If the State's Audit Selection Is a Black Box, Yours Cannot Be

On September 4, 2026, Greg Matson published a short column on SALT Shaker arguing that algorithmic audit selection has outrun the ability of tax agencies to explain it. Read it twice if you operate AI agents, because the standard it proposes for the state applies with equal force to you: agent operators are the first class of taxpayer whose own tax determinations are produced by a machine, at a volume no human reviewed.

Matson is counsel at Eversheds Sutherland in Washington. Before that, per his firm biography, he "spent nearly two decades with the Multistate Tax Commission (MTC), including more than ten years as its Executive Director." Not an outside critic of state tax administration, then — someone who spent a decade coordinating it.

What He Actually Said

The column opens by defining the problem plainly. The black box problem, he writes, "describes the difficulty of understanding how a complex system reached a conclusion, even when all the inputs are known."

He is careful about what he is not objecting to: "The challenge is not tax administration use of algorithms," he writes, since such systems "have been instrumental in combatting refund fraud." Automated selection is old news — he traces it to the IRS's Discriminant Function System, in use since the 1960s. What has changed is the input side and the oversight side at once. Agencies now hold, in his list, "enormous data sets (like transaction-level sales data), third-party information reports, non-tax agency databases," and "multi-state information sharing arrangements" — alongside "fewer actual human beings available to oversee the process."

The conclusion is the sentence to keep: "No taxpayer should receive an audit notice that effectively says the taxpayer was selected by an unexplained automated system that no remaining human at the agency can meaningfully defend."

His column the week before, on August 27, is the practical half. Taxpayers should ask why they were selected, he wrote, because "while many states provide general descriptions of audit selection methods, some expressly recognize a taxpayer's right to know the reasons or criteria behind the selection." His instruction was one word: "Ask."

Why Agent Operators Sit At the Front of This Queue

Read that list again with agent commerce in mind. Transaction-level sales data, third-party reports, multi-state sharing: an agent operator generates all three at a density a comparably sized human business does not approach. That is the structural exposure, and it has nothing to do with misconduct. Anomaly detection finds outliers, and an operator whose transaction count is three orders of magnitude above its revenue peers, whose average ticket is under a dollar, and whose jurisdiction mix changes weekly is an outlier by construction. You do not have to be doing anything wrong to be interesting to a model trained on returns filed by dentists and distributors.

We have written before about state revenue departments turning AI audit systems on. Matson adds what that piece did not: selection is becoming less explicable at the same moment agencies lose the staff who could explain it.

The Standard Runs Both Ways

If an operator wants the state to open its box, that operator has to be able to open its own. An agent that charged 8.25 percent on a compute purchase in Austin and nothing on an equivalent purchase in Columbia, and whose operator cannot say why, is asking a courtesy it is not prepared to return. "Our system determined it" is the sentence Matson is objecting to, and it does not improve because a taxpayer says it.

The symmetry is operational, not rhetorical. The audit conference he describes — where scope, periods and issues get fixed — requires someone on your side to explain a determination made eighteen months ago by software since updated four times. If the only surviving artifact is a total, that conversation ends before it starts.

What an Explainable Determination Actually Contains

The abstract version of this advice is useless, so here is the concrete one, taken from our engine's output on a live run this morning. A buyer-side compute purchase of $1,000 in Austin returns $66 of use tax; alongside the number, the response carries an audit_trail object of thirteen fields. Four do the real work.

The jurisdiction chain decomposes the 8.25 percent rather than asserting it: a state entry (Texas, 6.25 percent, from the rate table), a local entry (Austin, 2 percent, from the zip table, with the zip recorded), and a special-rule entry recording that the taxable base was reduced to 80 percent, its note field naming the Texas statutory exemption the engine relied on. Three numbers, three provenances, one arithmetic path.

The classification block separates what the caller sent from what the engine decided. It records the input transaction type, the input work type, the resolved work type, and the source of that resolution: explicit when the caller stated it, inferred when the engine did. Most classification disputes are really disputes about who chose the category, and this field settles that without argument.

