An auditor's real work takes about ten seconds.
It's the moment they sign. Everything before it — the weeks of building workbooks, tying figures back to support, selecting samples, chasing a balance that won't agree — is scaffolding for those ten seconds, when a person puts their name on a file and says, I looked at this, and it holds.
The signature is the product. Everything else is in service of it.
Yet all of the auditor's time goes into the scaffolding.
We've written before about how only 28% of inspected audits met the relevant standards last year, and how the number of registered auditors in South Africa fell almost 15% in the last six years. The failures aren't in the planning or the conclusion. They're in the middle: the manual grind of building the file. Because auditors are so worn down from building the scaffolding, the judgement gets starved of time.
AI can build the scaffolding. Then what?
Many auditors realise it's becoming more obvious by the day that AI can take on the grind. The biggest firms are spending billions of dollars each to level up their audit game. The question that follows is: what does it leave you with when the manual work is done?
There are two versions of the answer. They look very similar, but there is an important distinction.
In the first version, you hand over the file and the system hands back a conclusion. A clean workbook, a tidy answer, a green tick. All you have to do is sign.
No auditor will ever make a conclusion on faith alone. A tidy answer without transparent supporting work behind it won't ever get signed. You need to be able to go and look. You're paid to not quite believe things, and the confident answer from a system you can't see into is exactly the kind of thing you're trained not to quite believe. Nobody needs to warn an auditor about a black box.
So the problem with the first version was never that an auditor would be talked into trusting it. The problem is that in this version it leaves your scepticism with nothing to land on.
Imagine trying to review this version's output. You might ask why the revenue sample was accepted, and the system could give you a paragraph perfectly explaining its reasoning. But it would be a description of the work and not a record of the work. An auditor knows the weight of this distinction. The first is a story of what's been done, while the second is a verifiable audit trail of all actions taken and decisions made.
And if this audit trail of actions and assumptions isn't visible to you, then you only have two options. Either you take the story at face value — which you won't, because scepticism is your entire business. Or you rebuild the work to find out whether the story is true — which puts you back where you started.
A tool you can't see into just redirects your scepticism. Instead of the company you're supposed to be auditing, you end up auditing the software you're using. The tool that's meant to save you time and take work off your plate starts piling more onto it.
In the second version, the system does the same grind — but you can see into all of it. Every figure stays tied to the source it came from. Every test records what it looked at, against which standard, and what it found, at the moment it ran. Where the evidence is thin or where the numbers won't reconcile, it tells you. No phantom ticks. You don't have to read each and every line. But you know you can read any of them if you need to, and when something doesn't sit right, checking it takes seconds.
The difference between the two isn't how much work the machine does. In both, it does almost all of it. The difference is whether the work stays clearly visible after it's done.
Where the scepticism goes
When building the workbook took three days of keying figures, the scepticism you could exercise when you reached the end was limited by time constraints. When the file is built in minutes and every number is traceable, you can use the file in a whole new way. You ask the evidence a question and see the answer. You notice the thing that doesn't sit right and push on it. The client restates their numbers, as so often they do, and instead of three more days, you hand over the changes and the workbook updates. The auditor's real job — looking hard at the evidence and making a call on whether to believe it — finally gets the time it never had.
The machine shifts where the auditor's focus lives. It moves it two steps downstream. First, to the judgement: what's the real risk here, is this enough evidence, does this truly reconcile or just appear as if it does. And second, to the responsibility — because when IRBA opens the file, it's your name on it, not the vendor's and not the agent's. The one thing you can't delegate is the thing you're liable for.
Software went through this a few years early, and it's worth watching what actually happened. The best coding tools didn't turn engineers into people who approve code they can't read. The AI typed more of the code while making the code easier to inspect, to trace, to step into. The engineers didn't disappear. Their attention shifted too — to design, taste, review and judgement. Audit is the same shift arriving later, and with even more at stake, because a signed audit opinion carries a weight a line of code never will.
What we won't build
This is the line we're building Relay along.
Relay does the manual work and shows you everything behind it. Every figure traces to its source. Every test is recorded. When it isn't sure, it tells you. And its job is not to conclude — the judgement, the sign-off, and the opinion stay with you.
An auditor's authority has only ever come from two things: contact with the evidence and applying their mind to it. Not the hours, not the number of ticks, not how big the file is, but from having actually looked, and actually thought. A tool worth using brings you closer to both. It does not stand between you and the evidence.
The grind can go. Looking at the evidence can't. We're building Relay so that when you sign, you've seen the thing you're standing behind.