After 3 Invoices, the AI Recognises Your Supplier. After 10, It Stops Asking.
The first diesel invoice from a new supplier may need a reviewer. When the reviewer corrects a match, that wording is saved as an alias, so later invoices that describe the material the same way match through it. Here's how that learning works and where it stops.
Your sustainability analyst opens a diesel invoice from a fuel supplier they've never seen before. New format. Different layout. The consumption figure is on page two, buried between a delivery address and an ABN. The analyst finds the litres, looks up the NGA emission factor (2.71 kg CO2-e per litre for diesel oil), classifies it as Scope 1, and creates the record. That takes about four minutes.
Now multiply that by 2,000 invoices a quarter. That's over 130 hours of someone's life spent doing the same cognitive task: find the number, match the material, apply the factor, confirm the scope. Every single time. Even when 60% of those invoices come from the same twelve suppliers, in the same format, for the same materials.
This is the part of carbon accounting that nobody talks about at conferences. Not the strategy. Not the science. The mind-numbing repetition.
Static Tools vs. Systems That Learn
Many document workflows treat every document the same way, every time. Upload an invoice. Extract the data. Present it for review. Doesn't matter if you've seen that exact supplier format 400 times. The system has no memory. It doesn't get better with use.
That's fine for a company processing 50 invoices a year. It's not fine for a Tier 1 contractor processing 2,000 a quarter across 200+ suppliers, or a property manager with 40 buildings generating 640 utility documents every three months.
We built Carbonly's material matching to remember corrections. When a reviewer corrects a material match, or assigns a material to a line the system could not match, the wording of that line is saved as an alias on that material for your organisation. Later documents that use the same wording match through the alias. Aliases stay inside your organisation; nothing learned from your documents is shared with another tenant.
The first invoice from a new diesel supplier? The system extracts the data and suggests a match to a factor from the NGA Factors workbook. Low-confidence matches go to a reviewer to confirm or correct.
If the reviewer corrects the match, that supplier's wording becomes an alias. The next invoice that describes the material the same way matches through it, instead of landing in the queue for the same reason.
Accepting a match that was already right does not add an alias. Learning comes from corrections, not approvals, and there is no count after which the system stops asking. A low-confidence match still goes to a reviewer.
That is the difference between a static tool and one that remembers how your suppliers describe what they sell.
The Maths on 2,000 Invoices a Quarter
Let's be specific about what this means for a company with real volume.
A mid-to-large construction company or property portfolio might process 2,000 emission-relevant documents per quarter. At an average of three minutes per manual review (finding the consumption figure, confirming the material, checking the emission factor, verifying the scope classification), that's 6,000 minutes. One hundred hours. Two and a half full working weeks of someone's time, every quarter, just reviewing data that's already been extracted.
How much review time that saves depends on how consistently your suppliers word their invoices and how often a correction is needed. We have not published a figure for it, and we would be wary of anyone who quotes one without seeing your documents. What should stay in the queue are the records that need human judgment: new suppliers, unusual materials, wording the system has not seen, and low-confidence matches.
The goal isn't zero review. It's spending review time on the records that need it.
How the Learning Actually Works
When your analyst reviews a low-confidence record and corrects the match ("no, this line is diesel oil"), the correction is saved.
The system saves the line text as an alias on the corrected material, for your organisation only. It doesn't just remember the supplier name. It remembers the way the material was described on the invoice. So if one supplier writes "automotive diesel" and another writes "diesel fuel", once a reviewer has mapped each wording, later lines with that wording match to the same material.
Aliases are capped at 500 per material, and they only help when the wording repeats. New phrasing goes through the normal matching tiers and, if confidence is low, to a reviewer.
Here's the part that matters for compliance: whenever confidence is low, a match is still a suggestion with a reviewer behind it. In carbon reporting under NGER and AASB S2, a false positive (an incorrect emission factor applied without review) is far worse than a false negative (asking a human to review something that turns out to be fine).
What Still Goes to a Reviewer
Not everything should be automated. We're pretty clear-eyed about this.
Low-confidence matches go to a reviewer, however many aliases the system holds. So do units that cannot be converted without a figure nobody stated, such as cubic metres to tonnes with no density. When the NGA Factors workbook gets updated (it happens annually, and the 2025-26 NGER legislation amendments changed several factors), each record keeps the emission factor version it was calculated with, so you can see which records used which edition.
