California Asked Industry How to Tax Software. Seventy-Three Pages Came Back, and Six Lines Mention AI.

Key Takeaway: The CDTFA posted Exhibit 9 to its S.B. 122 discussion paper - 73 pages of written comments from the companies and trade groups that will pay this tax. The record is detailed, competent and almost entirely about employees. Artificial intelligence is raised in six places across all of it, and no comment asks how to source a software transaction with no human user. The draft rules answer the questions that were asked. The meeting is September 10; written comments close September 24.

What Exhibit 9 Is

When the CDTFA published its discussion paper and eight draft regulations on September 1, the package arrived with a ninth attachment that got less attention: the written comments already received, reproduced in full. The cover letter from Sandy Barrow, Chief of the Tax Policy Bureau, sets the calendar - an interested parties meeting on September 10 in Sacramento with a Teams option, and a September 24 deadline for written suggestions, "including any proposed regulatory language."

Exhibit 9 is the input side of that process: letters from the California Taxpayers Association, CCIA, the Council On State Taxation, the Silicon Valley Tax Directors Group, the Silicon Valley Leadership Group, TechNet, Salesforce, Baker Tilly, Eversheds Sutherland, the realtors, a credit union coalition, an irrigation district and a public library, plus the questions typed into the chat during the July 21 workshop.

I read all 73 pages. The reason to read it is not gossip about who asked what. It is that a draft regulation is downstream of its comment file, and this file tells you why the draft looks the way it does.

The Question That Got Asked, and Answered

Start with the case where the process visibly worked.

Multiple commenters wanted infrastructure out. CalTax asked the regulations to "address backend computing resources (e.g., managed databases, machine learning environments, computing APIs) and clarify that they qualify, even if accessed via a proprietary web UI." The Silicon Valley Leadership Group supplied a drafted Infrastructure as a Service definition and asked that it be "explicitly include[d]." Three chat participants asked the same thing, one plaintively: "Hi all did we understand that as IaaS will be taxable? Read this as out of scope. But that comment threw me off."

Draft Regulation 1600 answers them by name. Its list of exclusions includes "Digital infrastructure, as defined in RTC section 6016.1, such as cloud-based Infrastructure as a Service (IaaS) and Platform as a Service (PaaS) that allow customers to create, deploy, or run their own software application."

That is a clean loop: the industry asked, the Department wrote it down. Keep it as the control for everything below.

The Question That Got Asked Narrowly

Artificial intelligence appears in Exhibit 9 six times. Two are a company describing itself and a membership roster. The other four are asks.

CCIA wrote that the framework "must explicitly address AI-Enabled Offerings (such as autonomous decision engines, generative models, machine learning APIs, and integrated assistant tooling)." CalTax wrote that "it would also be helpful to have clarity on generative AI and AI models, one way or the other," and asked whether prewritten computer software "include[s] SaaS, PaaS, AI products?" In the workshop chat, one participant asked the whole question in six words: "Will AI tools be created as Saas?"

The draft package's answer is a single worked example, in new Regulation 1502.2 on custom computer software. A vendor offers an "Artificial Intelligence (AI)-powered tax-analysis tool" inside a prewritten compliance platform and trains it on California's statutes, regulations and legal opinions so its suggestions are better for California users. The Department's conclusion: those outputs "are not the result of computer software prepared to the special order of a single customer" but "trained, automated outputs produced by a generalized machine-learning system that is sold or licensed to many customers," and so are not custom software.

That is a real answer and a useful one - fine-tuning on a customer's jurisdiction does not buy the custom-software exemption. But notice its shape. It answers is my AI product custom? with a no. It says nothing about where an AI product is used, who counts as using it, or how a charge is divided between California and everywhere else, which is the question that decides the bill.

The Question Nobody Asked

Draft Regulation 1600.2 governs digital products purchased for multiple points of use. Its definitions open with this: "'User' means an employee or agent of the purchaser that is authorized by the purchaser to use the digital product in the performance of their duties as an employee or other agent of the purchaser."

Duties. Employment. The word "agent" there is agency law, not software.

