Making AI Trust and Value Claims Verifiable

The agent has been live for a quarter. The value ledger shows processing time down by almost a third. The trust ledger shows no rise in complaints. Two well-kept records, both favorable, both pointing the same way.

The proposal on the table is to extend the agent to a second channel and widen its authority.

But the same two numbers fit at least four different stories — and only one of them justifies expansion. Neither the favorable reading nor the unfavorable ones should be assumed. The evidence has to tell them apart, and neither ledger, on its own, can.

Governance creates economic value when it helps an enterprise make better decisions about what AI may do, where its contribution is worthwhile and when its authority should change. Demonstrating that contribution requires evidence that connects controls to decisions, and decisions to consequences.

KPMG in India’s 2026 paper, AI Governance: The Architecture of Trust in an Age of Intelligent Power, offers a substantial foundation. It addresses embedded and third-party AI, information integrity, agent permissions and controls throughout the lifecycle. It also proposes a value ledger and a trust ledger for material use cases.[1] Our response begins with those commitments and asks how they become a defensible basis for delegating autonomy.

For Causum, the central issue is the quality of the claim an enterprise asks others to accept. Not the number of controls in place or the maturity score of the program, but the precise statement it puts forward — and whether that statement would survive examination by someone with no stake in the answer.

We propose that the next step is to make those questions answerable, through connected evidence, proportionate adjudication and explicit decisions about value.

A governance program that can answer them is producing evidence. One that cannot is producing documentation.

A value record and a trust record are each useful. Read side by side, they invite a conclusion neither can support alone. Their joint interpretation depends on whether they cover the same population, period, workflow and operating conditions — and an improvement in one service channel cannot establish the safety or value of another.

We propose connecting benefit and control evidence to a common unit of assessment: a defined decision or workflow, its authorized scope and its observed consequences.

Relevant records should identify the baseline, affected population, evidence source, uncertainty and accountable owner. The decision to expand autonomy should reference that context — including the exceptions that remain unresolved, which are usually the most informative entries in either ledger.

Separate measures remain useful. Combining them into a single trust-and-value score can obscure trade-offs that leadership must decide explicitly. A composite is easy to report and hard to argue with, which is precisely the problem: it averages away the cases leadership most needs to see.

A severe breach of an agreed boundary may prohibit an action even when average productivity improves. The appropriate outcome may then be continued operation under narrower authority, pending evidence that the problem has been corrected — neither a shutdown nor business as usual.

Our value economics lens begins with a question benefit reports rarely ask: what would plausibly have happened without the intervention?

Where feasible, comparison groups and staged introduction can help separate these effects. Where the evidence is weaker, the claim should be narrower and competing explanations should remain visible.

The same discipline applies to governance. Avoiding an incident has value, but a period without incidents does not demonstrate how many losses were prevented. A quiet quarter is an observation, not a measurement of prevented loss. Estimated risk reduction should therefore retain its assumptions and uncertainty, rather than hardening into a figure on a slide.

Realized financial benefits should account for review effort, integration, monitoring, remediation and the operating costs of the AI itself — and three kinds of benefit should be recorded distinctly to prevent double counting.

This also means evaluating the control. A review process may reduce errors while imposing excessive delay on routine work. Governance is itself an intervention, with costs and effects, and it should face the same test it applies to AI. The objective is to find an effective level of intervention for the consequences at stake, then reassess it as evidence accumulates.

KPMG’s paper goes beyond naming agents: its recommendations include bounded permissions, business ownership, approvals, logging, exception handling and mechanisms to stop execution.[1] Those measures create useful conditions for control.

A further question arises at the moment of action: does the available evidence justify using this permission in this situation? It is the gap between being allowed to act and having good reason to.

