built into the deployment architecture is measurement. Every AI application gets a cost centre owner, a revenue attribution rule, and a unit economics target before its first production token. McKinsey’s 2025 research found organisations seeing meaningful AI returns were twice as likely to have redesigned end-to-end workflows before selecting a model — attribution architecture is part of that redesign, not an afterthought bolted on.

Token-level governance rather than account-level reporting. Cloud account-level reporting shows total spend. Token-level governance shows spend per application, per model, per business process, and whether that granular spend is generating a proportionate return — the difference between a utility bill and an actual business case. Token density (the ratio of tokens processed to output generated) identifies model waste at the application level, and cost-per-inference lets a CFO assess the margin impact of a specific AI capability rather than treating AI as one opaque shared infrastructure line.

AI unit economics as a genuine capital allocation tool. This is the discipline of calculating cost, revenue contribution, and gross margin impact at the level of a single AI-driven transaction — one inference call, one AI-assisted customer interaction, one automated decision. Enterprises with mature unit-economics governance can actually answer what boards are asking: what does each capability cost to operate at scale, what does it generate, and does the unit-level return justify continued or expanded investment. That’s the same portfolio discipline a board applies to any other capital programme — scaling what has strong unit economics, governing what has acceptable economics, retiring what’s negative — and it’s what actually converts AI from a technology experiment into a governed investment.

The board won’t wait for better models to prove the case. They’re waiting for the financial infrastructure that attributes what current models are costing and returning, at the workload level, in the language of capital allocation — and that’s a 2026 governance requirement, not a future capability.

See what an AI-Specific P&L looks like against your own AI estate. Book a 30-minute session and we’ll map real token-level attribution against what you’re actually spending. Request a demo →

Frequently asked questions

What is board-level AI ROI, and why does it matter in 2026? It’s an organisation’s ability to demonstrate, at board level, that its AI investment is generating measurable financial return — expressed in the language of capital allocation, gross margin, and revenue attribution, rather than activity metrics like queries processed or tokens consumed. With enterprise AI application software spend projected to nearly triple to $270 billion in 2026, AI has become a material capital allocation decision subject to the same governance discipline as any other board-level investment.

Why do most enterprise AI investments fail to show board-visible ROI? Primarily a measurement architecture problem rather than a technology one — AI cost disperses across cloud bills, data platform charges, and SaaS model fees with no attribution to the specific business outcome those workloads serve. Deloitte’s 2026 State of AI survey found 74% of enterprises want AI to drive revenue growth, but only 20% have achieved it. Shadow AI compounds this: enterprises typically discover 150-plus AI applications in active use against an expectation of roughly 30, with most of that estate carrying no attribution infrastructure at all.

What metrics should a CFO actually present to the board on AI? Four categories: AI unit economics (cost per inference, cost per AI-assisted transaction), AI gross margin contribution (net margin impact after full cost attribution), AI cost-to-revenue ratio (AI infrastructure spend as a share of AI-attributable revenue), and attribution coverage (the share of total AI spend assigned to a responsible cost centre or product line, with the unattributed portion quantified explicitly as a risk rather than hidden).

What is an AI-Specific P&L, and why do boards actually need one? A discrete financial ledger attributing AI infrastructure cost — tokens, GPU compute, orchestration, data platform charges — to the revenue lines and business units those workloads serve. Boards need one because AI cost behaves structurally differently from traditional infrastructure cost: a single prompt or model version change can spike token consumption 100x overnight, only 15% of enterprises can forecast AI cost within ±10%, and 84% report AI-driven gross margin erosion exceeding 6%. No other cost category with that variance profile would be governed without a dedicated financial ledger.

How does AI cost governance differ from traditional cloud FinOps? Traditional cloud FinOps governs provisioned-resource cost running at predictable rates. AI cost governance addresses a structurally different challenge — cost determined by token consumption, model selection, and autonomous agent behaviour, all of which can change dramatically with no provisioning approval gate at all. 98% of FinOps teams now manage AI spend, up from 63% in 2025, and the mechanisms this actually requires — token budget caps, per-application attribution, agentic kill-switches, inference-level anomaly detection — largely don’t exist in cloud FinOps tooling built before the generative AI era.