Document quality
A native PDF exported from a billing system is close to a solved problem. A photographed carbon copy, skewed on a desk under office lighting, is not. Most published figures are measured on the first kind. Your inbox contains both.
Accuracy
We don’t publish an accuracy percentage.
Every vendor in this category quotes a number in the high nineties. We could too — pick a clean test set, measure the easy fields, round up. It would tell you nothing about your invoices, and you’d have no way to check it.
Here is what actually determines the number, what Odyssin does when it isn’t confident, and how to measure it on your own documents in an afternoon.
A single number hides the thing you care about
“99% accurate” is an average across a test set someone else chose. Your accounts payable inbox is not that test set. Three things move the result far more than the headline suggests, and only one of them is about the software.
Which field
Vendor name, invoice number, dates and totals sit in predictable places and read reliably. Line items are a different problem: variable row counts, wrapped descriptions, subtotals that look like line totals. A blended average buries that difference.
Your vendor mix
Odyssin uses no templates and no per-vendor setup, so a brand-new vendor’s first invoice processes like one you have received a hundred times. But if most of your spend arrives from a handful of suppliers with unusual layouts, your result reflects those suppliers — not an average.
The alternative
Confidence per field, not a promise per vendor
The useful question isn’t “how often is it right?” but “what happens when it isn’t sure?” A system that is right 99% of the time and silently wrong the rest is worse than one that is right less often and tells you which is which.
-
Step 01
Every field gets a confidence score
Not one score per document — one per field. A total read cleanly and a due date read from a smudge are not treated the same way just because they arrived on the same page.
-
Step 02
You set the threshold
Anything scoring below your threshold is routed to a person instead of posted quietly. Where you put that line is a business decision about how much review you want, and it is yours to move.
-
Step 03
Validation catches what confidence can’t
A perfectly legible invoice can still be a duplicate, fail its own arithmetic, come from an unapproved vendor, or price above contract. Those checks run on every document regardless of how confidently it was read.
-
Step 04
Corrections stick
Exceptions land in a queue with the original document alongside, so a fix takes seconds rather than a hunt through email. When you correct something, Odyssin remembers that judgement for that vendor.
Do this instead
Measure it on your own invoices
Run a month of your own invoices through it rather than a curated demo set. This takes an afternoon and gives you a number you can defend internally.
-
1
Take your last 50 invoices — all of them
Not a curated sample. Include the handwritten one, the crooked scan, the vendor whose layout nobody can explain. Those are the invoices that decide whether this works for you.
-
2
Run them through and count corrections
Every field you have to fix is one data point. The count of corrections divided by the count of fields is your real accuracy figure, measured on your documents rather than ours.
-
3
Split header fields from line items
Score them separately. If you only need vendor, date and total posted correctly, line-item performance may not matter to you at all — and a blended number would have hidden that.
-
4
Check what the exceptions cost you
Time the review queue. Thirty seconds on eight flagged invoices is a different proposition from five minutes on forty. This is the number that decides whether the thing pays for itself.
-
5
Then compare vendors on the same set
Run the identical 50 documents through anyone else you are evaluating. Published figures become irrelevant the moment you have your own on a like-for-like set.
Worth asking any vendor, including us
If a number is quoted at you, these are the questions that turn it back into information.
-
Measured on which documents? A public dataset of clean PDFs, or invoices resembling the ones I actually receive?
-
Which fields count? Header fields only, or line items too — and what is the figure for each separately?
-
What counts as correct? An exact string match, or close enough after normalisation? The two produce very different percentages.
-
What happens on the misses? Are low-confidence reads flagged, or posted with everything else and discovered later?
-
Can I reproduce it? On my own documents, before paying, without a sales engineer driving.
Get your own number
Run your last fifty through it. Bring the messy ones — they are the only ones that tell you anything.