Choosing an AP Automation App for Business Central

This page is a set of questions, not a scorecard. They apply to any candidate — including this one — and each of them separates products that behave differently under pressure.

We do not name other products anywhere in this Help Center. The questions below are meant to be asked of whatever you are looking at, by you.

1. Where does the number that gets posted come from?

The most important question, and the one most easily answered vaguely.

There are two fundamentally different designs:

• Transcription. The tool reads the vendor’s invoice, produces a purchase invoice from what it read, and optionally checks it afterwards. The posted figure originates from the PDF.
• Reconciliation. The tool reads the vendor’s invoice as a claim, compares it against your purchase order and posted receipt, and lets Business Central build the document from your own records. The posted figure originates from your ledger.

Ask which one you are looking at. The answer determines what happens when the reading is wrong, and that is a question of when, not if.

Under a reconciliation design, a misread number causes an exception. Under a transcription design, a misread number that happens to fall within tolerance is a posted number.

2. What happens on the day the model is wrong?

Follow up on question 1 with a concrete case. Ask for a demonstration where the invoice claims more than was received.

Watch what the product does. Specifically:

• Does it stop, and say precisely what disagreed and by how much?
• If the exception is overridden, what does the resulting invoice contain — the vendor’s figures, or yours?
• Is there any path by which the vendor’s figure reaches the ledger?

A product that answers “our accuracy is 99%” is answering a different question. Accuracy tells you how often the reading is right. It tells you nothing about what happens the rest of the time.

3. Is there a per-vendor setup or training phase?

Ask what happens with a brand-new vendor’s first invoice, and what happens when an existing vendor changes their layout.

Coordinate-mapped approaches need a template per vendor and break when a layout shifts. Semantic approaches do not. The difference is invisible in a demo — demos use vendors that are already configured — and very visible in month three. See templates and semantic extraction.

4. What does it tell you when something fails?

Ask to see an exception on screen.

A confidence score tells you how sure the model is. It does not tell you what is wrong, it is not comparable between documents, and it gives a reviewer nothing to act on except “look harder”.

A reason code with the arithmetic behind it — expected 5, actual 8, delta +3 — tells a reviewer what to do next. Ask whether the codes are a fixed, documented set, or free text generated per document. A fixed set can be reported on, filtered and trended. Free text cannot. See reason codes, not confidence scores.

5. What does it do about duplicates?

Duplicate payments are one of the largest real losses in accounts payable, and duplicate detection is easy to do badly.

Ask specifically:

• Does it compare file content, or only the invoice number?
• Does it check against invoices already posted in Business Central, not just against its own queue?
• What happens when a vendor re-issues a corrected invoice under the same number?

That last question is the revealing one. A product that treats a corrected re-issue as a duplicate and discards it has destroyed the only correct copy of a bill somebody has to pay. The safe behaviour is to route it to a person.

6. Does it post anything on its own?

Ask whether there is any configuration — any flag, any “full automation” mode — under which the product posts to your ledger without a person.

There are reasonable products that offer this. But you should know the answer deliberately rather than discover it, and you should know what the guard is if the answer is yes.

Also separate two words that get used interchangeably: registering (creating an unposted Business Central document) and posting (committing to the ledger). “Touchless” means one of them. Ask which.

7. Does it use Business Central’s own approval framework?

Business Central has approval workflows, approval entries and Requests to Approve natively.

A product that routes through them means your approvers use one interface for everything and your approval history lives in one place. A product that builds its own parallel approval system means two queues, two sets of notifications, and an audit trail split across two places.

Business Central also allows only one enabled workflow per document type — so ask what happens if you already have one running.

8. Where do the invoice files end up, and who holds them?

Invoices are records most organisations must retain for years — longer than any software subscription is guaranteed to last.

Ask:

• Where do the PDFs physically live?
• Do they count against your Business Central database capacity?
• If the subscription lapses, can you still get to your own documents?
• Is there an export route, and how long is it available for?

If the answer is that the vendor holds your documents and access is tied to an active subscription, that is not automatically wrong — but it is a retention decision, and it should be made deliberately by whoever owns your retention policy. See where your documents are stored for how DocumentQ answers this, including the parts that are a window rather than a guarantee.

9. What leaves your tenant?

Any product that uses an AI service sends your invoice document somewhere. Ask exactly what else goes with it — vendor master data, chart of accounts, ledger entries? — and what the provider retains.

Ask whether anything is stored by the vendor afterwards, and get it in writing rather than in a demo.

10. What has actually been observed working?

The most useful question, and the least often asked.

There is a real difference between built, deployed and observed working on a real tenant. Software can pass every test it has and still never have been watched doing the thing it is sold for.

Ask which specific scenarios have been run end to end on a live tenant, and be more interested in a candid list with gaps than in a confident one without. A vendor who can tell you precisely where the boundary of their own evidence sits is telling you something useful about how they will behave when you hit a problem.

DocumentQ’s own answer to this is in the FAQ, including what has not been observed.