An AI drafted a tax return from a W-2. Now price your work

In the launch material for GPT-6 Astra on 3 September 2026, OpenAI showed the model filling in a federal tax return in a browser, working from a W-2. Not summarising the rules. Not explaining what a W-2 is. Opening the form and completing it.

OpenAI called the output a draft, which is the honest word for it. But every firm owner who saw that demo did the same arithmetic in their head, and it is worth doing that arithmetic properly rather than in a panic at two in the morning during season.

What that demo actually automated

Be precise about the task, because the precision is the whole answer.

The model took a structured source document with known fields in known positions, and moved those fields into a form with known positions. It is the most automatable task in tax, and it has been getting automated in pieces for twenty years through OCR, direct payroll feeds and pre-population from the revenue authority. A single-W-2 return is the simplest return the US system produces.

What the demo did not show, and what nobody has yet demonstrated at professional standard, is any of the following: deciding which of two defensible positions to take, knowing that a client's side business has changed character, spotting that the numbers in front of you are wrong because you know the client, or signing a document and accepting the consequences of it being wrong.

So the correct summary is narrow. A cheap machine got better at the cheapest part of the job. That has happened before in this profession, repeatedly, and the firms that got hurt each time were the ones that had accidentally been selling the cheap part.

The mistake that is about to be very common

The mistake is straightforward: work out how much time the automation saves, and reduce what you ask for by roughly that much.

It feels honest. It is how a lot of good operators think. And it hands your entire productivity gain to the client in exchange for nothing, permanently, on the assumption that what you sell is hours.

The reason it is a mistake has nothing to do with greed. It is that the thing you were paid for was never the keystrokes. When a client pays for a return, they are buying three separate things that happen to arrive in one invoice:

• Keystrokes. Getting the numbers from where they are into where they need to be. This is the part machines have been eating for two decades, and Astra just took another bite.
• Judgement. Which treatment, which election, what this transaction actually was, what the client failed to mention. This requires knowing the client and the law, and it is not what the demo showed.
• Risk. Someone with a registration, a professional body and insurance puts their name on it. If it is wrong, they answer for it. No model carries this and no vendor has offered to.

If your engagement has quietly become mostly the first item, the machine is a genuine threat and cutting what you ask for will not save you, because the machine will always be cheaper. If it is mostly the second and third, the machine is a margin improvement and there is no reason for the number to move at all.

Most firms genuinely do not know which of those describes their compliance work. That is the actual problem this release exposes.

How to find out, using files you already have

You do not need a consultant or a framework for this. You need last season and an honest afternoon.

Take twenty completed returns across your normal spread. For each one, split the recorded time into the three buckets above. Not estimated from memory. Reconstructed from the file: what did the notes, the queries and the review comments actually concern?

You are looking for one ratio. What share of the total went into moving data around, versus deciding things and being accountable for them?

Firms that run this exercise are usually surprised twice. The keystroke share is higher than the partners believe, and the judgement work is concentrated in a small number of jobs that were never separated out or recognised in what the client was asked for. Both findings are useful, and neither is visible from a report on average job profitability.

Then decide what you are actually selling

Once you know the ratio, the decision in front of you is a positioning one and only you can make it. There is no benchmark to copy and this article is not going to give you a number, because a number would be worthless: your clients, your risk, your market and your capacity are not the ones next door.

What can be said is what the options are.

You can hold the total and change what sits inside it, moving the recovered hours into review, planning conversations or advisory work the client can actually feel. This is the option most firms say they want and few execute, because it requires the client to see something new rather than simply be told the work got faster.

You can unbundle, and separate the mechanical return from the thinking, so the client can see there are two products and choose. This is the most honest option and the most administratively painful, and it is the one that survives contact with a client who has read about AI.

You can compete on the mechanical work deliberately, at volume, with the automation doing the heavy lifting. That is a real business and some firms will do very well at it. It is a different business from the one most practices are currently running, and drifting into it by accident is how firms end up busy and unprofitable.

What you cannot do is nothing, and describe that as a decision. The tooling around pricing decisions in a firm is worth having in place before the conversation lands, not after.

The part that does not get cheaper

One thing is fixed regardless of what any model can do. A return prepared for a client is signed by a person, and that person carries the obligation.

In the US, a paid preparer works under Circular 230 and the due diligence expectations attached to it. Nothing about a model completing the form transfers any of that. In Australia, the Tax Practitioners Board has already said the Code follows the work rather than the tool, and the equivalent principle applies wherever you are registered. The specifics of what US firms have to observe when AI touches a return are worth reading alongside this. None of this is legal advice, and your professional body is the source of truth for your own obligations.

This matters commercially, not just as compliance. Risk-carrying is a service, it is the one part of the engagement no software company has offered to take on, and firms consistently fail to name it as something the client is paying for. A client who does not know they are buying your signature will assume they are buying your typing, and will price you accordingly the moment typing gets cheap.

What to do before the next season

1. Run the twenty-file split. Nothing else on this list works without the ratio.
2. Name the three components in your engagement letter, so the client has seen them written down before anyone renegotiates anything.
3. Test the automation on returns you have already completed and know the answer to. You are measuring your review time, not the model's speed.
4. Work out what a review-only engagement looks like for the simplest clients, because some of them will arrive next season with a return already drafted by something.
5. Decide the positioning question deliberately, at partner level, and write the decision down. A decision nobody recorded is a drift.

The demo was a genuine milestone and also a narrow one. Watching a machine complete a form from a W-2 tells you very little about what your firm is worth. Twenty of your own files will tell you almost everything. More of our coverage on where this is heading sits on the AI for accounting firms hub, including what stays with the practitioner when AI handles the tax work.

Frequently asked questions

Can AI actually prepare a tax return now?

It can produce a draft from a structured source document, which is what OpenAI demonstrated with a W-2 on 3 September 2026. That is not the same as preparing a return to professional standard, which involves establishing the facts, choosing treatments, checking for what the client did not mention, and signing. Treat it as a fast first pass that a qualified person still has to work through.

Should I reduce what I ask clients to pay because AI made the work faster?

Not automatically, and not before you know what proportion of the engagement was ever mechanical. Reducing the total on the assumption you sell hours converts a productivity gain into a permanent giveaway. Work out the split between data entry, judgement and risk in your own files first, then make it a deliberate positioning decision.

What do I tell a client who says an AI could do their return?

Agree with the part that is true, because arguing with it damages your credibility. Something can produce a draft. Then be specific about what they are actually buying from you: someone who knows their circumstances, chooses between defensible positions, and carries the consequence of the return being wrong. If you have never separated those things out in writing, the conversation is much harder.

Does using AI on a client return create a compliance problem?

It creates obligations rather than a prohibition. Registered practitioners remain responsible for the output regardless of what produced it, confidentiality still applies to whatever you put into the tool, and several jurisdictions now expect disclosure to the client. Check your own professional body's current guidance rather than a general article, and get advice if anything about your situation is unusual.

Which clients are most exposed to this?

The ones whose returns are simple, repetitive and driven by a single structured document. They were always the least profitable and the most substitutable, and they are the group most likely to try a machine first. The useful response is knowing what that segment contributes now, so you can decide whether to defend it, reprice it or let it go, rather than discovering the answer when they leave.