The mechanics follow. Subdivision (d)(3) presumes that "calculating the measure of tax based on the number of users or computer inside this State and outside this State is a reasonable alternative method." Subdivision (d)(4) is one sentence: "A method of calculating the measure of tax based on the location of the servers where the software is installed is not considered reasonable."

For an autonomous workload that is the entire problem. There are no seats. There may be three human beings on the account while the software executes millions of times across regions with no person in any of them. The one physical fact such a workload has is where the compute happened, and the draft rejects it by name.

Exhibit 9 explains why nobody fought about it. Every worked example in the file is a person at a desk:

• The Silicon Valley Tax Directors Group's central illustration is a company with employees in all 50 states, billed at its California headquarters, whose platform "is used exclusively by software engineers located in Texas." Their point, a good one, is that use follows the engineers rather than the invoice.

• COST asked the Department to consider a multiple-points-of-use certificate, citing Washington's formula: "the number of users in this state divided by the number of users everywhere."

• The Silicon Valley Leadership Group came closest and stopped just short, writing that location of use "is not determined solely by the customer's billing address, contract execution location, or server location, except to the extent those facts reliably establish where the software is used." Server location survives there as a fact that might sometimes be reliable. Draft 1600.2(d)(4) then removed it as a category.

• Salesforce filed for the same seat-based mechanism, in a letter that describes its own products as ones that "bring autonomous agents, unified data, analytics, and sector-specific applications together." The company that sells agents asked California for a headcount rule, because headcount is what its customers' finance systems can produce.

That last one is not hypocrisy; it is the shape of the whole record. Everyone who showed up is a business whose software is used by staff, so the rule they asked for measures staff, and the Department wrote what it heard. A rule for workloads without users was not rejected. It was never proposed.

What Our Engine Does Today, and What It Cannot Do

A $1,000 SaaS charge to a California buyer returns $0.00 today and $72.50 at 7.25% for a transaction dated after January 1, 2027, when S.B. 122 takes effect. A $1,000 compute or cloud-infrastructure charge returns $0.00 in 2027 as well - the position draft Regulation 1600's IaaS/PaaS exclusion confirms rather than merely permits. An API access charge returns $0.00 in 2027 with a classification advisory attached, flagging that if the product is really prewritten software rather than raw infrastructure, it is taxable and should be sent as SaaS.

Two honest limits. We carry no California ZIP-level district rates, so a San Francisco transaction returns the 7.25% state rate with a ZIP_UNKNOWN advisory rather than the higher combined rate - that gap is ours, not California's. And we compute one jurisdiction's tax on one transaction; we do not apportion a charge across states, because no engine can apportion by a method the state has not adopted. If your workload has no seats, there is no compliant denominator to compute with.

What to Do Before September 24

If you operate agents sold into or used in California in 2027, the comment window is the cheapest lever you will get. Once the Department submits to the Office of Administrative Law, the questions in the file are the questions that exist.

• Ask the sourcing question in plain terms. Comments go to PPDD-BTC.InformationRequests@cdtfa.ca.gov by September 24. Ask whether a purchaser whose use of a digital product is executed by automated processes rather than by authorized employees may use a usage-based or computer-based denominator, and how (d)(4) applies when servers are the only locus of use.

• Read 1600.2(d)(2) before assuming you are stuck. The permitted methods include "the number of users or computers," and counting client computers that access a product is not the same act as locating the servers where software is installed. That is a reading, not a stated rule, and it sits beside a prohibition an auditor could stretch over it - which is why it should be asked out loud rather than relied on quietly.

• Decide now whether your product is infrastructure or software. The draft treats them very differently, and that boundary is where most agent products sit.

• Do not read Example 2 as broader than it is. It decides custom versus prewritten, and nothing about sourcing.

We are not a filer here and have no standing to be. What we can do is say plainly that the record has a hole in it, three days before the meeting, while it can still be filled.

For where agent activity sits in each state's tax base, see our AI agent sales tax guide; for how the figures above are derived, see our tax methodology. California would be the fourth US jurisdiction running a software apportionment regime, after Michigan's repealed experiment, Chicago's operating one and Massachusetts. None of the four counts machines.

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.