Every carbon accounting platform now claims to be “AI-powered”. Strip away the marketing language and a more nuanced picture emerges. AI is genuinely changing parts of carbon management, but it isn't a magic wand. Knowing what it can and can't do is essential when you choose your carbon accounting tools.
The current state of AI in carbon management
In carbon accounting, AI isn't a single technology. It's an umbrella term for several distinct capabilities, each at a different stage of maturity. The most useful applications today fall into four areas: natural language processing for data input, document intelligence for extraction, emission factor matching, and automatic data-quality checks.
These aren't theoretical. They're in use today on real companies' data. But it's worth understanding each one in detail, to know where AI adds value and where human judgement remains essential.
What AI can actually do today
Natural language data input
Traditional carbon accounting asks users to work through complex forms, choosing emission sources, fuel types, units and factors from dropdown menus. Natural language processing lets users describe an activity in plain words, such as “2 diesel vans doing 20,000 miles each per year”, and turns it into structured data.
This is more than a convenience. It lowers the barrier to entry for carbon accounting, so people who aren't specialists can contribute data accurately. Modern language models can handle ambiguous inputs, infer missing context and ask clarifying questions when they need to.
Document intelligence and extraction
Organisations generate huge numbers of documents that hold emissions-relevant data: utility bills, fuel receipts, logistics invoices. Transcribing them by hand is tedious, error-prone and expensive.
AI document reading can:
- Extract the key data points from invoices and bills, including quantities, units, dates and supplier details.
- Read tables from PDFs and scanned documents.
- Find relevant figures in unstructured text, such as a narrative report.
- Handle several document formats and languages.
Emission factor matching
One of the most error-prone steps in carbon accounting is choosing the right emission factor for an activity. With thousands of factors across government and commercial databases, each updated on its own cycle, choosing the right one takes expertise.
AI can match an activity to a suitable emission factor from a plain description, taking into account the type of activity, geography, time period and how specific the data is. That turns a job that needed a specialist into a suggestion a person can check, as long as the tool shows which factor it chose and how confident it is.
Automatic data-quality checks
Software is good at spotting patterns, which makes it well suited to finding unusual data points. Has a supplier reported emissions ten times higher than last year? Has electricity use dropped to zero at a site that should be operating? These might be data entry errors, changes of method or genuine operational shifts. Automatic checks can flag them for a person to review.
The best AI systems don't replace human judgement. They amplify it, by surfacing the signals that matter from the noise of raw data.
Where AI still falls short
For all its promise, AI has clear limits in carbon accounting that responsible practitioners must acknowledge:
Methodology decisions
Carbon accounting involves many choices of method: organisational boundaries (operational or financial control), how to allocate shared emissions, how to treat biogenic carbon, and which global warming potentials to use. These are judgement calls that need an understanding of the business, the regulations and what stakeholders expect. AI can present the options, but a person must make the decision.
Data completeness
AI can't create data that doesn't exist. If a company doesn't track its refrigerant losses, no amount of AI will produce accurate Scope 1 figures for leaked gases. AI can estimate from proxies, industry averages and modelling, but estimates aren't the same as measured data, and the difference matters for assurance.
Regulatory interpretation
Sustainability regulation changes quickly and is often open to interpretation. Whether an activity falls within the scope of the EU's Corporate Sustainability Reporting Directive (CSRD), how to apply the phase-in provisions of an ESRS standard, or whether a green claim would stand up under consumer protection law: these questions need legal and specialist expertise that AI can't reliably provide.
Assurance and accountability
Auditors and assurance providers need to understand and check the method behind reported figures. “The AI did it” isn't an acceptable answer. Any AI-powered system must keep a clear record of how inputs were processed, which factors were applied and what assumptions were made.
The future outlook
AI in carbon accounting is moving quickly. Several developments could widen what it does:
- Real-time emissions monitoring: links to sensors and operational systems could let emissions be calculated almost as they happen, rather than from periodic data collection.
- Predictive analytics: models could forecast emissions from planned activities, purchasing decisions and operational changes, so carbon management becomes proactive rather than reactive.
- Supply chain intelligence: as more companies report emissions, models of supply chain emissions could become more accurate, reducing reliance on spend-based estimates.
- Regulatory mapping: AI could map emissions data to the disclosure requirements of several frameworks at once, cutting the duplicated effort of reporting to each one separately.
Practical advice for adopting AI in carbon accounting
If you're evaluating AI-powered carbon accounting tools, these are the questions that matter:
- What specific tasks does AI handle? Be wary of vague claims. Ask for concrete examples of how AI is used in the product and where human input is still needed.
- How transparent is the method? Can you see which emission factors were applied and why? Can you override the AI when your judgement differs?
- What is the audit trail? Every AI-assisted calculation should be traceable to its inputs, assumptions and data sources. If it's a black box, it won't survive assurance.
- How does it handle uncertainty? Good AI systems show how confident they are. A spend-based Scope 3 estimate (Scope 3 being the emissions in your supply chain) shouldn't be presented with the same confidence as a directly measured Scope 1 figure.
- Does it learn from corrections? Ask what happens when you fix a reading. The best systems remember the correction so you don't make it twice.
AI is a powerful tool in carbon accounting, but it's exactly that: a tool. The companies that lead on climate action will be those that combine AI's efficiency with human expertise and a real commitment to data quality.
At Zoru, AI reads your bills, invoices, receipts and spreadsheets, turns an activity you describe in your own words (in English, French, German or Spanish) into an entry in the Scope 1 and 2 sections and Scope 3 categories 1 to 7, and drafts action plans for you to edit. Zoru uses Anthropic's Claude models with its own instructions and trains none. While you review, each reading shows a confidence level and the reasons for it, and lines Zoru isn't sure about are held rather than guessed. When you import a spreadsheet or ledger, automatic checks flag outliers, duplicates, missing months and big year-on-year changes, and corrections you make to an uploaded spreadsheet are remembered. Every entry starts as a draft: it counts in your totals, marked not yet reviewed, until someone approves it, and the year can't be closed until every draft is reviewed.