What An AI Actually Did On A Real File

The deadline was five days away.

A voluntary disclosure to Inland Revenue. Two tax years to reconcile. A penalty argument to make. A formal letter on letterhead and tax info pack sent by a fixed date. The kind of file most accountants and many tax-aware lawyers would recognise.

A job like this would normally be given to a junior to research, draft and package, then partner review and amendments, the whole process usually takes about a week to finalise. Last week, with Claude the time it took was closer to 2 hours.

The interesting part is not the time saved. It's what got compressed and what didn't.

For those of you who don't know me well, I'm also a chartered accountant and run an accountancy practice alongside Law Firm Profit Success. And for those of you who haven't come across Claude before, it's an AI assistant from a company called Anthropic, similar to ChatGPT.

What Claude actually did

Claude built the reconciliation tables for both tax years. Compared previously filed tax returns with the new tax position, updated share scheme income, dividends, and the additional core tax to pay. It cross-checked the bracket position of the revised taxable income against the personal tax thresholds. It ran a second independent method on the core tax number, using additional income times the marginal rate less new credits, and reconciled to the same figure to the cent.

It surfaced the right sections of the Tax Administration Act. Section 141G(3)(a)(i) for 100 percent reduction in shortfall penalty. Section 183A for remission of the late payment penalty and section 183D for remission of use-of-money interest, the only path under the Act for that. It mirrored the structure and wording of two precedent disclosure letters from prior years, so the new draft read in our house style rather than as a fresh assembly.

It flagged the verification steps to run before send. Reconfirm the figures against the most recent income summary. Reconfirm the prior reported numbers from the original signed returns. Spell-check the letterhead, the IRD number, the date, the email and the phone.

That is a lot of structured work. None of it requires judgement. All of it has to be done. All of it has to be right.

What stayed with me

Every judgement call stayed with the team.

There was a question about whether an IRD officer's phone call counts as notification of a pending tax audit or investigation, which would mean imposition of shortfall penalties. Claude could lay out the test. It could not decide. The call we make is that a phone call from IRD's Customer Compliance team inviting a voluntary disclosure is not the same thing as being advised of a pending audit. Different function, different purpose. That is a position we are willing to take and defend. An AI can suggest it. A human has to own it.

There was a question of scope. The dividend in the earlier year had not yet been picked up by IRD. We could have left it out. We chose to bundle it in, on the view that voluntary disclosure of a known omission carries weight, and that selective disclosure does not. Claude helped us think it through. The decision was ours.

There was a question of how to frame the background. The client had personal circumstances during the relevant period that mattered to the explanation. The temptation in this kind of writing is to lean into excuse. The right tone is explanation, not defence. Choosing that tone is judgement.

The lesson sitting underneath

In March I wrote about how AI is taking over the structured work in legal services while the judgement work stays with humans. Last week's file is the cleanest version of that argument I've seen in my own practice.

The structured work compressed fast. The judgement work did not compress at all. The hours saved came from table building, citation pulls, precedent mirroring, and verification steps. The hours spent went into the things only a human can do. Reading the room with the regulator. Choosing scope. Picking the tone. Owning the position.

That is the shape of the next ten years of professional services work in my view. The leverage on structured output goes up sharply. The premium on judgement goes up too. The middle layer, the part where you charged a client for assembling the table or pulling the citation, gets squeezed out of the bill.

What this means for a boutique firm

A boutique firm can pilot this on a real matter this quarter. Unlike a large firm, there's no need for an internal review committee, no procurement cycle. Just one partner, one file, one afternoon.

The starting point is not the tech, the starting point is documented work. The cleaner your precedents, your SOPs, and your prior matter records, the more an AI can do for you. Firms that have already done the quiet structural work I wrote about previously are the firms that get the AI dividend first.

If you are sitting on a file this week that is structurally heavy and judgement-light at the edges, that is the file to try this on.

More next Tuesday,

Gordon