Open Letter: Tax Policy Needs an Update for AI

To: State Tax Administrators, the Multistate Tax Commission, and the Streamlined Sales Tax Governing Board

From: AgentTax Research

Re: The growing compliance gap in AI agent commerce — and what policy can do about it

We are writing to flag an accelerating gap between U.S. sales and use tax policy and the reality of modern AI-powered commerce.

The tax code's core frameworks — nexus, product classification, collection obligations, filing — were designed for a world of physical stores, human buyers, and discrete, infrequent transactions. That world is changing fast. AI agents are now buying and selling digital services autonomously, at machine speed, across every state jurisdiction simultaneously. The compliance infrastructure hasn't kept pace, and the gap is widening daily.

This isn't a theoretical concern. McKinsey projects agentic commerce could generate $3 to $5 trillion globally by 2030 (McKinsey, October 2025). Visa, Google, J.P. Morgan, and dozens of other companies are actively building the payment and protocol infrastructure for agent-initiated transactions (Visa, 2025; DevPro Journal, January 2026). The AI agents market itself is projected to exceed $10.9 billion in 2026 (Grand View Research).

The Compliance Gaps

We see five specific areas where current policy creates structural compliance challenges for AI agent commerce:

1. Product Classification Is Ambiguous for AI Services

When an AI agent sells "inference compute" or "model training time," how should it be classified? As data processing (taxable in Texas at 80%)? As a digital automated service (taxable in Washington)? As an information service (potentially different treatment)? As a non-enumerated service (exempt in many states)?

The same AI service can reasonably fit multiple classifications depending on how a state interprets its own statutes. This isn't a problem that better compliance tools can solve — it requires clearer legislative guidance on how AI-native service categories map to existing tax frameworks.

Recommendation: The Streamlined Sales Tax Governing Board should consider adding AI-specific service categories to the Taxability Information Code (TIC) system, providing consistent classification for common AI transaction types: compute, API access, model training, agent labor, and data services.

2. Economic Nexus Thresholds Were Designed for Human-Speed Commerce

The standard $100,000 economic nexus threshold was calibrated for businesses that grow gradually into new markets. AI agents can cross that threshold in a single state within days of launching. An agent selling compute at moderate volume can trigger nexus in a dozen states within its first quarter.

This isn't inherently a policy problem — the thresholds work as designed. But the speed at which agents trigger nexus means the window between obligation-trigger and first-filing is extremely compressed. Businesses using AI agents need to monitor nexus in real time, not at quarterly intervals.

Recommendation: States should clarify that economic nexus monitoring obligations extend to automated sales channels, and should provide machine-readable APIs for threshold lookups and registration, enabling compliance tools to monitor and alert in real time.

3. Use Tax Enforcement Relies on Buyer Self-Assessment

Use tax has always depended on buyer compliance, and compliance has always been low. In agent-to-agent commerce, the problem is compounded: buyer agents have no mechanism to check whether a seller collected tax, and no built-in process for self-assessing use tax when the seller didn't.

The result is a growing pool of untaxed transactions that neither the seller nor the buyer is tracking. States are deploying AI-powered audit tools (OurTaxPartner, 2026), but enforcement after the fact is far less efficient than enabling compliance at the point of transaction.

Recommendation: Encourage the development of seller-remittance verification standards — a way for buyer agents to query, at the point of transaction, whether the seller is registered and collecting in the buyer's state. This would allow use tax obligations to be identified immediately rather than discovered in an audit years later.

4. Filing Burden Discourages Small-Operator Compliance

A GAO report found that some businesses incurred $1,500 in monthly compliance costs to remit less than $500 in sales taxes (TaxConnex, 2024). For small AI agent operators — solo developers, small startups — the cost of compliance in 20+ states can exceed the tax liability itself.

This economic irrationality discourages compliance. Rational economic actors will delay registration and collection when the cost of compliance exceeds the perceived risk of enforcement. This is bad for revenue collection and bad for the competitive landscape, as compliant businesses bear costs that non-compliant businesses avoid.

Recommendation: Expand the Streamlined Sales Tax program's free-filing provisions for small sellers. Consider a federal small-seller safe harbor, as the GAO has recommended, that exempts businesses below a meaningful national threshold from multi-state collection obligations. And invest in making registration and filing APIs that compliance tools can integrate with directly, reducing the cost of compliance for automated systems.

5. No Standard for Agent Tax Identity

In human commerce, the buyer's tax-relevant identity is established at checkout — shipping address, exemption certificate, entity type. In agent-to-agent commerce, there's no checkout. There's an API call.

Agents need a standardized way to communicate tax-relevant information: jurisdiction, entity type, exemption status, nexus registrations. Without this, every transaction requires manual follow-up or assumptions that may be incorrect.

Recommendation: Work with the emerging agent commerce standards bodies — including the Linux Foundation's Agentic AI Foundation, which counts Anthropic, Google, Microsoft, and OpenAI among its members (McKinsey, January 2026) — to define tax identity fields that can be included in agent-to-agent protocols like A2A and MCP.

The Opportunity

The agent economy presents a unique opportunity for tax administration. Unlike human commerce, where compliance depends on individual behavior, agent commerce is programmable. If the rules are clear and the infrastructure is accessible, compliance can be built into the transaction protocol itself — automatically, at scale, with perfect accuracy.

This is actually better for tax collection than human-mediated commerce. Agents don't forget to collect tax. They don't misclassify products out of ignorance. They don't delay registration because the paperwork is annoying. They follow whatever rules are programmed in.

The limiting factor isn't technology — it's policy clarity. If states provide clear, machine-readable guidance on classification, nexus, and rates, the compliance tools will implement it. If guidance remains ambiguous, fragmented, and accessible only through legal research, the gap will continue to widen.

We're building AgentTax to solve this problem from the private sector side — giving AI builders a single API to handle tax calculation, classification, and reporting. But the system works best when the rules it implements are clear, consistent, and designed for the speed of modern commerce.

The agent economy is here. The tax framework should be too.

Sources:

• McKinsey, "The Agentic Commerce Opportunity," October 2025

• McKinsey, "The Automation Curve in Agentic Commerce," January 2026

• Grand View Research, "AI Agents Market Size and Share," 2025

• Visa, "Visa and Partners Complete Secure AI Transactions," 2025

• DevPro Journal, "NRF 2026: The Rise of Agentic Commerce," January 2026

• OurTaxPartner, "2026 State Sales Tax Rates," February 2026

• TaxConnex, "As Wayfair Turns 6," 2024

• South Dakota v. Wayfair, Inc., 585 U.S. ___ (2018)