AI Invoice Matching for Fabricators: How We Saved One Shop 12 Hours a Week

This is the second post in the Applied AI for Fabricators series. The first post walked through six production AI systems we built for one client. This one zooms in on a single project — invoice matching and PO reconciliation — because it’s the one most fabrication shops should look at first. It’s high-impact, low-risk, and the savings are visible inside the first month.

The setup

Like every fabrication shop running real volume, our client (Kingsland Fabrications, an aluminium window and door manufacturer near Warrington) had a steady stream of supplier invoices arriving every week. Aluminium profiles. Glass. Hardware. Powder coating. Consumables. Each invoice needed checking against the original purchase order before it got paid — quantities, prices, part numbers, line-by-line.

The office team knew the drill. Pull up the PO, open the invoice, work through it. 15 to 30 minutes per PO. On a busy week, dozens of invoices. The maths was ugly: comfortably 10-15 hours a week of senior office staff time disappearing into a job that produces no windows.

It was also error-prone. Human attention slips on the fortieth invoice of the week — and slips most when invoices look mostly right but contain a single subtle discrepancy. A part number a digit off. A quantity that matches the PO but not the actual delivery. A price that’s gone up since the quote was agreed. These are the discrepancies that cost real money, and they’re exactly the ones that get missed when someone is trying to clear a backlog before lunch.

The failure mode that mattered most

The cost wasn’t really the staff time. It was what slipped through.

Wrong-quantity deliveries were the most expensive. Suppose 100 metres of profile gets ordered, 95 metres turn up, and the invoice is for 100 metres. If nobody catches it at invoice stage, it’s been signed off and paid by the time the shop floor realises they’re 5 metres short on a job. Now they’re either delaying the job, paying for emergency stock, or eating the loss. Three of these a quarter and you’ve lost more than the invoice automation would cost.

The wrong-price discrepancies were sneakier. Suppliers occasionally update prices without flagging it on the quote. If the invoice price is £2.40/m and the PO is at £2.20/m, that 9% delta multiplies across every invoice from that supplier until someone notices. Often nobody does. The first time anyone catches it is when the year-end accountant reviews supplier spend, by which point the cost is locked in.

Both failure modes are exactly what AI is good at: looking at every invoice with the same level of attention, every time, without the cognitive fatigue that makes humans miss subtle discrepancies on volume.

What we built

The pipeline runs in three passes:

Pass 1 — Read the invoice. A document model reads the PDF or photo of the supplier invoice and extracts every line item: part number, description, quantity, unit price, line total. It also reads the invoice header — supplier, invoice number, date, PO reference if present.

Pass 2 — Match against the PO. The system pulls up the original PO from the accounting system and runs a three-pass match:

1. Strict match — same part number, same quantity, same unit price.
2. Tolerance match — same part number, quantity within 2% (catches partial deliveries flagged as such), unit price within 1% (catches rounding differences).
3. Discrepancy flag — anything that doesn’t match either of the above.

Pass 3 — Surface the result. Office staff get a queue. Strict matches show one-click approve. Tolerance matches show the small discrepancy clearly (“invoiced 98m, ordered 100m, supplier annotated as partial delivery”) with one-click approve or escalate. Real discrepancies show the specific issue (“part X-447 invoiced quantity 12, ordered 10”) and route to the right person — usually whoever raised the original PO.

The whole match runs in seconds per invoice. Office staff went from doing 15-30 minutes of manual reconciliation per PO to spending 10 seconds glancing at a queue and clicking through. The job most weeks just disappears.

The bonus pipeline: material check-in

Once the PO matching engine was running, we extended it. Goods inward staff started photographing delivery notes when materials arrived. The same pipeline now does a fourth pass: matching what physically turned up against the PO, before it ever gets to the invoice stage.

This catches the wrong-quantity-delivery problem at the gate, not three weeks later when it’s already on the shop floor. The shop never starts a job short on stock without someone in the office being told first.

