AI Emission Factor Matching: How It Works, When to Override It and What Happens to Your Data

Halfway through a platform evaluation, someone from IT or procurement asks the question that decides the shortlist: "Show me why the AI picked this factor for this line. Can I change it? And is our spend data being used to train your model?"

These questions came up again and again in evaluations we took part in this month, from food and pharma to chemicals distribution. They are the right questions. Emission factor matching is where most of a Scope 3 inventory's accuracy is won or lost, and AI only helps if you can see and control what it does.

What is emission factor matching?

Emission factor matching is the step that links each activity record, such as an invoice line, a purchase order or a shipment, to an emission factor that converts it into CO2e. For purchased goods and services, many companies start with spend-based factors, which express emissions per unit of money spent.

The choice of factor matters because databases are granular. The US EPA's Supply Chain Greenhouse Gas Emission Factors v1.3, for example, cover 1,016 US commodities defined by the 2017 NAICS classification, expressed in kilograms of CO2e per 2022 US dollar at purchaser prices, with separate factors with and without margins (US EPA via Data.gov). Picking the wrong commodity, or the wrong margin treatment, can shift a category's emissions significantly even when the spend data is perfect.

Doing this by hand across tens of thousands of lines is slow and inconsistent. That is the problem AI matching is meant to solve.

How does AI emission factor matching work?

Most AI matching follows the same four steps, whatever the vendor:

1. Normalise the input. Clean supplier names, item descriptions, units and categories, often in several languages.
2. Classify the activity. Map each line to a product or service category, using the description, the general ledger account, the supplier's industry and the unit.
3. Retrieve and rank candidate factors. Pull candidate factors from one or more databases and rank them by fit, including geography, year and unit compatibility.
4. Score confidence and route. High-confidence matches are applied; low-confidence or high-impact matches go to a reviewer.

The quality difference between tools sits mostly in steps 2 and 4: how well the classifier handles messy descriptions, and whether the system knows when it is unsure.

Can you see and override AI matches?

You should be able to, and you should insist on it. The GHG Protocol's accuracy principle asks that emissions be "systematically neither over nor under actual emissions", and its transparency principle requires "a clear audit trail" (Deloitte summary of the GHG Protocol principles). A match nobody can explain fails both tests.

Use these questions when you evaluate any AI matching capability:

• Can I see why each factor was chosen? Why it matters: Auditors and reviewers need the reasoning, not just the result; What a good answer looks like: Matched category, factor source and version, and the input fields used.
• Is there a confidence score? Why it matters: You need to know where to focus review time; What a good answer looks like: Low-confidence and high-emission lines are flagged for review.
• Can I override a match and keep the override? Why it matters: Corrections must survive the next data load; What a good answer looks like: Overrides become rules applied to future lines, with an author and date.
• Can I bring my own factor library? Why it matters: Many companies have sector-specific or internal factors; What a good answer looks like: Custom factors sit alongside public databases with clear precedence rules.
• What happens when a supplier provides its own data? Why it matters: Supplier-specific data should replace averages; What a good answer looks like: Supplier values take precedence, and the change is traceable.
• Is my data used to train models for other customers? Why it matters: Spend data is commercially sensitive; What a good answer looks like: A clear contractual answer, not a marketing statement.

Comparing AI carbon accounting tools? See the four questions chemical companies ask about AI-powered carbon footprinting.

Is your data used to train a generic AI model?

That depends entirely on the vendor's terms, so ask for it in writing. Buyers increasingly treat this as a security question rather than a sustainability one, and it belongs in your RFP alongside hosting, access control and data retention.

Three points to clarify:

• Training scope: is your data used only to improve matching for your own tenant, or pooled across customers?
• Consent: is any pooled use opt-in, and can you withdraw it?
• Third-party models: if a general-purpose AI model is involved, do the vendor's agreements with that provider exclude your data from training?

Review the vendor's security documentation, such as CO2 AI's platform security page, and have the answers reflected in your contract.

When should you bring your own emission factors?

Bring your own factors when you have better information than a generic database: supplier-specific product footprints, industry association factors for your sector, or internal LCA results for key materials. Generic databases are a starting point, not an end state.

The GHG Protocol Scope 3 Standard requires companies to report, for each category, "the percentage of emissions calculated using data obtained from suppliers or other value chain partners" (GHG Protocol Scope 3 Standard). A matching system that lets supplier and custom factors override averages, with a record of each change, is how that percentage grows over time.

Expect arbitration cases. A supplier may report a value that differs sharply from the database average. Neither number wins automatically: check the supplier's boundary, allocation and year before accepting it, and keep both values on record.

What AI matching looks like in practice with CO2 AI

One global specialty ingredients manufacturer tracks around 21,000 purchased raw materials. It holds itself to an internal footprint quality KPI: 95% of materials matched to a primary or secondary emission factor, and never below 90%. About 100 new material codes enter its inventory every month, so each quarterly refresh lands several thousand unmatched lines. Most are not new at all. They are known materials under a new ERP code, a new supplier or a spelling variant.

Matching runs in three layers:

1. Validated rules first. Existing approved matches are applied.
2. Fuzzy matching. Emission factors the company has already validated are copied onto spelling variants of the same material and supplier. This layer never picks a new factor.
3. AI matching for truly new materials. Each suggestion comes with its reasoning and a confidence score. Below the threshold, the system returns “no match” rather than a weak guess, and the material goes to the company’s LCA experts to model.

The experts then review only what moves the number, roughly 30 to 40 materials a month ranked by quantity. Approved matches become rules for the next refresh. The point is not to match everything. It is to protect coverage without forcing a poor factor onto a material that is bought in small volumes but carries a very high footprint.

How does CO2 AI approach emission factor matching?

CO2 AI is an AI-powered carbon management platform used by enterprises such as Volvo Group, Klöckner & Co and Accor to measure and reduce corporate and product emissions. Its AI capabilities match activity data to emission factors with transparent, editable matching rules, support custom factor libraries alongside public databases, and route uncertain matches to reviewers so that every override is recorded.