When the Sellers Write the Scorecard: Who Defines Your AI ROI
As boards demand proof that AI spending pays off, the firms that profit from that spending have formed a body to define the metric itself. Nothing is published yet, but read whatever it ships as a supplier document, and keep a number they never touched.
Every measurement standard has an author, and the author's incentives leak into the reading. So when a new body forms to define AI ROI, the first question isn't whether its eventual yardstick looks sophisticated. It's who calibrated it, and what happens to their revenue when the number goes up.
In 2026 the Linux Foundation launched the Tokenomics Foundation, whose stated mission is to define the economics and return on investment (ROI) of AI value. Its site leads with the line "driving AI value and ROI across all AI costs", and it defines its own coinage plainly: "AI token economics, or AI tokenomics, is the emerging practice of managing the production, consumption, and value of AI to generate business outcomes." Among the founding members are Accenture, IBM, JP Morgan Chase, Oracle and SAP. That is where the evidence stops. No specification has been published yet; the name, that stated mission and that membership are what there is to go on. They are enough to predict the shape of what's coming. Read the roster against the mission and a neutral standard already looks like a supplier document, before a line of methodology is written.
Why does it matter who defines AI ROI?
Fix the unit of account and you settle the argument before it starts. To its credit, the definition doesn't stop at consumption: it ends "to generate business outcomes", which names the right destination and is the strongest reply to the charge that the frame is consumption-denominated. Grant it fully. Then look at what actually does the counting. The practice is called tokenomics, the definition puts "production, consumption, and value" ahead of that outcomes clause, and a token is the unit vendors bill for. Outcomes are named as the goal; consumption is what the frame is built to measure. If the canonical number rises with consumption, a metric that climbs as you spend reads as success, and a buyer who adopts per-token throughput as the proxy for value has quietly agreed to be measured on the axis that flatters the seller.
The buyer's real question sits on a different axis. What did it cost to complete an outcome a customer or a P&L actually recognises: a resolved support case, a correct underwriting decision, a document processed to an accepted standard? Those numbers can fall while token consumption soars, and climb while consumption looks efficient. They measure different things, and no amount of consortium branding collapses the gap.
How should a board read a vendor-backed AI ROI standard?
Standards bodies borrow credibility by association. Put enough recognised names on a masthead and the output reads as impartial by default, because the members plainly are serious institutions. That is exactly where the care is needed: the credibility of the members and the independence of the metric are separate properties, and only one of them is on display so far. Any specification this membership authors would amount to consensus among interested parties, a useful input that falls well short of an audit.
The incentive is old, even if this particular body is new. A firm that bills by a unit has every reason to favour a measurement frame denominated in that unit, and a buyer who accepts the frame inherits the bias baked into it. This is an argument about incentives, not a charge against any member's competence: serious institutions can still author a metric that happens to serve their revenue. Token-denominated AI value fits that shape, and nothing about a distinguished membership changes the direction the incentive points.
The cost lands when the metric spreads. Once a vendor-defined measure becomes the language of board reporting, procurement, budgeting and vendor comparison all inherit its bias. Bake per-token value into the reporting stack and you've bought switching costs you never negotiated: every future comparison then runs on the incumbent's axis. Tuning a fleet of models against consumption-flattering metrics can raise spend while the outcomes meant to justify it stay flat, and nobody in the reporting chain notices, because the chain was built to watch the wrong number.
What would change the verdict?
Here's the evidence that would flip this position. If a value specification were governed by parties with no financial stake in AI consumption, published its methodology in full, and defined its headline metric as cost per completed outcome rather than per unit consumed, it would earn the benefit of the doubt. If independent auditors, not member firms, computed and attested the figures, better still. Those are testable conditions, and they matter precisely because none of them can be checked yet: there is no document to hold them against. Apply them to whatever the Tokenomics Foundation eventually publishes, or to any framework you're handed. A standard authored by its beneficiaries tends to fail the first condition on its own.
All of this argues for one move: own the ruler that measures the spending. Getting there starts before procurement, with a clear view of which outcomes the technology is supposed to move, work that belongs to an honest readiness assessment rather than a vendor deck. It continues with keeping a human hand on the metrics that decide budgets, and with a technical strategy that treats the measurement layer as your own asset.
The defensible posture keeps any future standard in its place: one scorecard among several, never the only one. Keep one outcome-based measure no consortium designed for you: cost per solved task on a private, held-out workload that mirrors your real business, owned and computed outside anyone's framework. That number is cheap to maintain and hard to game, precisely because the sellers never touched its definition, and it's the number that should decide renewals.
There's an asymmetry here that rarely gets priced. The party that defines the winning metric wins every negotiation that references it, quietly, for years. That's worth far more to a supplier than any single contract, which is why the definition is being written now, by the firms with the most to gain from it. The cheapest defensive move a board can make is to write down its own number first.
Questions people ask
What is a better metric than tokens consumed for AI ROI?
Cost per completed business outcome: the fully loaded cost of getting one unit of recognised work done to an accepted standard, such as a resolved case, an approved decision or a processed document. Unlike token throughput, it can be tracked against a real profit-and-loss line and can't be inflated by simply consuming more.
Is a vendor-backed AI standard automatically untrustworthy?
Not automatically, but it should be read as an interested party's proposal rather than an audited yardstick. Test it against three conditions: independent governance, a fully published methodology, and a headline metric denominated in outcomes rather than units the authors sell. Frameworks that pass all three deserve more trust than those that pass none.
How can a buyer avoid being locked into a supplier's measurement frame?
Keep a benchmark you own: a held-out slice of your real workload, priced by what it costs to finish one unit of recognised work, computed by you rather than read off a vendor's dashboard. Use that figure as the reference point for renewals and vendor comparisons, so switching costs stay yours to control.
Related
- Retention Bonuses Don't Reduce Key Person Risk. Doctrine Does.
- Bot Traffic Is Splitting Your Ad Budget, and a Human Signed It Off
- Your Cloud Provider Is Financing Its Own AI Revenue. Do the Financial Due Diligence
- Data & Strategy
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.