How AgentTax Classifies AI-to-AI Transactions

Tax law doesn't have a category for "one AI agent buying inference from another AI agent." It has categories like "data processing services," "prewritten software," "specified digital products," and "information services." The gap between what agents actually do and how tax law classifies what they do is where compliance errors happen.

AgentTax bridges this gap with a classification system that maps the six primary types of agent-to-agent transactions to the appropriate tax categories in each state. Get this right, and your tax calculations are accurate. Get it wrong, and you're either overpaying (eroding margins) or underpaying (creating liability).

The Classification Problem

Consider a simple transaction: Agent A purchases 100 GPU-hours from Agent B.

How should this be classified for tax purposes? Depending on the state, it could be:

• Data processing services (Texas) — taxable at 6.25%, with the 80% rule reducing the effective rate to 5.0%

• Computer services (Mississippi) — taxable at 7.0%

• Information services (New York) — exempt, unless the information is considered a "digital good"

• Tangible personal property (if the buyer is downloading a processed result) — taxable in most states

• Non-enumerated service (some states) — exempt because the state hasn't specifically listed it as taxable

The same transaction. Five different potential classifications. Five different tax outcomes.

This is why "just use the sales tax rate" doesn't work. The rate is only half the equation. The other half is whether the transaction is taxable at all — and that depends on classification.

AgentTax's Six Transaction Types

We've identified six primary transaction types that cover the vast majority of agent-to-agent commerce. Each maps to one of three tax categories:

1. Compute / Processing → digital_service

What it covers: Cloud compute, GPU time, batch processing, inference jobs, data transformation.

Tax category: Digital service / data processing. Taxable in states that tax data processing services (~30 states).

Examples: An agent renting GPU capacity from a compute marketplace. An agent paying for batch inference on a model. An agent purchasing data transformation services.

Key state variations:

• Texas: "Data processing services" — 80% taxable rule applies

• Ohio: "Automatic data processing" — fully taxable

• California: Not taxed (most digital services exempt)

2. API Access → digital_service

What it covers: Calls to external AI/ML APIs, inference endpoints, search APIs, embedding services.

Tax category: Digital service. The distinction from compute is subtle but important — API access is metered per-call rather than per-resource, and some states treat subscription vs. metered access differently.

Examples: An agent calling OpenAI's API for completions. An agent accessing a vector search endpoint. An agent using a classification API.

Key state variations:

• Most states treat API access identically to other digital services

• Connecticut's 1% SaaS rate may apply if the API is accessed via a subscription

• Some states distinguish "real-time" data feeds (exempt as "information services") from processing APIs (taxable)

3. Data Purchase → digital_good

What it covers: Buying datasets, training data, pre-computed embeddings, labeled data, research reports.

Tax category: Digital good / specified digital product. This is important — more states tax digital goods than digital services. A state that exempts SaaS might still tax a dataset purchase.

Examples: An agent licensing a training dataset. An agent purchasing pre-computed vector embeddings. An agent buying labeled data for fine-tuning.

Key state variations:

• New Jersey: Exempts SaaS but taxes "specified digital products" — a data purchase would be taxable

• Virginia: Exempts SaaS but may tax "digital personal property"

• Georgia: Generally exempts digital services but has evolving positions on digital goods

4. SaaS / Software → digital_service

What it covers: Software subscriptions, cloud tools, monitoring platforms, analytics services, CI/CD tools.

Tax category: Digital service (SaaS). The most heavily legislated category — about 30 states now explicitly tax SaaS.

Examples: An agent with a subscription to a monitoring service. A development agent using CI/CD tooling. An agent fleet managed through a cloud orchestration platform.

5. AI Agent Labor → service

What it covers: Work performed by one AI agent for another. Code review, content generation, analysis, research, data annotation.

Tax category: Service. This is the gray area. Is it a "data processing service"? An "information service"? A "professional service"? The answer depends on the state and has almost no precedent for AI-performed work.

Our classification rationale: We classify AI labor as a general service rather than a digital_service because the buyer is paying for the output of work, not for access to software or data. This aligns with how most states treat human-performed services (consulting, analysis, research) and avoids the digital service classification that would make it taxable in more states.

Key state variations:

• Hawaii and New Mexico tax virtually all services under their broad-based systems

• Most states with sales tax do not tax non-enumerated services

• If the labor produces a deliverable (a document, dataset, or report), it could be reclassified as a digital good in some states

6. Storage → digital_service

What it covers: Data storage, model hosting, file systems, object storage, vector stores, model registries.

Tax category: Digital service. Storage is generally treated as a cloud computing service, taxable where digital services are taxable.

Examples: An agent storing data in an S3-compatible service. A model hosted on a registry. Vector data stored in a managed database.

How Classification Affects Your Tax Bill

Here's a concrete example showing how the same $10,000 transaction is taxed differently based on classification and state:

| State | Compute (digital_service) | Data Purchase (digital_good) | AI Labor (service) |
|---|---|---|---|
| Texas | $500 (6.25% × 80%) | $625 (6.25%) | $0 (services not taxed) |
| California | $0 (exempt) | $0 (exempt) | $0 (exempt) |
| New York | $0 (info services exempt) | $400 (4% digital goods) | $0 (not enumerated) |
| Washington | $650 (6.5%) | $650 (6.5%) | $650 (B&O tax applies) |
| New Jersey | $0 (SaaS exempt) | $663 (6.625% digital goods) | $0 (not enumerated) |

The difference between classifying a transaction as compute vs. data_purchase could be $625 in Texas or $663 in New Jersey. Across thousands of transactions per year, classification accuracy directly impacts your bottom line.

Using Classification in the API

You specify the transaction type in your API call, and AgentTax handles the state-specific mapping:

POST /api/v1/calculate
{
  "role": "seller",
  "amount": 10000.00,
  "buyer_state": "NJ",
  "transaction_type": "data_purchase",
  "counterparty_id": "agent_research_lab"
}

// Response
{
  "sales_tax": {
    "jurisdiction": "New Jersey",
    "state": "NJ",
    "rate": 0.06625,
    "amount": 662.50,
    "type": "Digital Goods Tax",
    "reason": "NJ taxes specified digital products including datasets"
  }
}

Change transaction_type to "compute" for the same buyer state, and the result changes:

{
  "sales_tax": {
    "jurisdiction": "New Jersey",
    "rate": 0.06625,
    "amount": 0.00,
    "type": "Exempt",
    "reason": "NJ does not tax SaaS / digital services"
  },
  "exemption": {
    "type": "digital_exempt",
    "detail": "Digital services not taxable in NJ"
  }
}

Same amount, same state, different classification, different result. This is why getting the classification right matters.

When Classification Is Ambiguous

Sometimes a transaction doesn't fit cleanly into one category. An agent might purchase "GPU compute that produces a dataset" — is that compute (digital service) or a data purchase (digital good)?

Our general rule: classify based on what the buyer is paying for, not what the seller is doing.

• If the buyer is paying for access to a resource (compute time, API calls, storage capacity): it's a digital_service

• If the buyer is paying for a deliverable (a dataset, a report, a trained model): it's a digital_good

• If the buyer is paying for work to be done (analysis, code review, content creation): it's a service

When in doubt, the more conservative classification (the one that results in higher tax) is safer from an audit perspective. You can always get a refund for overpaid tax; you can't un-ring the bell on underpaid tax.

Explore transaction types in detail →

Next: API-First Tax: Why AI Agents Need a POST, Not a Portal