Why This Matters for AASB S2 and NGER
Group 2 entities under ASRS start mandatory climate reporting for financial years beginning 1 July 2026. Group 3 follows from 1 July 2027. Both groups need to report Scope 1 and 2 emissions from day one, with Scope 3 required from the second reporting period. NGER reporters already face annual deadlines by 31 October, with civil penalties of up to $728,000 per contravention under Section 19 of the NGER Act.
The ANAO's performance audit of the NGER scheme found that 72% of 545 reports examined contained errors. Seventeen percent had significant errors. These were organisations using manual processes, spreadsheets, and consultants. The data quality problem in Australian emissions reporting isn't theoretical. It's documented by the government's own auditor.
A system that remembers corrections doesn't just save time. It reduces the error surface. Once a reviewer has mapped a supplier's wording to the right material, later lines with that wording match the same way instead of being re-keyed by hand. Every line that is not re-keyed is one less chance to fat-finger an emission factor, misclassify diesel as Scope 2, or apply the Victorian grid factor (0.78 kg CO2-e/kWh) to a facility in Tasmania (0.20 kg CO2-e/kWh).
That's worth something to an auditor. Under ASSA 5000, the assurance practitioner needs to understand your internal controls over sustainability information. A system that keeps the source document and verbatim line text behind each record, records typed audit events on ledger writes, and locks reporting periods? That's a control environment. A spreadsheet with 2,000 rows and no trace of who entered what? That's a risk.
The First Three Months Are the Hardest
We won't pretend the onboarding period is effortless. The first months involve more corrections than later ones.
In the first few weeks, more records need a reviewer. The system is seeing your suppliers' wording for the first time. Your analyst will spend time confirming matches, correcting the occasional misclassification, and teaching the system how your supply chain describes things.
This is the investment period. And it's tempting to look at the review queue in week two and think "this isn't saving me any time." It isn't. Yet.
But corrections accumulate. Each one saved as an alias is a correction you should not have to make again for the same wording. How quickly the review queue shrinks depends on how consistent your suppliers' wording is.
For organisations with many regular suppliers, there is a crossover point where the system saves more time than it costs to review. When it arrives depends on supplier count and document mix. After that, the gap should widen every month.
Consider a property manager tracking emissions across 40 buildings. They have maybe eight energy retailers across three states. Once each retailer's wording has been mapped, the quarterly electricity data that used to take two full days of manual entry should need far less review. The time goes to the things that actually require judgment: new tenants, unusual consumption spikes, reconciling estimated vs. actual meter reads.
The Audit Trail Makes It Defensible
Every record in Carbonly carries its history. Not just the final emission figure. The chain: source document, the verbatim line text, matched material, applied emission factor and its version, and calculated emissions.
When an assurance practitioner (or the Clean Energy Regulator) asks how a specific record was generated, you don't say "the AI did it." You show them the source document, the line text the record came from, the material it matched and the emission factor version applied.
That's the difference between using AI as a black box and using AI as a documented, auditable control system. The ACCC's continued focus on greenwashing (including an $8.25 million penalty against Clorox in 2025 for misleading environmental claims) means your emissions numbers need to be traceable. Not approximately right. Demonstrably right, with evidence.
What Happens When Suppliers Change
Suppliers change their invoice formats. It happens. A fuel company updates their billing system. A concrete supplier gets acquired and rebadged. An energy retailer redesigns their invoice template mid-year.
When a known supplier changes its wording, the old alias no longer matches. The line goes through normal matching and, if confidence is low, to a reviewer. A correction saves the new wording as another alias.
This is intentional friction. We'd rather ask your team to re-confirm a supplier whose format changed than silently extract the wrong number from a layout the system hasn't seen. The cost of a five-second manual confirmation is negligible. The cost of an incorrect emission factor flowing into your NGER return is not.
The same caution applies when emission factors change. DCCEEW updates the NGA Factors workbook annually. Each record stores the emission factor version it was calculated with, so it is clear which records used which edition.
Start With Your Highest-Volume Suppliers
If you're evaluating whether this approach works for your organisation, here's a concrete way to test it.
Pick your top twenty suppliers by invoice volume. For most companies, those twenty suppliers account for 60-70% of all emission-relevant documents. Upload three months of invoices from those suppliers. Correct the matches carefully as you go, then upload the next month and see how many lines now match through wording you have already corrected.
Then look at the numbers. How many lines needed a correction in month two compared with month one? How much review time did you save? What's the error rate on matched records compared with manually entered ones?
That data tells you whether a learning system fits your reporting workflow better than a static one. If you want to run that test, reach out at hello@carbonly.ai and we'll walk you through it.