Forward a Bill, Get Emissions Data: How Email Ingestion Kills the Carbon Accounting Bottleneck

Here's a scene that plays out every week at hundreds of Australian companies. A site manager on a construction project gets a diesel delivery docket. They fold it in half, shove it in their hi-vis pocket, and drive to two more sites before heading home. The docket sits in their glovebox for a fortnight. Eventually, someone in the sustainability team sends a chasing email. The docket arrives as a photo taken under fluorescent light at 9pm on a Tuesday. Someone types the litres into a spreadsheet. Maybe correctly, maybe not.

That's not carbon accounting. That's a scavenger hunt.

We built carbon accounting automation email ingestion into Carbonly because we got sick of watching this happen. Not in theory — in practice, during our own pilot with a construction company generating 10,000 fuel receipts per quarter. The problem was never the maths. The maths is multiplication. The problem was getting the numbers from source documents into a system that could do anything useful with them.

The real bottleneck isn't calculation. It's collection.

BCG's GAMMA surveys found that only 9% of organisations can measure their greenhouse gas emissions with any real frequency or accuracy. That number barely moved between 2021 and 2024. And the same research put estimated error rates at 25-30% across emissions measurements. Not because the emission factors are wrong (though that happens too — we wrote about how factor selection creates errors). Because the source data never makes it from the document to the database intact.

The GHG Protocol's own findings back this up: 83% of companies reporting climate disclosures struggle to access their emissions data. Not calculate it. Access it. The raw numbers sitting in PDFs, Excel files, scanned invoices, and yes — photos of crumpled dockets in someone's pocket.

For Australian companies reporting under NGER, this has real consequences. The ANAO audited 545 NGER submissions and found 72% contained errors, 17% of them significant. The most common problems were gaps in electricity data and missing sources. These aren't sophisticated methodology disputes. They're "we didn't have the bill" problems.

With ASRS Group 2 reporting starting from July 2026 and Scope 3 disclosures kicking in for those entities in their second year, the volume of documents flowing into emissions systems is about to multiply. You can't solve a document volume problem by hiring more people to type faster.

What email ingestion actually means

Every project in Carbonly gets a unique email address. Something like project-ABC@mail.carbonly.ai. You forward a utility bill, fuel docket, waste manifest, or supplier invoice to that address. Done.

No logging into a platform. No navigating to the right project folder. No selecting a file type or filling in metadata fields. You hit forward, and the document enters the processing pipeline.

The site manager with the diesel docket? They photograph it with their phone and email it to the project address. Takes about 15 seconds. The accounts payable team gets an electricity bill PDF from Origin? They set up an auto-forward rule once. Every subsequent bill goes straight to Carbonly without anyone touching it.

This matters because the friction in carbon data collection isn't processing time. It's participation. Getting the person who holds the document to actually send it somewhere useful. Every additional step you add — log in, navigate, select project, choose document type, upload — is a step where someone decides it can wait until tomorrow. And tomorrow becomes next week. And next week becomes a frantic scramble in September because the NGER deadline is 31 October and half the Q2 bills are missing.

An email address is the lowest friction data channel that exists. Everyone already knows how to forward an email.

What happens after you hit send

Here's the part where this stops being about email and starts being about what the system does with the document. Because forwarding a bill to an inbox that just stores files would be pointless — you'd just be moving the data entry problem to a different desk.

When a document arrives via email, it enters the same 7-phase AI pipeline that handles every other ingestion channel. The system queues it in the Document Hub with a tracked processing stage and estimated completion time. Multiple documents process in parallel — up to 10 at once — so a batch of monthly bills from 30 sites doesn't create a single-file bottleneck.

The AI extracts specific fields: date, material description, quantity, unit of measurement, dollar amount, currency, and supplier name. Each extraction carries a confidence score. High-confidence items proceed automatically. Low-confidence items get flagged for human review — the system doesn't silently guess when it's uncertain. That distinction matters more than people realise. A tool that's 95% accurate and knows when it's in the 5% is genuinely useful. A tool that's 95% accurate and presents everything with equal certainty is dangerous.

After extraction, the 5-tier material matching system maps each line item to the correct emission factor from the NGA Factors database, state-based grid factors, or other relevant sources. This is where knowing that "diesel — automotive" on a fuel receipt means NGA Table 4, energy content factor 38.6 GJ/kL, and emission factor 69.9 kg CO2-e/GJ actually matters.

Every step is logged. Original document, extracted values, matched factor, calculated emissions — the full chain. When an ASRS auditor asks to trace a number back to its source, you don't hand them a spreadsheet with a note saying "from site manager email." You hand them a document ID that links to the original PDF, the extracted fields, the confidence scores, and the factor match.

The eight formats nobody thinks about

Email ingestion would be simple if every document arrived as a clean PDF with a consistent layout. They don't. Fuel dockets come as photos. Waste manifests arrive in Excel. Some suppliers send CSVs. Older subcontractors occasionally send Word documents. We've seen PowerPoint files containing tabulated consumption data (don't ask).

