A Tax That Does Not Go Away When the Debt Is Paid

On a formal model separating two costs of deploying stochastic agents that governance conversations have been treating as one.

On May 26, 2026, Muhammad Zia Hydari, Raja Iqbal, and Narayan Ramasubbu published Modeling Agentic Technical Debt and Stochastic Tax on arXiv, a paper that does something the agentic AI governance conversation has largely lacked: it gives a name and a measurable structure to the cost of using probabilistic systems in business workflows, and it insists that cost is not the same thing as technical debt.

The paper's central distinction is precise. Agentic Technical Debt is a stock of accumulated design and governance liability, the familiar category: shortcuts taken during implementation, missing controls, governance work deferred. Stochastic Tax is different in kind. It is a recurring flow of operating burden that arises when stochastic agents are used in business workflows. Debt is something an organization can pay down. Tax is something an organization pays repeatedly, as an ongoing condition of operating a system whose outputs are not deterministic.

Why the Distinction Matters

The paper is explicit that the two constructs are related, but they are not the same: debt can amplify the tax, while the tax can remain positive even when debt is minimized. That second clause is the paper's most consequential claim. An organization that fully remediates its agentic technical debt, closes every governance gap, fixes every architectural shortcut, has not thereby eliminated its operating cost. The tax persists because the underlying mechanism generating it, probabilistic reasoning applied to delegated action, is not itself a defect to be fixed. It is a structural property of the system category. Paying down the debt does not zero out the tax.

This reframes a conversation that governance teams often collapse into a single line item: "technical debt from AI adoption." The paper's model says that line item actually contains two economically distinct components, one that behaves like a balance sheet liability and one that behaves like an ongoing operating expense that never disappears entirely, regardless of how much governance investment is made.

A Model Built to Be Used, Not Just Cited

The paper does not stop at defining the two constructs. It starts from a compact dashboard expression, expands it into a fuller structural model, defines all variables and parameters, and shows how each cost category can be estimated from operational data. This is a deliberate move toward operational usability rather than pure theory. The paper illustrates the framework with an accounts-payable simulation and companion spreadsheet, a specific, auditable business process chosen precisely because it involves delegated, repeated, rule-governed decisions, the exact conditions under which stochastic tax accrues.

What This Gives the Rest of the Arc

This model provides something the broader governance conversation around agentic AI has lacked: a named, structural, measurable variable for the cost of stochasticity itself, independent of any particular failure or incident. Much of the public conversation about agentic AI risk treats cost in terms of specific bad outcomes, a breach, a compliance violation, a hallucinated output that caused harm. The Hydari, Iqbal, and Ramasubbu model treats the recurring burden of operating a stochastic system as a cost category in its own right, present even in the absence of any specific incident, because it is a property of the decision-making mechanism rather than a consequence of any particular failure.

This lands adjacent to what Shangding Gu's harness-scaling paper named as verification cost, a measurable magnitude a system architect has to budget for rather than a binary compliance gate. Both papers, arriving within a day of each other in late May 2026, are converging on the same underlying claim from different directions: the cost of governing a stochastic system is not a one-time expenditure that resolves the problem. It is an ongoing structural feature of the system category, and it needs to be measured as a flow, not written off as a stock.

What the Model Does Not Settle

A dashboard expression and a companion spreadsheet make the stochastic tax measurable in principle. They do not, on their own, establish who inside an organization is accountable for tracking that measurement over time, at what cadence it should be recalculated as workflows change, or what threshold would trigger a decision to stop using a stochastic agent for a given task because the tax has become too costly relative to the alternative. The paper provides the instrument. It does not specify the governance process that would use the instrument on an ongoing basis.

Open Questions

• If Stochastic Tax persists even when Agentic Technical Debt is minimized, what is the actual floor cost of using a probabilistic agent for a given workflow, and has any organization measured it directly?
• Who inside an enterprise is accountable for tracking the stochastic tax over time, given that it is a recurring flow rather than a one-time remediation project?
• At what tax threshold does a workflow stop being a candidate for agentic automation, and who sets that threshold?
• How does the accounts-payable simulation's cost estimation generalize to workflows with less structured, less auditable decision points than accounts payable?
• What standard applies when an organization reports having addressed its "AI technical debt" without distinguishing which portion of the reported cost was debt and which was tax?
• Does the existence of a measurable stochastic tax change the calculus for deploying agents in domains where the tax cannot be reliably estimated in advance?

The loop closed around an oversight function that was never instrumented.