Most comparisons of this software are feature tables where every product has a tick in every row. The ticks aren't false — they're just measuring things at a resolution where everything looks the same.
These are the questions that separate them.
Start here: what are you actually trying to remove?
Worth being specific, because "invoice capture" covers several different jobs and tools are good at different ones.
Filing. Getting documents out of email into somewhere findable. Largely solved, often free with your accounting software.
Data entry. Getting the values off the document and into your ledger, coded. This is where the time goes.
Approval. Getting the right person to sign off before payment.
Completeness. Knowing nothing's been missed.
If the actual pain is entry, a tool that files beautifully and leaves you typing hasn't helped. Pin this down before looking at anything.
The questions
1. Does it read line items, or just the header?
Header capture gets you supplier, date, total. If you code purchases to a single account per supplier, that's enough. If you split invoices across accounts, jobs or VAT rates, the work is in the lines and header capture leaves it with you.
Ask for extracted output from a multi-line invoice you supply. Not a demo video.
2. Does it tell you which fields it's unsure about?
A single accuracy figure per document is an average, and averages hide distribution. The useful behaviour is per-field confidence: this total is certain, this bank account isn't.
Ask what happens when a field reads badly. If the answer is "it's 99% accurate", that isn't an answer to the question.
3. Where does the data end up?
The distinction that matters: exported versus posted.
Exported means a file, or a folder of PDFs, or a CSV you import. Posted means a coded entry in your ledger.
A tool that exports to a folder has moved the documents, not the work.
Ask specifically about your accounting software. "Integrates with QuickBooks" can mean a live connection or a CSV you download and import. Those are different products.
4. What does it do with a photograph?
Every tool demos on a clean PDF. Your actual post includes phone photos taken at angles in poor light.
Send your three worst documents during evaluation. Any vendor confident in their extraction will welcome this. Hesitation is informative.
5. Does anything post without a human seeing it?
Some tools post automatically once confidence passes a threshold. Whether that's good or alarming depends entirely on your business — but you need to know which it does, and be able to change it.
Ask what the approval chain looks like and whether it's recorded. "Who approved this and when" should be answerable from the system.
6. What happens to related documents?
An invoice, its delivery note and a credit note against it are one commercial event. Most tools treat them as three unrelated captures, leaving you to connect them.
Ask how a credit note gets associated with the invoice it credits.
7. Does it learn your coding?
The first invoice from a supplier needs a coding decision. The fiftieth shouldn't.
Ask whether corrections you make are remembered, per supplier.
8. What does it cost at your actual volume?
Per-document pricing looks cheap at demo volumes and less so at 500 a month. Per-seat pricing looks cheap for one bookkeeper and less so once approvers need logins.
Ask what it costs at your real monthly volume, with everyone who needs access. Ask what happens when you exceed the allowance.
9. What does it cost to leave?
The data is your accounting records. You need to be able to take it.
Ask whether you can export processed data and original documents in bulk, and in what format.
10. What's the actual risk if it's wrong?
Not a features question, but the one underneath all the others.
The realistic answer is that no tool reads every document perfectly, so what matters is whether errors surface before they cost you anything. A wrong value that you're told about is an inconvenience. A wrong value silently posted is a problem you find later.
The checklist
| Question | A weak answer sounds like | A strong answer sounds like |
|---|---|---|
| Line items or header? | "We extract all your data" | "Here's the output from your invoice" |
| Which fields is it unsure about? | "We're 99% accurate" | "Per-field confidence, flagged below threshold" |
| Exported or posted? | "Integrates with your accounts" | "Posted as a coded entry; here's the API" |
| What about a photo? | "Best results with clean scans" | "Send us your worst three" |
| Does anything post unseen? | "It's fully automatic" | "Configurable, and every approval is recorded" |
| Related documents? | "Each is captured separately" | "Credit notes link to the invoice" |
| Does it learn? | "It uses AI" | "Corrections are remembered per supplier" |
| Cost at my volume? | "From £X per month" | A figure for your documents and seats |
| Cost to leave? | Hesitation | "Bulk export, documents and data" |
What the feature table won't tell you
Whether it handles your suppliers' actual documents. Extraction quality varies by layout, language and print quality in ways no spec sheet captures. Only testing on your own post answers this.
How much residual work is left. The honest comparison isn't against your current process at zero automation. It's against the work still remaining — reviewing flagged fields, handling awkward documents, coding new suppliers.
Whether the people who need to approve will actually use it. A tool that requires desktop login will not be used by someone who's on a site.
A sensible evaluation
Take twenty real documents, chosen to include your worst. A photo, a multi-line invoice, a credit note, a statement, something in a language you don't read.
Run them through each candidate. Then count: how many could you post without opening the original?
That number is the product. Everything else on the pricing page is commentary.
If you'd like to run that test on Cribble, send us your worst three documents. No install, and you'll get the extracted data back with per-field confidence so you can see exactly where it was and wasn't certain.