AI for tax workpapers

A workpaper has to do two jobs. It has to show the number, and it has to show why the number is that. Most software aimed at this problem is built for the first job, demos well on the first job, and quietly leaves the second one to the preparer, which is where the time actually goes.

The three things it can do, in increasing order of difficulty

Extraction, which is largely solved

Reading a W-2, a 1099, a K-1 or a brokerage statement into structured values is no longer the hard part, including on scans and photographs. Any serious tool will do this well on clean documents and acceptably on bad ones. It is table stakes and it is not a reason to choose anything.

The distinguishing question inside extraction is not accuracy on a good scan. It is behaviour on a bad one: does the tool say it could not read the box, or does it return a plausible number with nothing marking it as a guess?

Assembly, which is mostly template plumbing

Putting extracted values into your firm’s own schedule, in your own format, with your own subtotals, is genuinely useful and mostly engineering rather than intelligence. The thing to check is whether it fills your template or its own. A tool that produces a beautiful workpaper in its house format has moved the work rather than removed it, because somebody still has to transcribe it into the format the file actually uses.

Tie-out and review, which is where tools differ

This is the job. A reviewer looking at a depreciation schedule needs to get from a line on that schedule to the document and the box that line came from, in one step, without searching. If that step does not exist, the reviewer does the tie-out manually, and the tool has automated the typing while leaving the expensive part untouched.

It matters beyond convenience. Circular 230 section 10.35 requires competence in the matter, and section 10.36 puts an obligation on whoever has principal authority over the firm’s tax practice to ensure adequate procedures are actually in place. Neither is satisfiable over output nobody can check.

The test to run in a demo

Bring your own document, ideally a bad scan of something ordinary, and ask for four things in order.

• Read this. Watch what it does with the boxes it cannot make out.
• Put it in this template. Yours, not theirs.
• Now show me where that figure came from. It should land on the page and the box, not open the document at page one.
• Now show me who accepted it. A name and a timestamp, per figure.

Most tools are strong on the first two and vague on the second two. The vagueness is the answer.

What it should not do

Two boundaries worth writing into your own evaluation, because software will cross them if allowed.

It should not decide a position. Recording what a document says is not the same as choosing a treatment. Transcribing box 1 of a W-2 is bookkeeping; deciding whether an expense is ordinary and necessary is practising. A tool that blurs the two has taken on a judgement that belongs to a person who signs.

It should not resolve a conflict quietly. When a new document disagrees with something already in the file, no confidence score should settle it. Being sure of a reading says nothing about which of two readings should win, and the facts where two sources disagree are precisely the ones a reviewer most needed to see.

That second point is the one most worth pressing a vendor on, and it is covered at length in the piece on what human in the loop actually means.

Where the time actually goes

If you are evaluating this to save time, it is worth knowing which time. Extraction saves typing, which is real but is the cheapest hour in the engagement. Assembly saves formatting. The hour worth recovering is the reviewer’s, and that hour is only recovered if the review can be done from the workpaper itself rather than by opening the source documents alongside it.

Before any of it, the documents have to be asked for, and there is a free request list builder here that will produce the ask for a 1040 or an 1120-S.

Which is why the tie-out question is not a nice-to-have feature comparison item. It is the one that determines whether the rest of it was worth buying. Before any of it happens, though, the documents have to arrive, which is its own problem.