Automatic invoice capture is mostly a solved problem: point a scanner or an inbox at it, and a draft transaction appears. That part of the evaluation takes ten minutes. What's harder to see in a demo is what the tool does with a document it's unsure about, and what it does about the parts of an invoice that were never about reading at all. Here's what to actually ask before you rely on a tool's output for a client's return.
1. Does it tell you what it isn't sure of?
A tool that returns a flat result with no sign of how sure it was is asking you to trust every field equally: the crisp printed invoice number and the smudged handwritten total get treated the same. Ask whether extraction flags each field it is unsure of, and whether a doubtful field is shown to you rather than quietly accepted into the ledger.
2. What happens after the read, not during it?
Reading a document and checking it are different jobs. Ask specifically: does the tool run the required-particulars checks from HMRC's VAT guide (Notice 700) on every document (supplier details, tax point, line-item detail), or is that still a manual step at quarter end? Does it check the arithmetic, that net plus VAT equals the total and the lines sum to the subtotal? And does it look up the supplier's VAT registration, or only confirm the number is formatted plausibly? Format and registration are different questions: a number can be perfectly formatted and belong to nobody. (NoticeDesk looks the number up on HMRC's Check-a-UK-VAT-number service for GB and VIES for EU numbers. A lookup that can't be completed is recorded as unverified, never silently dropped.)
3. Can you produce the evidence later, not just the number?
If a return is ever queried, "the software imported it" isn't a working paper. Ask what's retained: is the field-level result stored, is there a timestamped record of what was checked and when, and can it be exported as evidence if HMRC asks? A verdict that exists only in memory during the import is gone by the time anyone needs it.
4. What does a failure actually look like?
Every extraction tool misreads something eventually: a rotated scan, an unexpected currency, a layout it's never seen. The question is what happens next: does a failed read surface for manual review, or does it silently write a best guess into the ledger with nothing to flag it? Ask to see the failure path, not just the happy one.
5. Where does the checked data actually land?
A clean CSV export isn't the same as an import your bookkeeping software accepts without a manual fix-up pass. Ask whether exports map to your accounts package's own import specification (including how credit notes are signed), or whether reconciling a generic dump against the chart of accounts is left to you, every time.
This is the same due diligence you'd apply to anything a client's VAT return depends on. NoticeDesk was built around these five answers specifically: a marker on every doubtful field, 16 compliance checks on every document (a 17th on credit notes and on receipts) with the gov.uk reference attached wherever an HMRC rule applies, a stored evidence trail per invoice, a manual-review path for anything it can't parse, and exports laid out to your accounts package's own import specification, with credit notes signed the way it expects.