A Causum perspective on proving the value of agentic AI — and on why autonomy should expand only in proportion to the quality of the evidence supporting its use.
Monthly benefits report · 1,000 hours released · $50 an hour
The report says $50,000 a month. Payroll has not changed.
AI released a thousand staff hours this month. Management valued each one at fifty dollars and reported the product as benefit. No expense was removed. No additional useful work was delivered.
The figure will appear in the board pack. It will be cited in the case for the next deployment, and the one after that.
What it describes is capacity, valued at an assumed rate. An enterprise can deploy capable agents, assign human owners and report higher productivity — while still being unable to establish whether the resulting decisions were justified, or whether the organization became better off.
Our position is that autonomy should expand in proportion to the quality of the evidence supporting its use. That evidence must connect authority, decisions and consequences to outcomes that matter to the people affected.
Accenture and Wharton’s 2026 report, The Age of Co-Intelligence, provides a useful starting point. It connects the redeployment of released capacity to growth, retains human responsibility for agentic work, and calls for governance proportionate to decision risk. Its methodology describes economic estimates as conditional scenarios rather than guaranteed outcomes.[1] We build on that argument by asking what would make these commitments demonstrable in an operating enterprise.
For Causum, value economics examines how benefits arise, how they are attributed, what it costs to obtain them, and who receives the benefits or bears the consequences.
Can the enterprise substantiate a claim of beneficial, authorized performance strongly enough for another party to rely on it?
Time saved creates an opportunity. Converting that opportunity into value requires a further decision: remove an actual expense, meet previously unmet demand, improve service quality, reduce a relevant exposure, or undertake work that could not otherwise be completed.
Each route needs its own evidence. An employee completing a task sooner does not by itself establish a payroll saving or an increase in revenue.
A defensible assessment therefore separates three things that a benefits report tends to merge.
| Claim | What it can honestly state |
|---|---|
| Estimated opportunity | A forecast, with an explicit time horizon and stated uncertainty |
| Observed operational improvement | What measurably changed in the operation, compared with a credible alternative |
| Realized economic contribution | Expense actually avoided, or margin actually attributable — net of integration, inference, supervision, rework and transition costs |
Material effects on service quality, people or rights should remain visible even when they cannot be credibly converted into money.
What changed because of this deployment, compared with a credible alternative — and what evidence would cause us to revise the benefit claim?
Naming a human owner establishes a point of responsibility. Effective control additionally requires sufficient knowledge, time and authority to intervene.
Receives hundreds of decisions. Sees only machine-generated explanations. Cannot suspend execution.
Carries responsibility without meaningful influence over the outcome.
Knowledge of the evidence, time to examine it, and the power to withhold, contest or stop.
Demonstrated under realistic workload, not asserted at sign-off.
We propose assessing oversight through observable capabilities rather than organizational charts.
Can the responsible person see the evidence behind a consequential recommendation?
Can they contest an interpretation, withhold permission and stop an action before an irreversible consequence?
Are escalation routes staffed — and can unresolved concerns reach someone with the authority to act?
Do interventions actually alter what the system does next?
These capabilities should be tested under realistic workload and failure conditions. A successful demonstration at low volume does not establish that the same review arrangement works at enterprise scale. Where supervision becomes the constraint, the response may be narrower permissions, slower execution or a redesigned workflow.
Expanding autonomy to conceal an oversight bottleneck makes the evidence weaker precisely when reliance becomes greater.
NIST’s AI RMF 1.0 supports clear responsibilities for human–AI oversight and regular assessment involving experts outside the frontline development team or independent assessors.[2] We extend that principle into a practical test of whether authority remains effective during operation.
An execution log can show that a tool ran. It may leave unanswered whether the source information was current, the interpretation was appropriate, the action was authorized for this situation, or the eventual result was beneficial.
Evaluation therefore needs a record that connects the claim being made to the evidence available when the decision occurred — and to what happened afterward.
