The Rationale That Is Running Out of Runway
Enterprise AI investment has been justified for the past three years on a rationale that combines strategic positioning, competitive necessity, and optionality: we need to invest in AI capability because the organisations that do not will fall behind those that do. This rationale does not require return measurement. It requires belief in the strategic importance of AI and willingness to invest ahead of demonstrated return.
This rationale was appropriate in 2023 and 2024, when the pace of AI capability change made it genuinely difficult to predict which investments would produce returns and which would not, and when the competitive landscape was sufficiently uncertain that strategic positioning investment was defensible. It is less defensible in 2026. Three years of enterprise AI investment has produced enough evidence — across enough organisations, in enough use cases, at enough scales — to have a more honest conversation about what AI investment actually returns.
The conversation that is needed is not “AI is not worth investing in.” The return evidence across well-designed AI programmes is positive. The conversation that is needed is “these are the specific AI investments that produce demonstrated returns, these are the ones that have not produced demonstrable returns despite significant investment, and this is how we reallocate between them.” That is the portfolio management conversation that enterprise boards apply to every other significant investment category, and it is the conversation that AI investment now deserves.
What the Return Evidence Actually Shows
The AI investment return evidence across European enterprises that have been serious AI investors since 2022 or 2023 shows a pattern that is more nuanced than either the AI optimists or the AI sceptics acknowledge.
The use cases that are producing demonstrable, measurable return in 2026 are narrower than the AI investment portfolio of most enterprises. Knowledge worker productivity augmentation in well-defined task categories — document drafting, research summarisation, code review, data analysis — is producing measured productivity improvement of twenty to forty percent for the specific task types the AI is applied to, in the organisations that have redesigned workflows around AI assistance rather than layering AI tools onto unmodified workflows. The return is real, measurable, and at the investment levels required for well-designed productivity augmentation programmes, strongly positive.
Customer-facing AI in well-scoped, high-volume, routine interaction categories — customer service routing and triage, product recommendation, search and discovery — is producing measurable improvement in customer experience metrics and in the operational efficiency of the human teams that handle the non-routine interactions that AI routes to them. Again, the return is measurable and positive where the implementation design is well-executed.
The use cases that are not producing demonstrable return at the investment levels they have received are the more ambitious AI applications: autonomous multi-step reasoning for complex knowledge work, AI-assisted strategic decision support at executive level, and AI systems that were designed to reduce headcount in knowledge work roles without redesigning the work around the reduced headcount. These use cases are absorbing significant investment and producing either no measurable return or returns that are significantly below the projections that justified the investment.
The Measurement Infrastructure That Most Enterprises Have Not Built
The reason most enterprises cannot have the honest return conversation with their boards is not that the return is absent. It is that the measurement infrastructure to demonstrate the return — or its absence — was never built.
The standard AI investment governance model approves the investment based on projected returns, deploys the AI capability, and then measures adoption (how many users are using the tool?) and operational metrics (system availability, inference latency, error rate) rather than business outcomes (what changed in the business as a result of the AI capability?). The adoption and operational metrics are necessary. They are not sufficient to demonstrate return.
Business outcome measurement for AI investment requires: a baseline measurement of the outcome before AI deployment, a measurement mechanism that tracks the specific outcome the AI is designed to improve, and attribution logic that connects the outcome change to the AI investment rather than to other concurrent changes. These three requirements are individually straightforward and collectively rarely implemented.
The baseline is missed because the measurement framework is defined after deployment rather than before. The outcome measurement is missed because the organisation measures AI system metrics rather than business outcomes. The attribution is missed because it is genuinely hard in an environment with many concurrent changes, and because implementing it requires more analytical investment than most AI programmes allocate.
The Board Conversation That Should Have Happened Sooner
The board conversation that the return evidence enables is not a retrospective on past AI investment performance, though that is a legitimate component. It is a prospective investment allocation conversation based on what the return evidence shows.
The conversation has three components. First, an honest portfolio review: across the AI investments approved in the past three years, what does the return evidence show? Which use cases have produced the projected returns? Which have produced returns below projection? Which have produced no demonstrable return? The portfolio review requires the measurement infrastructure described above; where that infrastructure is absent, the honest answer is that the return is unknown, which is itself useful information.
Second, an investment reallocation proposal based on the portfolio review. The AI use cases with demonstrated positive returns deserve increased investment to scale them further. The use cases with negative or unknown returns deserve either a defined measurement programme that will produce return evidence within a defined timeframe, or reallocation of their budget to use cases with better evidence.
Third, a return measurement commitment for new AI investments: no new AI investment approved without a defined return measurement framework that includes baseline, measurement mechanism, and attribution approach, with the measurement results reported at a defined cadence.
This commitment changes the governance quality of future AI investment decisions. The board that approves AI investments with defined return measurement is in a better position to allocate capital optimally across the AI portfolio than the board that continues to approve investments on strategic positioning rationale alone.
The Technology Leader Who Has This Conversation
The technology leader who brings the honest AI investment return conversation to their board is taking a reputational risk: the review of past AI investment performance may not show the returns that the previous investment cases projected. This is uncomfortable.
It is also the right thing to do, and it produces a qualitatively better board relationship than the alternative. The board that is managed with optimistic AI investment narratives that do not reflect the actual return evidence will eventually lose confidence in the technology leader’s assessments. The board that is brought an honest portfolio review, with clear-eyed identification of what is working and what is not, and a credible investment reallocation proposal, gains confidence in the technology leader’s judgment.
The honest conversation is harder in the short term and better for the long term. Technology leaders who have it are building the board relationship that makes future technology investment decisions easier, not harder.
