Most advice on this starts with choosing software, which is the third or fourth thing you should do rather than the first. Automating a process you haven't tidied usually automates the mess.
This is the sequence in the order that works, with the parts that stay manual named honestly.
What "invoice processing" actually contains
Six distinct jobs. They automate to very different degrees, and lumping them together is why expectations end up wrong.
| Step | How well it automates |
|---|---|
| Receiving documents | Completely |
| Identifying what each one is | Mostly |
| Extracting the data | Mostly, with review |
| Coding to accounts | Partly — improves over time |
| Approval | The routing, not the decision |
| Posting to the ledger | Completely, once the rest is right |
The middle rows are where the value is and where the residual work lives.
Step 1: One address, before anything else
Every invoice arrives at a single dedicated address rather than at individuals.
This costs nothing, takes an afternoon, and is worth doing even if you automate nothing else. It removes the largest single cause of lost invoices, and it's the prerequisite for every later step — you can't automate a stream you can't see.
Tell suppliers once, then put the address on your remittances and purchase orders so it keeps being reinforced. Set up forwarding from the personal inboxes invoices currently reach, and expect that to run for months.
Step 2: Tidy the coding before you automate it
Automation learns your conventions. If your conventions are inconsistent, it learns the inconsistency.
Worth doing first:
- Prune the chart of accounts. Most small ledgers have codes nobody uses and pairs that overlap. Decide which of each pair survives.
- Fix supplier defaults. Set the default nominal code and VAT treatment per supplier in your accounting package. This is the highest-value setting most packages have and it's frequently unused — and it's the thing a capture tool builds on.
- Agree the conventions with budget holders rather than discovering them one rejected invoice at a time.
Half an hour here changes what any tool can do for you.
Step 3: Automate capture and extraction
Now the software step. What to look for, briefly, since this is covered properly in the buyer's checklist:
Line items, not just the header. Header capture gets you supplier, date and total. If you split invoices across accounts or VAT rates, the work is in the lines.
Per-field confidence. A single accuracy score per document tells you nothing about which value to check. Per-field scores turn review from re-reading the page into checking one row.
Posted, not exported. A folder of PDFs is filing. A coded entry in your ledger is data entry. Ask specifically which one you're buying.
Your documents, not their demo. Test on twenty of your own, including the worst.
Step 4: Automate the routing, not the decision
Approval automation means the invoice reaches the right person with the right context and its state is visible. It does not mean software approves things.
The mechanics worth setting up:
- A value threshold, so low-value recurring invoices from known suppliers don't get routed at all. Every routed invoice is a chance for a delay.
- Routing by who ordered it, not by department.
- A visible queue showing what's outstanding and with whom. This single change removes most silent delay, because email has no state and a list does.
- A rule for bank detail changes that always involves a human and a phone call to a number you already had. Keep this one manual deliberately.
Step 5: Post to the ledger
The last step, and the one that decides whether any of the above saved time.
If your accounting software is cloud-based, a capture tool can usually post directly. If it's desktop, the realistic mechanism is a CSV import — which still removes the typing but leaves an import step.
Decide explicitly whether anything posts without a human seeing it. Both answers are defensible; not knowing which one your tool does is not.
What stays manual
Worth being straight about, because a plan that assumes otherwise overruns.
Coding decisions on unfamiliar suppliers. The first invoice from a new supplier needs a judgment from someone who knows the business. The fiftieth shouldn't.
Documents that read badly. Photographs at an angle, faint thermal receipts, handwritten dockets. These get flagged rather than solved.
Anything that isn't an invoice. Statements, pro-formas, order acknowledgements and delivery notes all arrive in the same stream and none should be entered as invoices.
The approval decision itself. That's the control you're keeping.
Queries. Wrong quantity, price differs from the quote, no reference. These need a person and a conversation.
Realistically, expect to still spend 15–25% of the original time once things have settled, and more than that in the first month while you're checking everything twice.
A sensible order
If you do nothing else, do steps 1 and 2 — they're free and they improve the process whether or not you ever buy software.
- Week one: single address, forwarding set up, suppliers told
- Week two: chart of accounts pruned, supplier defaults set
- Week three: trial two capture tools on the same twenty real documents
- Week four: run one live alongside the old process, don't switch yet
- Month two: switch, keep the threshold conservative, widen it as trust builds
The parallel-running month is the step people skip and the one that catches the problems while they're still cheap.
How to tell whether it worked
Not by the vendor's accuracy figure. Two numbers:
How many documents could you post without opening the original? This is the residual, and it's the whole business case.
How long from invoice arriving to approved and ready to pay? If entry used to happen monthly and now happens on arrival, this drops by weeks — usually the larger benefit, and the one nobody predicts.
Measure both before you start. Most people never do, and then can't tell whether the change was worth it.
Cribble reads invoices from a dedicated address, extracts to the line item, scores every field separately so you know which values to check, and posts into Xero — or exports CSV for QuickBooks, Sage50 and DATEV. Send us your three most awkward documents if you want the residual figure for your own paperwork.