Economic ambition needs an evidential basis
An AI business case makes a claim about the future. Delegating authority to an AI system makes that claim consequential: resources are committed, decisions are taken and other people bear the results. The business case stops being a forecast and becomes a basis for action.
Causum’s position is that enterprises should connect these commitments through evidence showing what was authorized, what actually improved and why the resulting value claim deserves reliance.
EY’s 2026 white paper, Can AI Pay for Itself?, broadens the economic analysis beyond token expenditure and distinguishes private cost savings from economy-wide growth. Its illustrations estimate the output growth or labor-cost reductions needed to support a projected AI cost base, explicitly identifying the arithmetic as illustrative rather than predictive.[1] That candor is useful. It invites the next question: what must an enterprise establish before it treats anticipated returns as a reason to expand autonomy?
Our response brings that economic question together with EY’s public governance findings. A credible investment case needs an ongoing process for accepting, challenging and revising its claims as deployment produces evidence.
Justified value — beneficial outcomes that can be substantiated within a stated scope, with legitimate authority and visible consequences for affected parties.
A required return does not establish a worthwhile investment
Working backward from a projected expenditure identifies the performance necessary to justify it, under stated assumptions. It does not establish that the performance is attainable, that customers will pay for the additional output, or that this is the best use of the enterprise’s resources.
An enterprise should therefore examine alternatives before accepting a deployment target. A strong business case explains why the proposed approach is preferable under the organization’s actual constraints — and identifies what evidence would change that conclusion.
Supplier investment creates commercial pressure to earn revenue, but it does not guarantee that customers will accept the prices required to recover every investment. Competition, utilization and changing capabilities can alter who absorbs the shortfall. Our inference is that an enterprise should treat supplier cost-recovery assumptions as a scenario to test, not an obligation to expand its own AI consumption.
Leadership should also distinguish sunk expenditure from future commitments. Having funded a pilot may justify evaluating what was learned; it does not by itself justify scaling it. The next decision concerns the expected incremental benefit and cost of continuing, changing course or stopping.
A pilot already paid for is a reason to learn from it. It is not a reason to scale it.
The ability to withdraw is not a failure mode of the investment case. It is what preserves resources for a more promising use.
A scenario must retain its assumptions
EY’s white paper uses a 12 percent capital-recovery model and presents baseline and narrower geographic scenarios.[1] Those assumptions matter to the interpretation of its results.
They should remain attached to the headline conclusions when the analysis is discussed in a boardroom or translated into an enterprise investment case. Headlines travel; footnotes tend not to.
We propose testing the variables that could reverse the local decision.
A favorable central estimate can coexist with a financially difficult deployment if expenditure arrives early and useful output arrives much later. Different cost and benefit schedules should be compared consistently, with financing assumptions made explicit.
The cost boundary also needs discipline. When a vendor’s price already includes its infrastructure and operating costs, adding those same upstream costs again to the customer’s business case would count them twice. Shared enterprise capabilities require transparent allocation as well: a common evidence service may support several workflows, but each workflow cannot claim the full benefit while none carries an appropriate share of the cost.
The purpose of sensitivity analysis is to expose the decision’s dependence on uncertain conditions. Leadership can then choose a smaller commitment, negotiate different terms, collect further evidence or reject the proposal.
Precision in a model should never conceal uncertainty in the premises.
The unit of value is an accepted outcome
A cheaper model call can reduce one component of expenditure while increasing rework or degrading service. Conversely, a more expensive inference may be worthwhile if it materially improves a consequential decision.
Token efficiency helps manage consumption. Its economic significance depends on the outcome the consumption supports — which is why the statement at the top of this paper divides by accepted cases, not processed ones.
Acceptance criteria must preserve the purpose of the work. Otherwise, improving the reported metric can leave the intended service unimproved.
| What the metric records | What acceptance may also require |
|---|---|
| A ticket marked resolved | Confirmation that the underlying problem was fixed — not simply a closed status |
| A recommendation completed | An adequate evidential basis, and an authorized recipient |
Financial evaluation should then distinguish three things that a single “savings” line tends to blend:
It should compare like-for-like quality and workload, recognize delayed consequences and avoid counting the same benefit more than once. Nonfinancial outcomes can remain separately assessed where monetary conversion would be artificial.
