When we tell people BIZA has an AI assistant that can create invoices, record expenses, and capture supplier documents, the first question from anyone who has run a finance team is always the same: "And it just… posts them?"
No. It doesn't. And that single decision — AI drafts, humans post — is the most important design choice in the product. This post explains why we drew the line there, and why we think every AI feature that touches financial records should.
Your books are not a chat transcript
A conversation with an AI is low-stakes: if the answer is wrong, you ask again. Your general ledger is the opposite. It is the record you file taxes from, the record your auditor samples, the record a bank reads before extending credit. In most jurisdictions, tax invoices are legal documents with prescribed contents — under ZATCA Phase 2 in Saudi Arabia they're cleared with the authority in real time.
An AI that writes directly into that record has a failure mode with no good answer: when (not if) it misreads something — a similar customer name, a decimal point on a photographed receipt, a tax code on an unusual transaction — the error doesn't sit in a chat window. It sits in your books, compounding quietly until reconciliation day.
Drafts change the economics of errors
Here's the asymmetry that makes drafting the right model. Reviewing a prepared document is fast; preparing one is slow. Catching an error in a draft costs seconds; finding one in posted books costs hours.
So BIZA splits the work along exactly that line:
- The AI does the slow part. Reading the supplier invoice, resolving the customer name, filling twelve fields, applying the tax treatment, coding the lines against the supplier's history.
- You do the fast part. Looking at a completed draft — side by side with the source document, when there is one — and deciding: post, edit, or reject.
A wrong draft costs you the ten seconds it took to spot it. That's the entire downside. Compare that with the downside of a wrong posting, and the design picks itself.
Uncertainty is surfaced, not smoothed over
A drafting AI can afford to be honest about what it doesn't know, because a human is guaranteed to look. So when BIZA's assistant is unsure, it says so:
- Two customers with similar names? It asks which one, rather than picking.
- A captured total that doesn't match the printed VAT arithmetic? It flags the difference rather than silently "fixing" it.
- Fields it couldn't read confidently? Highlighted, so your eyes go there first.
An AI that must act autonomously has the opposite incentive — it has to resolve every ambiguity somehow, and its guesses are invisible precisely because no one reviews them. The approval step is what buys the honesty.
Accountability stays with a name
Every posting in BIZA carries the identity of the person who approved it, in the same audit log that records everything else. When someone asks "who approved this?", the answer is a person — someone who saw the draft and confirmed it — never "the model did."
This matters beyond blame. Approval rules, role permissions, and multi-level sign-offs are how organisations encode who is trusted with what. The assistant operates inside those rules, not above them: it can't approve its own drafts, and it can't draft its way past a permission your role doesn't have. AI capability grows; the control structure holds.
"Doesn't the review step slow you down?"
Less than you'd think, and the trade is worth measuring honestly.
Typing a supplier invoice by hand: a few minutes of careful field-by-field entry, with typos as a bonus. Reviewing a captured draft: around ten seconds when the document is routine, longer when the AI has flagged something — which is exactly when you want to slow down. Our walkthrough of the capture workflow shows the full loop; "under a minute, photo to posted entry" is the honest steady state, review included.
The review step is also tunable in the right dimension: routine, low-value, high-confidence drafts take one glance; large or unusual ones deserve a real look. What we won't do is make the glance optional. The moment "AI posts routine entries automatically" ships, the definition of routine becomes the attack surface — and the first bad batch erases every minute the shortcut ever saved.
The bottom line
The useful question about AI in accounting isn't "how much can it do?" — it's "what happens when it's wrong?" Drafting gives that question a boring answer: a human catches it, in seconds, before it touches the books. That's why every AI path in BIZA — talking to the assistant, capturing documents, generating records — converges on the same final step: a person, looking at a draft, deciding.
See the AI assistant and AI document capture in detail, or try the loop yourself.