The Last Mile Problem Every tax automation demo ends at the moment the document is read — and that's the cheap part. Why the expensive stretch of a tax return is the last mile no vendor demos, and the four steps real automation has to do in order.

Every tax automation demo I have ever sat through has the same shape.

Someone drops a W-2 onto a screen. A progress bar moves. Field values appear in a tidy panel — wages, withholding, the employer's EIN — and everybody in the room nods, because it worked. It read the document. That is the moment the demo is built around, and it is genuinely impressive the first time you see it.

Then the demo ends.

What nobody shows you is the next forty minutes. Because reading the document is not the job. The job is a finished return, and between "we have the document" and "the number is in the return" there is a stretch of work that no vendor puts on a slide. In logistics they call this the last mile — the short final leg that costs more than the entire journey before it. Tax has one too, and I would argue it is where most of a firm's August actually goes.

Four steps, and everyone sells you the first one

Real automation of a tax document has to do four things, in order.

Extract. A machine reads the document and pulls the field values off the page. This is the step everyone demos, and it is the step that has genuinely gotten good.

Validate. Those values get checked against the things that are checkable — is this the right shape for an EIN, does this box exist on this form, could this number even be right given the others around it. Structural sanity, not judgment.

Transfer. The data has to end up inside the firm's own tax software. Not in a panel next to it. Not in a spreadsheet the firm exports and re-imports. Inside the software where the return is actually being built.

Verify. A licensed human confirms the fields against the source before any of it is trusted on a return that will carry their signature.

Most tools do step one and describe it as automation. Step two shows up in some of them. Step three is where the marketing goes quiet, and step four gets handed back to the firm as though it were a courtesy — of course you'll want to review everything — when in practice it is being handed back because there was never a plan for it.

Here is the uncomfortable arithmetic. If a tool extracts perfectly and then leaves the transfer to a human, it has not saved that human the typing. It has moved the typing to later in the day and added a screen to look at. The firm now has an extra tool, an extra login, an extra vendor invoice, and the same person typing the same numbers into the same software at four in the afternoon.

Why the last mile is genuinely hard

It is worth being honest about why this step gets skipped, because it is not laziness.

Tax software was not designed to receive data from anything except a person. There are import paths, and they exist, and they are the obvious thing to build against. Generate a structured file, hand it over, let the software swallow it whole. On a whiteboard it is clean. In a demo it is beautiful.

It is also fragile in exactly the way that hurts most. Those formats are private, undocumented, and free to change. The software updates mid-season — as it must, because the tax law moved. The integration that worked flawlessly in January stops working in March, and it usually stops working quietly. In tax software, an integration that breaks silently in February is not a bug. It is a season.

The alternative is slower to build and much less elegant to describe: put the data in the way a person puts the data in — through the software's own interface, field by field, at machine speed. It does not depend on a private file format holding still. And when it fails, it fails the way a person fails: visibly, on the field it is on, where somebody can see it and fix it.

That is the path we took with Sophicor Forge, and I will admit it was not the fun decision. It was the one that survives a February.

The step we deliberately did not automate

There is a fifth thing that could be automated, and we chose not to.

After extraction, there is a review screen: every field that came off the document, waiting to be confirmed. The efficient version of that screen is a tidy list with an "approve all" button at the bottom. We built that version first. It was fast, and it was wrong, and it took us longer than it should have to see why.

A list of numbers with no document beside it does not get verified. It gets approved. We had accidentally built the thing we do not believe in.

The version we kept puts the source document in front of the person, field by field, so that confirming actually means looking. It is slightly slower. That is the point. The speed we are offering lives in the extraction and the transfer — the freight. The middle step is supposed to cost you the two seconds it takes to be sure.

I understand the commercial pull in the other direction. "Fully automated" is a better line than "automated, with a human check you cannot skip." But a confidently wrong number on a tax return is worse than a slow right one, and the signature at the bottom of that return belongs to a person, not a model. If verification is easy to skip, it is not verification.

What this means for a firm in the last three weeks of August

There is a practical version of all this, and September 15 makes it urgent.

That date carries two deadlines at once. Calendar-year partnerships and S corporations that timely extended have to file — Forms 1065 and 1120-S — and there is no second extension behind them. The third estimated tax installment for 2026 is due the same day. Both are on irs.gov if you want to check me.

Pull your open list this week and sort it, not by client, but by what is blocking each return. You will usually find that most of what is left is not judgment. It is assembly. Documents arriving in six formats. Numbers moving from a PDF to a screen by hand. Somebody checking that the moving was done correctly. That pile is the last mile, and it is the part of the work that a machine can genuinely take.

The part it cannot take is the reason clients pay you. Every profession that has been automated lost the transcription first and kept the judgment. Medicine did not lose the physician to software; it lost the dictation backlog to speech recognition. Law did not lose the lawyer; it lost the room of associates reading boxes for a keyword. Tax will be no different, with one wrinkle: because we sign returns, the freight and the judgment have always been billed as a single thing. They are not a single thing. Separating them is the opportunity, not the threat.

Reading is table stakes now. The last mile is the product.

Sources referenced: IRS Third-Quarter Tax Calendar and estimated-tax guidance — calendar-year partnership and S-corporation extended returns (Forms 1065 and 1120-S) and the third 2026 estimated-tax installment due September 15, irs.gov.