How Carbon+Alt+Delete uses AI in carbon accounting software

Carbon accounting is entering a phase of rapid scaling. Demand for greenhouse gas reporting is increasing due to regulation, investor expectations, and internal climate strategies. At the same time, the number of qualified carbon accountants is not growing at the same pace.

Internal analysis shows that both market demand and required capacity will need to increase significantly in the coming years . This creates a structural challenge: how to deliver more carbon footprints, with consistent quality, without proportionally increasing manual effort.

At Carbon+Alt+Delete, the working assumption is that this gap cannot be closed without automation—specifically, without applying AI to the most time-intensive parts of carbon accounting.

Where AI creates value in carbon accounting

AI is not a replacement for carbon accountants. Its value lies in supporting the repetitive, data-heavy parts of the workflow that typically consume most of the time in carbon accounting projects.

In practice, the highest impact is seen in:

AI is particularly effective in situations where data is both unstructured and repetitive. Invoice processing is a good example: large volumes of documents with similar patterns, but varying formats and levels of detail. This combination makes it well suited for machine learning and large language models.

From spend-based to activity-based carbon accounting

A key design choice is to focus on activity-based carbon accounting rather than spend-based approximations.

This requires extracting real activity data directly from primary sources such as invoices. Invoices are particularly relevant because they provide a centralised and relatively complete view of company operations. They are also closely linked to actual activities across Scope 1, Scope 2, and parts of Scope 3.

However, invoices are not designed for carbon accounting. They contain unstructured text, inconsistent terminology, and varying levels of detail. This makes manual processing time-consuming and difficult to standardise.

AI is used to translate this unstructured input into structured, usable carbon data.

How AI processes invoices into emissions

The AI model translates invoices into emissions through three sequential steps.

The model interprets the content of the invoice and determines which of the GHG Protocol activities it relates to. This requires understanding descriptions that are often incomplete or ambiguous and mapping them to categories such as fuel use, electricity consumption, or business travel.

Rather than relying on financial spend, the model extracts physical activity data, such as kilowatt-hours of electricity, litres of fuel, or kilometres travelled. This distinction is important because activity-based data allows for more accurate and actionable carbon accounting.

Based on the detected activity and extracted data, the model selects an appropriate emission factor. This is a complex task, as there are thousands of possible emission factors with varying levels of specificity, and there is often no single obvious match—particularly in Scope 3 contexts .

What the model can do today

Based on a dataset of more than 200 invoices, the current performance of the model provides a realistic view of what AI can already achieve:

Performance is not uniform across all categories. The model performs particularly well for Scope 1 and Scope 2 emissions, such as stationary combustion, mobile combustion, and electricity. Business travel also shows relatively high accuracy. In these areas, accuracy exceeds 80%, which means that most of the work can already be automated.

Scope 3 categories remain more challenging. Activities such as purchased goods and services or waste processing often lack clear, structured information, which makes consistent interpretation more difficult.

Why AI fails: context over volume

An important observation is that errors are rarely driven by incorrect quantitative data. In many cases, the model correctly captures volumes such as energy use or fuel consumption.

The main source of error lies in contextual interpretation. This includes situations where the model selects an emission factor that does not fully match the specific context, misinterprets contractual details such as renewable electricity, or incorrectly categorises the nature of a transaction.

This highlights a broader limitation of current AI systems. While they are effective at recognising patterns and extracting structured information, they still depend on the clarity and completeness of the input data.

It also explains why Scope 3 remains more difficult. Real-world transactions are often ambiguous, and invoices do not always provide sufficient context to resolve that ambiguity.

Efficiency gains in practice

Despite these limitations, the efficiency gains are substantial.

In the case study:

This does not mean that the AI produces a final, audit-ready result without human involvement. Instead, it generates a structured first version of the footprint, including extracted activity data and initial emission factor selections.

The role of the consultant then shifts toward validating the output, correcting edge cases, and ensuring consistency.

What this means for sustainability consultants

For sustainability consultants, AI changes the nature of the work rather than replacing it.

Instead of spending most of the time on data collection and manual processing, consultants increasingly focus on validation, interpretation, and methodological decisions. This includes reviewing AI-generated outputs, handling complex or ambiguous cases, and ensuring that results are robust and auditable.

At the same time, consultants need to understand how AI systems operate, including their limitations and typical failure modes. This becomes essential for maintaining quality and credibility in carbon accounting outputs.

Conclusion

AI in carbon accounting is no longer theoretical. It is already capable of automating a significant share of the workflow, particularly for structured data sources and Scope 1 and 2 emissions.

The main limitation remains Scope 3, where data availability and contextual complexity continue to constrain automation.

The overall direction is clear. Carbon accounting is moving toward a hybrid model in which AI handles data-heavy processes, while consultants focus on validation, interpretation, and decision-making.

Given the expected growth in demand, this shift is necessary to make carbon accounting scalable across organisations.