Multiple Points of Use: The Massachusetts Apportionment Rule California Is Copying, Read Against AI Agents

At its July 21 implementation workshop the CDTFA said it is evaluating apportionment approaches for S.B. 122 modeled on other states, Massachusetts among them, keyed to user or device location. We reported that signal and objected to the denominator. What we did not do — and what is worth doing with a discussion draft expected inside two weeks — is read the Massachusetts rule itself. It is the country's most developed multiple-points-of-use regime, it has been litigated to the Supreme Judicial Court, and it contains one provision that decides the question for agent operators before the arithmetic starts.

What Massachusetts taxes, and at what rate

Massachusetts treats prewritten computer software as tangible personal property under M.G.L. c. 64H regardless of the method of delivery. The implementing regulation, 830 CMR 64H.1.3, is explicit that the taxable transfers include licenses and leases, electronic delivery, load-and-leave, and — at (3)(a) — transfers of the right to use software installed on a remote server. The rate is 6.25 percent, with no local layer.

The regulation also disposes of the argument operators reach for first. Describing a product as software-as-a-service does not determine its taxability, and neither does the fact that the customer never downloads or installs anything; Letter Rulings 12-8 (cloud computing) and 12-10 (screen-sharing software) apply that reasoning to hosted offerings. Custom software developed to a purchaser's unique specifications remains exempt as a professional service transaction, but that exit is narrow — a hosted agent tool sold to many customers is not custom software because each deployment is configured differently.

For an agent operator buying model APIs, vector stores, and hosted tools, the starting assumption in Massachusetts should be that the purchase is taxable prewritten software — a well-supported position. The interesting question is not whether. It is how much of it belongs to Massachusetts.

The apportionment mechanics

830 CMR 64H.1.3(15) supplies the answer, and it is more generous than most states. Where software will be concurrently available for use in more than one jurisdiction, the tax may be apportioned. Practitioner summaries read the subsection as offering three routes:

Purchaser-initiated. The purchaser delivers a multiple points of use certificate — Form ST-12, the Exempt Use Certificate, which carries an MPU designation — to the vendor no later than the time the transaction is reported for sales or use tax purposes. On receipt the seller is relieved of the obligation to collect, and the purchaser remits the apportioned tax directly.

Vendor-initiated. Where the vendor knows the software will be concurrently available in more than one jurisdiction but has received no certificate, it may work with the purchaser to produce the correct apportionment. The purchaser certifies its accuracy; the vendor remits on the apportioned figure.

Pay and abate. Pay on the entire transaction and pursue apportionment through the general abatement process under G.L. c. 62C.

The method itself must be "any reasonable, but consistent and uniform, method of apportionment" supported by the purchaser's books and records as they exist when the transaction is reported. The regulation offers two illustrations: the number of licensed users in each jurisdiction, or the number of computers or terminals in each jurisdiction. And it draws one hard line — apportionment may not be based on the location of the servers where the software is installed.

The line that decides it

Set that against an agent operator's actual facts.

Licensed users: one, or zero, depending on whether the service account counts. Computers or terminals per jurisdiction: none in any sense the regulation contemplates. The workload runs in ephemeral containers a scheduler placed wherever capacity existed, and a config change relocates next week.

Server location: known precisely, logged continuously, and the one input the regulation forbids.

This is worse than a poor fit. Massachusetts permits two denominators the agent operator does not have and prohibits the only geography it produces naturally. The prohibition is not arbitrary — server placement is an artifact a taxpayer could shop for, and a rule keyed to it would let any purchaser move its tax base by changing regions. That reasoning survives contact with agent commerce intact. The consequence is simply that the operator is left without a permitted input.

The gap is not fatal. The regulation's standard is "reasonable, consistent and uniform," and the two illustrations are illustrations rather than an exhaustive list. An allocation keyed to where the economic benefit lands — the principals the agent transacts for, the jurisdictions of the end customers served during the period — is arguable on the regulation's own terms. I would treat that as an open argument, nothing more: it is untested, it requires records no one is currently keeping, and the Department has never addressed it.

Whether an agent operator qualifies at all

There is a prior question that cuts the operator's way. The trigger in (15) is concurrent availability, not measured use. A hosted tool an operator can invoke from any region is concurrently available in every jurisdiction the operator runs in, whether or not it was invoked there last month — an easier fact to establish than a user count, and one that requires reconstructing nothing.

So the eligibility prong is satisfiable and the method prong is where the exposure sits. That is a useful diagnosis: the compliance work is records, not argument.

Why the abatement route is the one that fits

The two certificate routes share a timing assumption: the purchaser knows its multistate footprint by the time the transaction is reported. An agent's footprint is not knowable then. It is realized over the billing period, as the agent decides which tools to call, how often, and for whom. We made this point about California's proposed real-time exemption certificate on August 1; it applies with equal force to Form ST-12.

Massachusetts has a backstop California does not yet have. In Oracle USA, Inc. v. Commissioner of Revenue, 487 Mass. 518 (2021), the Supreme Judicial Court held that the right to apportion is statutory, and that a taxpayer who did not follow the Commissioner's regulatory procedure may still pay on the entire transaction and seek apportionment through the general abatement process. Strict compliance with the regulation is not a precondition.

For an agent operator that is the mechanically correct path, and it should be the planning default in Massachusetts: pay on the full purchase price, keep contemporaneous per-jurisdiction records through the period, seek abatement once the denominator exists. It converts an up-front certification the operator cannot honestly make into an after-the-fact showing it can. Whether the Department accepts an agent-derived allocation as reasonable, consistent and uniform is the open question — genuinely unsettled — but the procedural route to ask is settled.

What travels to California

If the CDTFA copies Massachusetts, it inherits the server-location prohibition along with the user-and-device illustrations, and California's 2027 regime arrives unusable by agent operators on day one. That is a comment-period point, not a complaint. The discussion draft is expected by the end of August 2026 and had not circulated as of this morning, with an interested-parties meeting in late August or September. The ask is narrow: permit an allocation keyed to the location of the persons for whom the software is used, alongside users and devices, and provide a back-end refund path for purchasers whose footprint is realized after the reporting date. Massachusetts supplies the second half already, through Oracle. California would have to write it.

Maryland is the third data point, and we covered its Multiple Points of Use certificate in the buyer-type analysis four days ago. Three states, three certificate regimes, one shared assumption about what a purchaser can count.

What to watch

The CDTFA discussion draft remains the top item and preempts everything else when it lands. Beyond it: whether Massachusetts issues guidance applying (15) to subscription and consumption-priced software, the form nearly all agent inputs take and one the regulation's license-count framing does not describe; and whether Colorado's HB 26-1223 regime, also arriving January 1, 2027, includes an apportionment mechanism at all. A taxable base with no apportionment route is worse for multistate operators than a badly fitted one.

No state has issued guidance addressing agent-initiated commerce directly. Every determination above applies a human-drafted category to a machine transaction, which is analysis rather than settled law.

AgentTax computes Massachusetts at the 6.25 percent state rate with no local layer. It does not currently model multiple-points-of-use apportionment in any state, Massachusetts or Maryland — a calculation returns the full-price figure, which is the correct amount to remit before an apportionment claim and the wrong amount to treat as final if your workload spans jurisdictions. That gap is with our engineering team, and is disclosed here rather than left to be discovered. See the per-state logic at agenttax.io, or start with our AI agent sales tax hub.

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.