For an aluminium fabricator running tight job timelines on customer-set delivery dates, that single pipeline change is worth a multiple of what the whole project cost.

What it cost, what it saved

Build: roughly 3 weeks from kickoff to production. Fixed price. Single-pipeline focus.

Cost saved: office time recovered was ~10-15 hrs/week, which on its own pays the project back inside the first quarter. The harder-to-measure win is the wrong-quantity-delivery and wrong-price discrepancies that don’t slip through any more — but the team agrees this is the bigger number, even though it’s not on the invoice.

Cost shape: a focused single-pipeline build like this is £3,000 fixed price in our standard pricing. There’s no monthly subscription, no per-invoice fee, no usage tier — once it’s built, it runs.

We treat it as a starter project for fabrication clients. It’s the lowest-risk, highest-confidence applied AI work in manufacturing right now: the data it works on (invoices, POs) is already structured-ish, the failure modes it prevents are easy to put a number on, and the implementation is well-trodden territory.

When this won’t work for your shop

A few honest caveats:

• If your supplier invoice volume is below ~10/week, the staff time saved doesn’t justify a £3k build. The cleaner move is to do invoice checking properly with a checklist and a reminder.
• If your POs are still on paper or in a folder system not connected to anything, we have to do digitisation work first — adds 1-2 weeks and a bit more cost.
• If your suppliers send invoices in highly inconsistent formats (handwritten, scanned at angle, low resolution) the OCR layer needs a rougher-edge pipeline and validation tooling. Doable but worth knowing up front.

For most UK fabricators with 30+ supplier invoices a week and a basic accounting system, this is the AI project that should be top of the list.

How we’d start with your shop

If you run a fabrication shop and this sounds familiar, a 15-minute conversation usually establishes whether it’s a fit:

1. What’s your invoice volume per week? (Below 10 → not the right first project. 10-30 → likely fit. 30+ → strong fit.)
2. What system do POs live in? (Sage, Xero, QuickBooks, bespoke ERP — all fine.)
3. What’s the rough cost of a missed-quantity discrepancy? (If you can name a specific case in the last quarter, we can usually price it from there.)

That conversation is free. If we think the fit isn’t there, we’ll say so before you’ve spent a penny — there are usually simpler fixes that don’t need AI for shops below the volume threshold.

If you want that conversation: book a 15-minute intro call, or read more about how this fits into our broader AI document processing service.

Frequently asked questions

How accurate is AI invoice matching for fabrication suppliers?

In production at Kingsland we run above 95% accuracy on extraction and matching. The 5% that needs human review is exactly what you want a human reviewing — genuine discrepancies, edge cases, supplier formatting changes — not the routine 95% the system handles cleanly. Accuracy improves over time as the model sees more of your specific suppliers’ invoice formats.

Will it work with my accounting system?

Yes. We’ve integrated invoice matching with Sage, Xero, QuickBooks, and bespoke manufacturing ERPs. The pipeline pulls PO data via whatever interface your system supports — direct database access, exported reports, or API integration. No requirement to switch accounting systems.

What about handwritten or scanned invoices?

The OCR layer handles printed invoices, including PDFs sent by email and scans of varying quality. Genuinely handwritten invoices need a different model and more validation logic — possible but adds cost. Most of our clients’ invoices are PDFs from suppliers, which is where this works best out of the box.

Does it learn our suppliers’ specific formats?

Yes. Each supplier’s invoice format gets observed and tuned during the first month of running. After that, the pipeline is faster and more accurate on your specific supplier mix than a generic OCR tool would ever be.

What stops it from approving a fraudulent invoice?

The three-pass match validates against the original PO, which the supplier never sees. A fraudulent invoice for goods that weren’t ordered, or quantities that exceed the PO, will fail Pass 1 and route to a human. The system is paranoid by design — when in doubt, escalate.