In Causum’s proposed approach, a Knowledge Property Graph, or KPG, provides a domain-specific structure for knowledge assertions and their supporting relationships. Conflicting assertions should remain visible for examination.
Storing a claim in a graph does not make it true.
This structure can support an agent’s grounding and a reviewer’s assessment. It cannot guarantee the correctness of an answer or substitute for legitimate authority. Its usefulness depends on the quality of its sources, the adequacy of the domain model and the credibility of the people and processes maintaining it.
These are proposed design properties, not claims of demonstrated product performance.
A delivery team has legitimate incentives to demonstrate that its investment succeeds. A separate reviewer can help test those claims — provided the reviewer has access to relevant evidence, appropriate expertise and the freedom to reach an unfavorable conclusion.
Organizational separation alone does not establish independence; incentives, appointment arrangements and conflicts of interest also matter.
We use adjudication to mean a reasoned determination about a defined claim under stated criteria and delegated authority. It does not imply judicial power or formal audit assurance. One claim may concern whether an action respected its authorization; another may concern the reliability of the reported business benefit. Those claims require different expertise and should not be collapsed into one trust score.
This creates a possible role for a clearinghouse that coordinates evidence, assigns qualified assessors, records determinations and handles disputes across organizations. Its value would have to be demonstrated through better decisions, fewer unresolved disputes or lower verification costs. An extra approval layer that adds delay without improving those outcomes would fail the same value test.
Review should be proportionate.
An acceptable process must also specify appeal, correction and withdrawal of authorization.
A financial improvement for one enterprise can coexist with additional burdens elsewhere.
Faster employee output. A leaner service process.
Customers spend longer resolving errors. Output depends on sustained increases in cognitive load.
These are possibilities that evaluation should investigate, rather than assumptions that every deployment causes harm. We therefore propose that every value claim carry its own boundary.
| Stated with each value claim | The question it answers |
|---|---|
| Beneficiary | Who receives the benefit? |
| Affected groups | Who bears the consequences? |
| Measurement boundary | What was counted, and what was left outside? |
| Authority over trade-offs | Who may accept them on the organization’s behalf? |
| Route to challenge | How can affected people contest them? |
Financial contribution, service quality, working conditions and distributional effects can be assessed together without pretending they share a universal monetary unit.
This is where value economics meets political economy.
Decisions about whose interests count, which costs are recognized and who can contest an outcome shape the distribution of AI’s benefits. Technical performance informs those choices but cannot settle them.
An enterprise can begin with one consequential workflow.
Set out the intended outcome, current baseline, authorized scope, accountable owner and the evidence required for acceptance.
Connect permissions, relevant assertions, actions and outcomes. Use comparison groups or staged deployment where practical; where causal attribution is weaker, disclose competing explanations and use conservative claims.
Minimize sensitive data in the evidence record, with access and retention controls appropriate to its purpose.
A reviewer outside the delivery decision assesses the claim and its limitations. Leadership then decides what happens to the delegated authority.
The cost and delay of review are monitored as part of the operating economics, so that governance remains both effective and proportionate.
The decision to expand should rest on demonstrated performance within known conditions, supported by a credible ability to detect and respond when those conditions change.
Authorization should expire or be reassessed when the supporting evidence becomes stale or the deployment materially changes.
Accenture’s emphasis on deliberate operating-model change is a constructive foundation.[1] Our proposed extension is an explicit commitment to make claims about authority and value open to examination.
An enterprise should be able to explain why a consequential action was permitted, how its result was evaluated and who may challenge the conclusion.
What evidence would persuade a qualified, independent reviewer that this delegation of autonomy produced worthwhile outcomes within legitimate boundaries?
If that evidence cannot be assembled, strengthen the basis for reliance before expanding the scale of execution.
This independent Causum perspective does not imply endorsement by Accenture, Wharton or NIST. The service-operation figures are illustrative; the diagrams render the paper’s argument, not a measured dataset.
From Co-Intelligence to Justified Autonomy · Perspective Paper · October 2026
Ray Fatahi · Causum Research
© 2024–2026 Causum. All rights reserved.