Are you the human in the loop?

We are coming to expect the presence of AI in daily tasks. For example, let's consider an email. People use AI to draft replies, tidy up long messages and pull the actions out of a thread that has run to 20 messages, and that's a good thing.

What we're also seeing is the round trip. One person asks AI to write an email. The person receiving it asks their own AI to summarise it and draft a reply, then sends that back. Two people have signed off on the exchange. How much of it did either of them read? And if AI is checking AI, where's the worth?

The same round trip can run on a letter of advice, a matter note or the figures in a client report. Somewhere along the way a person needs to read the work properly, and the firm needs to know who that person is.

1Human in the loop

That person is known as "the human in the loop". The phrase is used in AI and automation to describe a person who reviews or corrects a system's output. In a firm, it usually means a person checks what AI produces before the firm relies on it.

But who is the human? At what point do they check, and what are they checking for? What we're seeing is that each person answers those questions for themselves, so the rule looks a little different at every desk, and sometimes it comes down to a quick look at the tone.

In practice, a lot of the checking tends to settle on one careful person. Let's take a look at Jane in accounts.

Jane reads everything before it goes out. Invoices, statements, the covering letter with the figures in it, and a fair few things that were never hers to check. She reads AI output with the same care she gives every draft, and she catches things. Ask the partners whether the firm has a human in the loop, and they may well think of Jane and say yes.

Jane is doing the firm a great service. However, she can only check the work that reaches her desk. Work that leaves the firm by another route goes out with whatever check its author gave it, and when Jane takes leave, her standard goes on leave with her.

2What the rules expect

Jane's careful read is close to what the NSW Supreme Court and the Tax Practitioners Board now expect in their respective areas. The Supreme Court of New South Wales issued Practice Note SC Gen 23, which took effect on 3 February 2025. Where AI has been used to prepare written submissions, clause 16 requires the author to verify that every citation and authority exists, is accurate and is relevant to the proceedings. Clause 17 adds: "Such verification must not be solely carried out by using a Gen AI tool or program."

The Tax Practitioners Board followed in July 2026. Its guidance statement TPB(GS) 55/2026 says tax practitioners "should verify and review AI generated content for accuracy throughout each step of the workflow". It also says that "tax practitioners are still ultimately responsible for the tax agent services they provide to their clients."

A check at each step is a lot to ask of one person in accounts, and it is easy to skip. In the Australian results of a 2025 study by the University of Melbourne and KPMG, 57% of employees said they had relied on AI output at work without evaluating its accuracy, and 59% said they had made mistakes in their work because of AI. The same study found 51% had presented AI-generated content as their own.

Those figures are about accuracy, which is most of what Jane checks. Is the figure the one on the file? Do the citations exist? Did the case say what the paragraph says it said? Some accuracy checks can be done directly against the source.

Then there's checking for judgement. Is this what the firm wants to say to this client right now? Does the tone suit a relationship that has run for 11 years? Does it pass the "pub test"?

That takes experience of the client and the matter, and it is a large part of what clients come to the firm for.

When one phrase covers both kinds, the work often goes to whoever is nearest. In a firm, that can mean a figure corrected after it has gone to a client, or a partner checking citations line by line.

3Where a person stops

When a partner finds a wrong figure in something that has already gone out, the first instinct can be to tighten everything. Every AI draft gets read twice, or AI comes off client work altogether.

But what happens a month later? What can happen is that the checking takes up the time AI saved, the people who were using AI well slow down, and some of that use moves to personal accounts where the firm can't see it. Jane, meanwhile, has more on her desk than before.

A firm gets further by deciding once where a person stops and reads the work, and who that person is.

Naming the checkpoints

Five steps set out the checkpoints on one page.

1. List the work that leaves the firm or goes on the file, such as letters of advice, file notes, research summaries, reports and invoices.

2. Mark the point where a person reads each piece. We usually start with four: before a client sees it, before it goes on the matter file, before advice is built on it, and before a figure leaves the firm.

3. Put a name or a role against each checkpoint, and say whether that person is checking for accuracy, judgement or both. A letter of advice might have the author checking the facts and the responsible partner checking the judgement.

4. Read down the names. Where one name sits against most of the work, that person is carrying the firm's standard the way Jane does, and some of that load can move to the people who wrote the work.

5. Keep the lightest touch for low-risk uses, such as formatting a document or transcribing a recording, where AI is not generating substantive content. A quick look is enough there. Anything the firm may later rely on, meeting notes included, belongs at a checkpoint.

When this is in place, people know which work is theirs to check, and they can use AI on low-risk work with more confidence. A new starter can be given the page as part of their AI induction. A client who asks how AI is used on their matter gets a clear answer. Jane can take a holiday without the firm's standard going with her.

4Where we come in

Getting to Jane's holiday starts with knowing how AI output gets checked today. Our investigation looks at how your firm's work moves from first draft to the client, where a person reads it along the way, and how long your everyday work takes, so you have a baseline of your own. It includes a draft AI use policy for your partners to finalise, which gives your checkpoints a home. The findings belong to your firm whether or not we go any further together.

If you would like to work out where a person stops in your firm, reach out to our team. The first conversation works out whether your firm is the right fit for the AI Uplift Program, and whether we are the right people to help.

Finding your human in the loop