Can AI Work Out Where You Owe Sales Tax? Not Quite

Ask any assistant whether you need to collect sales tax in Illinois and you will get a fluent, structurally correct answer about economic nexus thresholds. It will be a reasonable summary of the general rule and it will not tell you whether you owe anything, because that turns on your sales into that state over a period you have not told it about, on what your product legally is, and on whether the marketplace you sell through is already the seller of record.

This is the clearest example in small business of a question that looks like a knowledge question and is actually a data question. Worth being precise about, because the failure mode is expensive and delayed: you find out two years later, with interest.

What did the law actually change?

One rule, in one case, and the specifics are narrower than the folklore. South Dakota v. Wayfair, argued 17 April 2018 and decided 21 June 2018, held that the physical presence rule from Quill was unsound and incorrect, and overruled it.

The thresholds in the South Dakota statute matter because everyone copied them. That law reached only sellers delivering more than 100,000 dollars of goods or services into South Dakota in a year, or hitting 200 or more separate deliveries there. It also foreclosed retroactive application. Those two numbers became the template for most states that followed, which is why so much writing on this subject quotes them as if they were a national rule.

They are not, and the drift since is the whole practical problem. California's threshold is 500,000 dollars of combined sales of tangible personal property for delivery into the state, with no transaction count, and once you pass it you also have to collect district use tax for every district that imposes one. A seller reasoning from the 100,000 and 200 figures would be wrong about California in both directions: wrong about the amount, and unaware that crossing it brings local district obligations rather than a single state rate.

Why is product taxability harder than nexus?

Because nexus is arithmetic and taxability is definition. Once you know your sales into a state, the threshold question has a yes or no answer. What rate applies to what you sell is a question about how that state defines your product, and the definitions genuinely differ.

This is not folklore either. The Streamlined Sales Tax Governing Board publishes a state taxability matrix precisely because the treatment varies. The matrix identifies the definitions and tax administration practices adopted by the board that each member state must follow, and it was split in June 2021 into a library of definitions and a set of tax administration practices. Section C covers product definitions and records, per state, whether the state adopted the shared definition and whether the item is taxable or exempt, with a reference to the law or rule that says so.

Read that structure and the implication lands. There is a published, state by state answer to whether your category is taxable, maintained by the states themselves, and it exists because the answers disagree. Any tool that gives you one answer for a product across a country is either wrong or quietly consulting the same matrices and not telling you which.

The categories where this bites hardest are the ones with everyday names and technical definitions. Food and candy. Clothing and protective equipment. Digital goods, software and anything sold as a subscription. Dietary supplements. Kits and bundles containing a taxable item and an exempt one. If what you sell sits in one of those, the definition question is your primary risk and the threshold question is secondary.

Where does a marketplace change the answer?

Where it becomes the seller of record, which shifts the collection duty off you for those sales and leaves your own direct sales exactly where they were. This is the single most common source of confusion for a seller running both.

The practical consequence is that your threshold arithmetic has to be done on the right numerator. Depending on the state and the channel, marketplace sales may or may not count toward the threshold that determines whether you must register for your direct sales. That is a per state question with a documented answer, and it is the sort of question that a model will answer with a generalisation because the general pattern exists and the exceptions are the point.

So do the arithmetic twice. Once on total sales into a state, and once on direct sales only. If the two answers differ across a threshold, you have found a question worth paying somebody to answer properly, and you have found it before a state did.

What is the equivalent question in the EU?

Not whether you are registered, but which scheme you file under, and the answer is structured rather than threshold based. The One Stop Shop lets you declare and pay VAT for cross border consumer sales through a single member state instead of registering in each destination.

Three schemes exist and the differences are worth holding. The Union scheme covers intra community distance sales of goods and services to consumers in member states where you have no physical presence. The non Union scheme covers suppliers established outside the EU supplying services to consumers inside it. The import scheme covers distance sales of goods imported from outside the EU not exceeding 150 euro per consignment, excluding excise duty products.

Two operational details from the same source matter more than they look. Union and non Union returns are quarterly, due at the end of the month after each quarter, on 30 April, 31 July, 31 October and 31 January. Import scheme returns are monthly. And nil returns are required even in a period with no supplies, which is the rule that catches seasonal sellers who assume a quiet quarter needs no filing.

The 150 euro consignment ceiling on the import scheme is the one to design around if you ship from outside the EU, because a single order crossing it leaves the scheme and lands in ordinary import procedures, with the duty and clearance consequences that follow. The wider mechanics of that are in what changes when an order crosses a border.

So what is the machine genuinely for?

Three jobs, none of which is producing the answer. Turning your order data into the arithmetic, building the calendar, and reading the page you found so you do not misread it.

The arithmetic job is the most valuable and the least discussed. You need sales by destination state, by rolling twelve months and by calendar year, with a transaction count alongside the value because some states still use both and some use neither. That is a grouping exercise over data you already have, and it is exactly the sort of work where asking a tool to write the query beats asking it for a conclusion. This is the same distinction that makes asking for the formula rather than the answer the right habit in a spreadsheet.