The evaluated exemptions list the checks that did not fire: no-sales-tax (Texas has a 6.25 percent state rate), financial-service (category is data processing), B2B statutory (that exemption does not reach this category in Texas). A record of the roads not taken is the part an auditor asks about and the part almost nobody logs.

The final determination and policy references close the loop, naming the internal position the answer rests on and whether we treat it as settled. A separate confidence score carries a factor line worth reading literally — "Confidence reflects input data completeness, not legal certainty of classification" — because a strong score means we had your zip, not that a court agrees with us.

Reproducibility Is Not Logging

Logging tells you what a system did. Reproducibility tells you that you can derive it again. For an audit reaching back three years only the second is worth anything, and the difference shows up in how a system handles a date it does not model. Ours refuses. Asked to compute a Chicago SaaS transaction as of June 15, 2024, the engine returns no number at all: success: false, error code AS_OF_DATE_BEFORE_MODELED_WINDOW, naming Chicago, its earliest supported date of January 1, 2025, and the date requested. The message says why — "the engine's current rules do not state the law in force on that date" — and that it "refuses rather than silently applying current law to a historical period." An unparseable date is rejected on its own separate code rather than quietly resolved to today.

The override is instructive. Pass the engine-level opt-in and the same call returns $150, flagged is_estimate: true, with an advisory that the figures apply current law rather than the law in force on the requested date, an instruction not to persist it to any ledger or filing, and a confidence score dropped from 85 to 74. The estimate is available; passing it off as a determination is not.

The alternative is the quiet failure mode in every tax system I have examined: a 2024 transaction replayed through 2026 rate tables, producing a clean, wrong number indistinguishable from a right one. It survives internal review. It does not survive an auditor holding the 2024 schedule.

Where Our Own Box Is Dim

Reciprocity obliges me to name ours. We do not carry every zip code. Ask for $1,000 in Baton Rouge and the engine returns $50 — not because Louisiana's combined rate there is 5 percent, but because we hold the state rate and not the parish layer. What makes that defensible rather than negligent is that the response says so, in a ZIP_UNKNOWN advisory naming the zip, stating that only the state rate was applied, warning the actual combined rate may be higher, and telling the caller to verify with the local department before remitting. That is the honest shape of a gap: disclosed machine-readably at the moment of determination, not discovered by an auditor two years later.

What Agent Operators Should Do

• Persist the reasoning, not the result. A ledger holding a tax amount and a timestamp gives you a number and no defense. Store the determination artifact — chain, classification, evaluated exemptions, engine version — per transaction.

• Pin the engine version on every row. "The system changed since then" is a fact, not an excuse, and it is only usable if you recorded which version produced which row.

• Record what did not apply. Every exemption considered and rejected, with the reason. The cheapest documentation you will ever write and the most valuable one in an examination.

• Test your replay path before you need it. Recompute a transaction from eighteen months ago. If the number changes and your system does not say so, that is the finding, and you found it yourself.

• Take Matson's advice and ask. If a selection notice arrives, ask on what basis. Some states recognize your right to know, and the answer belongs in the file either way.

What to Watch

Watch whether any state publishes generative-AI selection criteria specific enough to be checked; Matson's column implies none has yet, and I have seen nothing to the contrary. Watch also for the first assessment defended on appeal with a machine-generated audit trail as the taxpayer's primary evidence — that case will set the documentation standard faster than any regulation.

One thing this does not resolve. Whether any state selection model treats an autonomous agent's transactions differently from a human's is unknown to me and, so far as I can tell, undisclosed — not a settled question I am simplifying, an unanswered one. What is knowable today is which side of the conversation you control.

To see what a determination artifact looks like against your own transactions, our tax methodology documents how each figure is derived, and the AI agent sales tax guide covers where agent activity sits in each state's tax base.

This analysis is for informational purposes only and does not constitute legal or tax advice. This post reflects AgentTax's current interpretation of evolving law. Consult a licensed tax professional for compliance decisions.