Accounts payable
The invoices your system hands back.
An AI model, trained on your own orders, contracts and past approvals and running on a server in your own office, works through the invoices that do not match. The exception queue stops being somebody’s whole week.
Your AP software already clears the invoices that match cleanly. What it cannot clear it hands to a person, and that pile is the job. Every invoice in it needs somebody who knows how your business buys to work out what it is, what it belongs to and whether it is right. That is the same judgement, made again and again, and it is why the AP team grows when the invoice volume does.
It runs on your own hardware. Your people approve and pay, and no invoice, contract or rate schedule leaves the office.
- Area
- Accounts payable, any industry
- Where it runs
- A server in your own office
- Working for you in
- About 8 to 12 weeks
- Investment
- About a junior’s salary, for a year
The challenge
Keying the number off an invoice is not the work. A scanner does that. The work is everything the match gets wrong, and every line below is a judgement somebody has to make, on every invoice, all day:
- Which order is this even against? It names no purchase order, or names the wrong one, or covers three, or is part of one still open.
- Is this line the same thing as that line? “Supply and install handrail, Level 3” on the invoice, “Balustrade works, stage 2” on the order. Same work, two companies describing it differently.
- Is the rate right? Against the contract, the schedule of rates, and the variation somebody approved by email in March.
- Is the quantity plausible? Against what was actually delivered, or what the progress claim says was done.
- Have we paid this already? The same work, billed again under a different reference.
- Whose job does it belong to? Which cost centre, which project, which contract.
Knowing the answers takes somebody who knows how your business buys. Do twice the invoices and you do twice of all of it, which is usually the point a business puts another person into accounts payable. The work that person does is the leftovers.
The queue is nothing but exceptions
Everything straightforward has already cleared. What reaches a desk is, by definition, only what did not, so the job is the difficult ones from start to finish.
It holds up the payment run
Suppliers ring about invoices nobody has reached yet, and somebody stops what they are doing to go and look. The call costs more than the check.
It is where money goes quietly
A duplicate paid under a second reference. A rate nobody checked against the variation. A quantity taken on trust. None of it ever appears as a loss.
Our approach
We do not replace the system you already have, and we do not put a chatbot on the desktop for people to remember to open. We train an AI model on your own orders, contracts and past approvals, and point it at the queue your system cannot clear.
- 1We go through a real month of exceptions with you and find the judgements that repeat. Most of a queue is a handful of situations happening over and over.
- 2The AI model learns from your own material: your orders, your contracts and rate schedules, your variations, and every invoice your team has already approved or sent back. That is how it learns the way your business buys. Nobody else has this one.
- 3We connect it to where your invoices and contracts already live, so nothing has to be filed anywhere new and the result lands in the system your team already approves in.
- 4Your people approve and pay. Every correction goes back into the training, so next month’s queue starts smaller.
There is no code in any of this, so there is nothing to re-code. No business analyst writing a spec every time something changes, no developer building it, no testing, no UAT, no release, and none of it again six months later. When the work changes we retrain the AI model on the new examples. It is also why an invoice that matches nothing in the system gets a sensible attempt instead of an error, where a rule set would simply stop.
We train the AI model the way you would train a junior, by showing it the work. The difference is that you train it once, it does not resign, and when the business changes you retrain it rather than hire and start again. What it learns stays with you: your own way of checking an invoice, held in an AI model you own.
What it handles
Finds what the invoice belongs to
The order, the contract and the job, even when the invoice names none of them, names them wrongly, or spans several at once.
Matches the line, not the wording
Your supplier’s description against your own, so the same work written two ways is recognised as one thing rather than sent back as a mismatch.
Checks the rate against the contract
The schedule of rates, the agreed variations and the adjustments that go with them, under your contract rather than a vendor’s idea of one.
Checks the quantity
Against the delivery, the progress claim or the work order, and says what it is comparing so a reviewer can agree or disagree in a moment.
Catches what you have paid already
The same work billed a second time under a different reference, a resubmitted invoice, a credit that was never applied.
Codes it to the job
Cost centre, project and contract, following the way your own team has coded the same kind of invoice before.
What stays with your people
The decision to pay. The disputes, and the conversation with the supplier that goes with them. The sign-off. The AI model presents what it found and what it is unsure about, with the order and the contract clause beside it, and waits rather than guessing. You set, task by task, what runs straight through and what always stops for a person, and that line is agreed in the scoping.
It runs in your office, so the question of what leaves it never arises
What your suppliers charge you, and what you have negotiated off their list price, is commercially yours. Putting it through somebody else’s system is a question you then have to answer. This one runs on your own hardware, so there is nothing to answer.
- No invoice, contract or rate schedule leaves for processing. No upload, no outside service, nobody else holding or learning from what you buy.
- No new third party sees what you pay your suppliers. Not a policy promise: there is simply nothing transmitted.
- No fee per invoice or per document. The AI model is yours and it runs on your server, so a busy month does not cost more than a quiet one.
- The approval trail stays inside your own systems, which is where your auditors already go looking for it.
The results
What the whole project costs, about a year of a junior’s pay, and it pays that back in a few months · indicative
The AI model absorbs exponentially more of the judgements that repeat, so the size of the AP team stops tracking the number of invoices
From the scoping call to the AI model doing the work. What moves the date is how easily your systems connect, not the training and tuning of the AI model · indicative
What that buys is not a smaller finance team. It is the same people, off the queue and onto the parts of the month that need them, and an invoice volume that can double without the headcount doubling behind it. The figures above are indicative; on the scoping call we work them out again with your own numbers.
Common questions
How is this different from the AP automation we already have?
It starts where that stops. AP automation reads an invoice and clears it when the numbers line up with the order. When they do not, it raises an exception and hands it to a person, and that queue is the cost. This is trained on your own contracts, variations and past approvals, so it works on the ones that did not line up. If your current system is clearing most of your invoices, it stays where it is and this takes the rest.
Does anything leave the office?
No. The AI model runs on a server in your own office. Your invoices, your contracts and your rates stay on it. There is no upload, no outside service holding them, and nothing for you to disclose as a new third party.
Who approves the payment?
A person, every time, unless you decide otherwise for a specific kind of invoice. The AI model does the finding and the checking and shows its working, with the order and the contract clause next to what it found. Where it is not confident it says so instead of filling in a number.
Our contracts and rate schedules are a mess. Is that a problem?
It is the normal starting point, and it is the reason the AI model is trained on your own material rather than configured against a template. A rate agreed in an email thread is still how you agreed it, and that is what it learns from. Where a document is genuinely unreadable it says so rather than guessing.
How long does it take?
About 8 to 12 weeks to a working system, with the scoping before it. What moves that is not the invoices. It is how your systems are set up and how much has to be built to reach your orders and contracts.
Is your first stretch the same as this?
A scoping call is a conversation about which of your tasks are worth taking on, and what they are costing you now. It is free, and there is no obligation.
Prefer to talk? Call +61 475 043 070