An authenticated procurement agent may be allowed to place an order below a financial threshold. That permission does not establish that demand is current, the supplier remains suitable or the proposed order respects an applicable exception. Three distinct things are being established, and only two of them are routinely recorded.

identity:       procurement_agent        authenticated
authorization:  order < financial threshold   within scope

basis for this order:
  - demand is current                ?
  - supplier remains suitable        ?
  - applicable exception respected   ?

if evidence disagrees or has expired:
    → review, not a confident recommendation

In the Causum approach, a Knowledge Property Graph can structure domain assertions together with provenance, conditions, temporal validity and the relationships needed to assess them. A statement about supplier eligibility can be linked to its evidence and the context in which it applies. Disagreement or expired evidence can trigger review rather than silently becoming a confident recommendation.

This is a proposed role for the KPG, not a claim that a graph guarantees truth. Its assertions still require credible sources and an appropriate acceptance process. The point is to make the basis for reliance inspectable and revisable — so that permissions are exercised with relevant knowledge and accepted uncertainty.

An enterprise can set conditions for a vendor and control the interfaces it owns. Its visibility into the vendor’s internal AI may remain limited. Every enterprise relying on a supplier’s model is, in part, relying on something it cannot see.

A service provider may expose an API and publish a statement about its model controls without exposing the decision process behind individual results. The customer can test outcomes, restrict submitted data and monitor calls through its own boundary. It cannot infer from those measures alone that every internal action met its policies.

Contractual obligations, vendor attestations, runtime observations and independently verified findings provide different kinds of evidence. A promise, a statement, an observation and a verification are not interchangeable, and a decision to rely on the service should record which of them it actually rests on.

We propose recording the limits of observation alongside the scope of assurance. For consequential workflows, leadership should decide whether the available evidence is sufficient, whether further review is required or whether the permitted use must be reduced.

This distinction is economically relevant. A cheaper service may require more independent checking, introduce exit costs or leave more residual exposure. The value assessment should include those consequences rather than comparing subscription or inference costs in isolation. The cheapest service on the invoice is not always the cheapest service to rely on.

An enterprise needs a process for resolving disagreement about whether an action was justified or a reported benefit was substantiated. Without one, disagreements tend to be settled by seniority, or simply outlasted.

We use adjudication to describe a reasoned determination under stated criteria and delegated authority. It does not imply judicial authority or formal audit assurance.

Those relying on the determination should understand which of these outcomes occurred. The third matters most: a process that cannot say not yet is not adjudicating. It is ratifying.

Independence must be designed for the claim being reviewed. A qualified team outside delivery may be sufficient for some uses; material external reliance may warrant a separate assessor. NIST AI RMF 1.0, MEASURE 1.3, supports regular assessment involving internal experts outside frontline development and/or independent assessors.[3] A new organizational label alone does not address conflicts of interest, lack of expertise or restricted access to evidence.

Causum’s proposed clearinghouse role is to coordinate this process across providers, enterprises and qualified reviewers — connecting evidence to determinations and retaining the history of challenges and revisions. Its own performance must be judged by whether it improves the credibility and timeliness of decisions at a proportionate cost.

Proportion cuts both ways. Routine, reversible actions should not need a new external approval every time they run — an adjudication layer that slows everything would fail the same value test it exists to apply.

Enterprise-level benefit can conceal a transfer of cost — out of the measured workflow and onto customers, employees or partners who appear in no ledger.

These are possibilities to investigate, not conclusions about any specific deployment.

A useful value claim therefore identifies its beneficiaries, affected groups and measurement boundary. Where interests conflict, leadership should state the trade-off and the basis for accepting it. People bearing material consequences need a workable route to correction or challenge. Financial metrics can inform that decision without settling every question of legitimacy.

This is the political economy within AI value economics. A governance architecture becomes more credible when those choices are visible and contestable.

None of this requires an enterprise-wide program to begin. It requires one workflow where the stakes are real, and the discipline to treat its claims as claims.

KPMG’s recommendations create a useful opening for this work. The next question for leadership is whether its trust and value records allow a qualified reviewer to establish that a particular delegation of autonomy remains justified.

This independent Causum perspective does not imply endorsement by KPMG or NIST. The claims operation and procurement agent are hypothetical; the diagrams render the paper’s argument, not a measured dataset.