The calendar job is unglamorous and prevents the most common failure. Registration in a state creates filing obligations that continue whether or not you sell there next quarter, including nil returns. A list of what is due when, with the registration date and the frequency, is a document a model can draft from your own facts in minutes, and it is the difference between a compliance cost and a penalty.

The reading job is the narrowest and still worth it. Paste the revenue department page you found and ask what conditions attach to the threshold, what the measurement period is, and what happens in the first period after you cross it. Those three details are where plain English state guidance is genuinely hard to parse, and a summary of a page you supplied is a much safer request than a question answered from memory.

What happens in the first period after you cross?

Usually not what people assume, and it is the detail most summaries omit. Crossing a threshold rarely makes you liable for everything you already sold. It starts an obligation from a point the state defines, and the point differs.

Three patterns exist across states. Some begin the obligation on the first day of the next transaction after you cross. Some begin it at the start of the following calendar month or quarter, which gives you a short administrative runway. Some measure against the previous calendar year rather than a rolling twelve months, which means your obligation for this year was already fixed by last year's sales and nothing you do now changes it.

The third pattern is the one that catches growing sellers, because it inverts the intuition. You can be under the threshold today, have been over it last year, and owe collection this entire year. A seller checking their current trailing figure would conclude they are safe and be wrong for twelve months. This is why the measurement period is the question to look for on the state page, not just the number.

The retroactivity point from the Wayfair opinion is worth keeping in view here too. The South Dakota act specifically foreclosed retroactive application of its requirement, and the Court noted it. That was a feature of that statute rather than a constitutional guarantee for every state, so the safe assumption is that the answer to what you owe for past periods is a question for an advisor rather than one you should resolve from a summary.

Does the same reasoning apply to digital products?

It applies more sharply, because there is no shipment to anchor the location and the definitions are newer and less settled. A downloadable file, a subscription to software, a recorded course and a live class delivered over video can carry four different treatments in the same state.

Two facts decide most of it. What the state considers the product to be, which is the definition problem again, and where the customer is deemed to be located, which for a digital sale is whatever address or billing evidence you collected. The second is a checkout design decision made long before the tax question arrives, and a shop that never captured a billing country has made the determination impossible after the fact.

So for digital sellers the practical work is upstream. Capture and store the evidence of customer location at the point of sale, keep it with the order rather than in a payment provider you may leave, and record what you sold in the terms the definitions use rather than in your own marketing language. A line item reading "annual plan" tells you nothing two years later. A line item recording that this was access to hosted software, sold to a consumer, billed to a specific country, answers the question on its own.

That record keeping discipline is the same one that makes any later audit survivable, and the same reason to keep your own order data rather than relying on a channel's reports. The wider version of that argument is in what moving product and order data between systems actually costs, since the exports you can get are what your records eventually consist of.

Which answers should you never accept from a model?

Any current threshold, any rate, and any conclusion about your own obligation. Each fails for a different reason and all three fail quietly.

A threshold recalled from training data may be a year out of date, and a wrong threshold does not look wrong. It looks like a number. Ask instead for the page where the state publishes it, then read that page and record the date you read it, because your file needs to show what you relied on and when.

A rate is worse, because rates vary by local district within a state, as the California district use tax requirement shows, and they change on schedules that have nothing to do with anything. Rates belong in a service that updates itself and sits in your checkout. No generated figure should ever reach a customer's invoice.

A conclusion about your obligation is the one that feels most helpful and carries the most risk. A model asked whether you need to register somewhere will answer, because the question is well formed and the pattern is familiar. It cannot know your sales, your channels, your product classification or your registration history. Pattern matching in place of missing facts is exactly the failure described in why language models produce confident wrong answers.

What does a sensible routine look like?

Quarterly, and it takes an hour once the query exists. Pull sales by destination for the trailing twelve months, sort descending, and look only at the states approaching or past a threshold. Most sellers have three or four that matter and thirty that do not.

For each state in the first group, check the threshold on the state's own page rather than a summary site, note the date, and record whether your figure is direct sales or total. If you are within a comfortable margin, note it and move on. If you are close, that is the point to get advice, because registering voluntarily before you cross is a cheaper conversation than registering afterwards.

Then look at the shape of your growth rather than the current number. A category with a definition problem plus fast growth into new states is the combination that produces a real bill, and it is visible a year ahead in your own data. Reading your own numbers for that signal is the same exercise as forecasting cash on a short trading history, and it fails in the same way if the underlying records are messy.

One honest limit on all of this. Nothing here replaces an accountant for a business genuinely crossing thresholds in several states, and the reason is not complexity, it is liability. The value of the routine is that it tells you when you need that conversation and gives you the numbers to walk in with. If you are choosing the systems that hold those records in the first place, the export and audit questions matter as much as the reports, which is part of why we publish what a build costs rather than making the data you can get out of it depend on a plan tier.