MIT’s Project NANDA found that 95 percent of enterprise generative AI pilots delivered no measurable impact on the P&L. Most of those systems worked. What failed was the measurement, because nothing had been set up to follow the money from the model to the result.
MIT Project NANDA, The GenAI Divide, 2025
Value has a long way to travel, from the first AI investment through to shareholder value, and it can leak at every step along the way. Most organizations measure the investment, the capability and the share price, and very little in between. Models deployed, use cases launched, staff enabled, prompts written and hours supposedly saved all describe activity, and none of them measures value.
Howard Rubin, banking AI value research, 2026
A few banks now publish the value AI has produced as a financial figure. The figures are not like-for-like, and the DBS figure counts analytics as well as AI.
S$1bn
DBS disclosed about S$1bn of economic value from AI and analytics in 2025.
€35m
Santander disclosed €35m of AI business value in the first quarter of 2026, and is aiming for more than €200m across the year.
£50m
Lloyds disclosed about £50m of generative AI value in 2025, and is aiming for more than £100m of additional P&L benefit in 2026.
Bank disclosures, as relayed by Howard Rubin, 2026
Cost per AI-assisted outcome only means something next to the cost of the same outcome handled by a person. Cost per inference only means something when it is fully loaded with the tokens, the platform underneath, the data and the people who review the exceptions. Divide a token bill by the number of outcomes and you get a cloud unit cost, which is a different thing from a return.
AI has moved from the CIO’s vocabulary into the language of the CFO and the investor, and proving what it returns now depends on the platform you measure it on.
A dashboard shows data, but a board needs the story behind it: what happened, why it happened, what it means and what happens next. Turning the data into that story is the step a dashboard cannot take.
They report AI spend by service tag, which is useful, but they cannot answer the question that matters most. If a new model raises inference cost by 30 percent and halves the cloud compute around it, is the business better off overall? That is a question about the whole estate, and a token meter cannot see the estate.
Benefits realization checks whether a project delivered what it promised. The case is written once, before the money is spent, and rarely reconciled afterwards, so it never reaches the continuing, portfolio-wide view a board needs: a technology P&L, return by program and the balance of run, grow and transform.
“Connecting AI investment to business outcomes is a hard problem. We’re going to take years to get there.”
Marco Meinardi, Gartner, FinOpsX 2026
The analysts are open about the fact that this is still unsolved. The organizations that start measuring properly now will be the ones with a record to show when the board asks.
In MOTEF, the MagicOrange Technology Economics Framework, this is the Prove step: return by program, read from live data and explained in the language of whoever is reading it.
Model, platform, data, labor and the AI built into the software you already pay for, well beyond the token line.
AI Economics, with Time Capture and Resource Planning for the people
Per inference, per AI-assisted transaction, per resolved ticket and per completed workflow.
Enterprise Allocation Modeling, AI Economics
The cost of the same work done by a person, so a productivity claim has something to be measured against.
Time Capture and Resource Planning, Enterprise Allocation Modeling
Run, grow and transform as a live mix, with each AI program tagged to the change it claims to make.
Budgeting and Forecasting, IT Financial Management
Plan against actual, told in the CFO’s terms for the CFO and in the board’s terms for the board, written from the governed model rather than a cost report.
Analytics and AI Studio, included with every capability
The model holds what was spent, on what, by whom and what that means for the enterprise. Whether a business process actually changed, and what the change produced, is recorded in the business and outside the finance system. Once that outcome data is supplied, MagicOrange makes the chain from cost to result auditable. It does not create the outcome data itself, and nobody can honestly claim to have measured the whole chain yet.
Instead of “we spent €4m on AI”, a figure like “€0.12 per successful customer interaction, down 18 percent this quarter”.
Illustrative. Not a customer result.
No customer has published an AI return figure yet.
Cost per inference and cost per successful outcome, fully loaded across tokens, platform, data and labor. It also shows return by program against plan and the current mix of run, grow and transform, with a written explanation generated from the governed model for whoever is reading it.
