A recent AccountingWEB and Sage survey of 600+ UK practice professionals found that only 7 per cent say AI has been transformational. Their verdict on AI in accounting: “a keen but scattershot junior that needs constant supervision.”
Many bookkeepers will recognise that. Automation was meant to save time, but plenty of firms ended up with a new job: supervising the software. In our recent webinar, Norman Ieng, Daria Antonova and Andrei Belov from Apron explained why this happens and what we’ve built to fix it.
Where bookkeeping AI falls short
Traditional extraction tools lean on supplier-level rules, and those rules break down on exceptions. Amazon is the classic case. A client can buy anything there, so no single nominal code fits. Mixed VAT rates on one receipt and changing line items cause the same trouble. You can’t set a hard yes-or-no rule for any of them, so someone ends up checking every line.
Context and plain English, not rigid rules
William AI, our AI bookkeeping agent, takes a different approach. With the latest update to William, Guidance 2.0, you instruct it in plain English, much as you’d brief a new starter. As Norman put it, you’d tell a new team member what to look out for rather than have them click publish, and you can now do the same with William.
In the live demo, Daria wrote guidance for two near-identical fuel receipts: diesel goes to cost of goods sold, petrol to direct expenses. A supplier rule can’t tell them apart. William coded each correctly, and the reasoning on each document showed which guidance it had applied. Guidance can also split a document by line item, grouped by VAT rate or category, or split an invoice 50/50 across two categories.
You don’t have to start with a blank page, either. William watches the corrections you make and proposes guidance in a new Suggestions tab in your inbox. Every suggestion links to the documents behind it. If William is confident, it writes the guidance for you to approve with one click. If it’s less sure, it asks a question first. Firms using guidance are seeing five times fewer corrections, and for some categories, such as nominal codes and tax rates, almost none.
Built for trust
Automation only helps if you trust it to publish. Instead of the old all-or-nothing auto-publish toggle on supplier rules, William reasons about each document and checks how confident it is. It runs more than 20 checks, including image quality, whether the figures add up, and whether the coding matches that supplier’s history. Some are hard limits: it never auto-publishes documents from new suppliers or suspected duplicates.
You stay in control. You can tell William never to auto-publish anything above £5,000, or to leave anything that looks like client entertainment for you to review. For clients with auto-publish enabled, around 50 per cent of documents are published without a single correction.
Questions from the room
Attendees asked plenty about the practical side:
Email search: William looks for missing documents and can search multiple connected inboxes. It only searches for transactions that are unreconciled in Xero or QuickBooks.
Where guidance applies: Guidance covers everything that comes into Apron, not just documents found by email search. At the moment it applies to newly uploaded documents.
Starting out: Andrei’s tip was to ask yourself each time you correct a document whether William needs some guidance, then add it straight from the document page. Norman’s was to start with Suggestions.
What’s next
Webinar 2 in the series brings in practices already using William to share how they onboarded, which guidance they set first and how they involved their teams.
Get started with Apron, book a demo call, or watch the Webinar 1 recording below.