I let AI clear two years of receipts. The leverage wasn't speed.

This week I cleared a backlog of roughly six hundred receipts and invoices — about two years' worth — that had been quietly piling up in an inbox. Scanned PDFs, in three languages, from a business I'm in the middle of winding down alongside my own freelance work. The kind of job everyone postpones because it's tedious, fiddly, and emotionally about as rewarding as flossing.

AI did most of it in an afternoon. But the speed isn't the interesting part. The interesting part is what the AI couldn't do, and why that's exactly where the value was.

The grunt work is genuinely solved now

Reading a PDF, finding the vendor, the date, the amount, the currency, the VAT line — that's a solved problem. Modern document AI now hits field-level accuracy in the high nineties on exactly this kind of multilingual invoice work, and it collapses the per-document cost from dollars to cents. McKinsey's broader read is the same: generative AI can automate a large share of finance activity, part of the trillions in annual value they keep modelling.

So I pointed vision-capable agents at the pile, in batches, and got back clean structured records for all six hundred documents. Zero missing, duplicates caught. Extraction, as a task, is done. If your bookkeeping pain is "someone has to type the numbers in," that pain is over.

The machine read every receipt perfectly. It still had no idea what any of them meant to me.

The value was in the rules, and the rules were mine

Here's the thing extraction doesn't touch. A receipt isn't just numbers — it's a decision. Does this expense belong to the business I'm closing, or to my freelance practice? Is it deductible or personal? Groceries and childcare are easy. But an invoice addressed to the old company, dated after I'd already gone full-time freelance, has to land in the old company's pile regardless of the date — because putting it in my freelance tax filing would be wrong, in the way that gets you a very unpleasant letter.

That rule — "the addressee always wins, except when the date says otherwise, except when it's clearly personal" — is judgement. It comes from knowing the actual situation: two entities, a specific handover date, a tax line I have to defend later. No off-the-shelf tool knows that. I had to define it, and refine it mid-run as edge cases surfaced. The AI executed the rule six hundred times without tiring. I had to be the one who knew what the rule was.

This is the pattern I keep seeing and keep writing about. The leverage of AI isn't that it replaces the expert — it's that it lets the expert apply their judgement at a scale that used to be impossible. It's the same lesson from when I pointed an AI agent at a live ad account: the tool surfaces the work, you supply the call. And it's the core of why I believe deep domain expertise is the only real leverage left — the person who knows which receipt goes where is now worth more, not less.

Knowing what to leave alone

There's a discipline question hiding in here too. With six hundred documents and a capable tool, the temptation is to fully automate the whole chain — auto-categorise, auto-post, auto-file, hands off. I didn't. The ambiguous cases, the ones where two valid interpretations existed, I flagged for human eyes rather than letting the machine guess confidently. A confidently miscategorised tax entry is worse than an honest "needs review."

That restraint — automating the certain, escalating the uncertain — is the same instinct I lean on across every mandate, the art of not fixing everything just because you can. The goal was a clean, defensible record, not a flashy "look, no hands" demo.

The afternoon I saved was nice. But it's not the point. The point is that the boring back-office work — the stuff people assume is beneath an experienced operator — is now exactly where that experience compounds. The machine handles the volume. You handle the meaning.

AI didn't do my bookkeeping. It did the typing. I still had to know what the numbers were for.

Sources & further reading

External
McKinsey — The economic potential of generative AI
iunera — The financial impact of AI receipt and invoice digitization

Related posts
The Death of Generic AI
I pointed the Meta MCP at a real ad account
The Art of Restraint