The policy exception is an economic decision
These findings describe an implementation tension; they do not demonstrate why each exception occurred or establish an industry-wide failure rate. But the pairing is instructive. Nearly everyone has the policy. Nearly half have stepped around it when it mattered.
Our interpretation is that urgency deserves examination as a decision about value and authority.
A business owner may perceive delay as costly, while the consequences of insufficient review fall on another function, customers or future operations. If the delivery benefit is immediately visible and the residual exposure is dispersed, the incentives to proceed can be stronger than the incentives to investigate. That asymmetry is an economic fact, and it should be priced as one.
A useful exception process makes the departure explicit rather than informal.
Repeated exceptions are themselves evidence. They should prompt examination of whether the standard process is impractical — or whether incentives encourage avoidable bypass.
The economic case for governance should include its own cost of delay and intervention. At the same time, a delivery deadline cannot automatically confer authority to disregard constraints.
Some boundaries express rights or obligations that the project owner is not empowered to trade away for a better financial result.
Review findings need to change the basis for reliance
EY’s public governance article emphasizes recurring evidence of control effectiveness and reports that formal reviews prompted changes, pauses and shutdowns.[3] This supports examining what review changes in practice.
Finding an issue demonstrates that the review detected something. The further question is whether the response corrected the condition — and improved the next decision. A finding that is logged and closed without changing what the enterprise relies on has produced paperwork, not governance.
We propose connecting each material finding to the relevant authorization, corrective action and evidence of resolution. Material changes to the model, tools, data, policy interpretation or operating environment should trigger reassessment where they undermine that basis for reliance.
In the Causum approach, a Knowledge Property Graph can structure the assertions on which a decision depends, with provenance, conditions and temporal validity. It can preserve the relationship between a benefit claim, the evidence supporting it and a reviewer’s determination. Conflicting or superseded assertions should remain distinguishable, so that later knowledge does not silently overwrite the historical basis for an earlier action.
A graph does not validate its own contents. Acceptance requires suitable evidence and a credible process. Causum’s proposed adjudication role is to coordinate a reasoned determination about a defined claim, under stated criteria and delegated authority.
Independent review needs relevant expertise, access to evidence and safeguards against conflicts of interest. It should complement execution controls and professional assurance — without implying that every action requires a new external approval.
Enterprise returns and public value have different boundaries
A company can earn a return by winning business from a competitor, while the economy experiences a redistribution of existing demand. EY draws the same distinction between private savings and economy-wide growth; it holds at the level of a single deployment too.
Improved quality or lower resource use can still benefit customers in that situation. Conversely, higher measured output does not establish that every affected group is better off. Our proposed value assessment therefore identifies both the beneficiary and the comparison being made.
The distribution of benefits and transition costs belongs in the decision. Leadership should make those effects visible without assuming that they cancel each other out or fit a single financial metric.
This is the connection to political economy: the rules determining ownership, bargaining power, acceptable trade-offs and recourse influence who captures AI’s contribution.
Evidence can clarify those choices. It cannot make them politically or ethically neutral.
A testable commitment before further scale
Start with one material workflow, and treat its business case as a hypothesis with a review date.
- i.Define what would justify further investmentThe outcome, the credible alternative and the scope of delegated authority. Record the assumptions that could change the decision.
- ii.Measure accepted outcomes and full delivery costOver an appropriate period, while preserving evidence of exceptions, interventions and unresolved consequences.
- iii.Have the claim examined from outside the delivery decisionA reviewer examines a specific value claim and its limitations — not the program in general.
- iv.Use the determinationAct on it, in one of four directions.ContinueNarrowExpandWithdraw
- v.Hold the review to its own value testEvaluate whether it improves decisions enough to justify its cost, and update the process as conditions change.