Carbonly handles eight file formats: PDF, Excel, Word, CSV, PowerPoint, RTF, images, and ZIP archives. Images go through OCR via Tesseract before entering the AI pipeline. ZIPs get unpacked and each file processed individually. The 25MB per-document limit handles virtually everything except massive consolidated billing spreadsheets — those occasionally need splitting.

One technical detail that sounds boring but prevents genuine compliance headaches: SHA-256 file hash deduplication. If the same document gets forwarded twice — because someone wasn't sure if the first email went through, or because the accounts team and the site manager both sent it — the system detects the duplicate and doesn't double-count it. That sounds trivial until you consider a property manager with 50 sites who might have three people forwarding bills, occasionally for the same property.

Date parsing handles four Australian formats: DD/MM/YYYY, YYYY-MM-DD, DD Month YYYY, and Month DD YYYY. Currency extraction recognises A$, US$, EUR, GBP, NZD, and five other currency codes. These details aren't glamorous, but they're the difference between an extraction that works on real Australian documents and one that was trained on American invoices and falls over when it encounters DD/MM date ordering.

Email isn't the only channel — but it's the one that changes behaviour

We built three ingestion channels because different organisations work differently. Automated emissions data collection isn't one-size-fits-all.

OneDrive and SharePoint sync works well for companies with existing document management. Connect a folder — it crawls recursively up to 5 levels deep and picks up 17 supported MIME types automatically. Your accounts team saves files where they always have. The carbon system picks them up without any workflow changes.

Manual upload through the Document Hub gives you direct control when you need it. Drag and drop. Batch processing. Priority queuing for urgent documents.

API upload handles programmatic integration for companies with their own systems feeding data in.

But email is the one that changed real behaviour at the organisations we've worked with. And the reason is simple: it meets people where they already are. A site manager doesn't need training. They don't need a login. They don't need to download an app. They need an email address to forward something to. That's a text message from the sustainability team, not a change management project.

A property manager told us their biggest data collection problem wasn't processing bills — it was getting bills from building managers in the first place. They set up email forwarding rules with their utility providers and had incoming documents flowing into Carbonly within a day. No more quarterly chasing emails. No more "I'll send it Monday" conversations.

Where it breaks (because it does break)

We're not going to pretend this is perfect. Email ingestion has real limitations and we'd rather tell you what they are than have you find out after setup.

Heavily degraded documents fail. A photo of a thermal-printed docket taken in poor light, at an angle, after the paper has been sitting on a dashboard for three weeks — sometimes the OCR layer can't recover enough text for the AI to work with. The system flags these rather than guessing. You still need a human for the worst 5-10% of documents. We're still working on bringing that number down, but physics and optics set a floor.

The sender needs to use the right project email address. If someone forwards a bill to the wrong project, the emissions end up allocated to the wrong project. This sounds obvious, but it happens. Auto-forward rules eliminate the problem for recurring bills. Ad-hoc forwarding by site managers is where misallocation risk lives. We're honest about this one — it's a training problem, not a technology problem.

Processing isn't instant. Documents queue and process in order, with parallel batching helping throughput. A single document typically processes in minutes. A batch of 500 takes longer. If you're uploading 10,000 fuel receipts the week before the NGER deadline, expect processing time measured in hours, not minutes. Plan for this.

The 25MB file size limit exists for a reason (server memory, processing stability), but it means consolidated annual billing summaries from large energy retailers occasionally need splitting. This affects maybe 2-3% of documents in practice.

And email ingestion doesn't solve data that doesn't exist in documents at all. Refrigerant leakage logs, fugitive emissions from equipment, and some Scope 3 categories don't have utility bills attached to them. Those still require manual input or direct measurement. We cover which Scope 3 categories are document-tractable and which aren't.

The 60-80% number

Sustainability teams typically spend 60-80% of their time on data collection and entry. We've heard this ratio from prospect after prospect, and the BCG research supports it — data gathering dominates sustainability workflows across industries.

Email ingestion attacks the collection half of that problem. It doesn't eliminate all manual work. But it takes the highest-friction step — getting documents from the people who have them into the system that needs them — and reduces it to a forwarded email.

For a company with 30 sites generating 4 document types per site per month, that's 1,440 documents per year. At the 8-10 minutes per document we've seen in manual workflows (including search, download, open, read, enter, verify), you're looking at roughly 200 hours of data entry annually. Even if email ingestion plus AI processing only handles 70% of those documents cleanly (the rest need human review), you've recovered 140 hours. That's three and a half working weeks your sustainability team gets back.

Three and a half weeks to actually analyse emissions trends, build reduction strategies, or prepare for the assurance engagement that ASRS requires. Instead of copying numbers from PDFs.

One action to take this week

Set up an email forwarding rule for one utility type — say, electricity bills — from your main energy retailer. Forward the last three months to a test project. See what comes out the other side. If the extracted kWh matches what's on the bill and the emission factor is correct for your state, you've just validated the workflow in about 10 minutes.

Then ask yourself how long it took someone to do that manually last quarter.

Related reading: