By Ted Kornish
For the past two years, most of the conversation about AI in corporate sustainability has focused on the wrong capability. The demos were chatbots: ask a question about your carbon footprint, get a fluent answer. Impressive, but largely beside the point. A fluent answer built on incomplete or unverifiable data is just a faster way to be wrong.
What’s changed, especially in just the last six months, is what the technology can actually do. Early conversations in this space were about systems that could talk about your data. Today’s systems can go get the data themselves, clean it, reconcile it, and hand back something an auditor could stand behind. That jump – from answering questions to doing the work – is the real shift happening in sustainability right now, and it’s mostly invisible from the outside, because it isn’t happening in a chat window. It’s happening in the unglamorous middle of the workflow: where data gets collected, cleaned, matched, and reconciled. That is where sustainability teams have always lost the majority of their time, and it is where the operational economics of this field are being rewritten.
The work was never the carbon math
Here’s an uncomfortable truth about carbon accounting: the calculation itself is the easy part. Multiplying activity data by an emission factor is arithmetic. Carbon accounting is primarily a data problem – what consumes sustainability teams is everything upstream of that multiplication.
An enterprise-scale company receives utility bills from dozens of providers in dozens of formats: PDFs, scanned paper, portal downloads, and spreadsheets forwarded by a site manager. Fleet fuel data lives in one system, refrigerant logs in another, supplier information in email threads. Before anyone calculates anything, someone has to extract the numbers, standardize the units, catch the anomalies, fill the gaps, and map everything to the right facility and reporting period. The numbers behind this are stark: a BCG study found that 86 percent of companies still record and report their emissions manually using spreadsheets, and only 9 percent can measure comprehensively across all scopes. The GHG Protocol reports that 83 percent of companies making climate disclosures struggle to access the emissions data they need. Practitioners will tell you the day-to-day reality matches the numbers.
This is exactly the category of work that a chatbot can’t touch but an agent can. Document extraction, entity matching, anomaly detection, and classification aren’t things you ask about – they’re things you need done, end to end, without a person in the loop for every step. That capability is what’s matured. Applied to the sustainability workflow, it collapses tasks that took analysts weeks into hours. The teams adopting these tools are not replacing judgment with automation. They are removing the clerical layer that stood between their people and the judgment work.
There is no such thing as sustainability data
The deeper reason this shift matters is that it exposes something the field has been slow to admit: there is no such thing as “sustainability data”. There are utility bills, fuel invoices, procurement records, meter readings, and lease agreements. This is operational and financial data that happens to be useful for emissions reporting, and it was generating business value, or failing to, long before any disclosure framework asked for it.
When data collection was manual, companies could only justify the effort by pointing at a compliance deadline. The information got gathered once a year, transformed into a report, and shelved. Treating it as a special category (“sustainability data,” handled by the sustainability team, for sustainability purposes) guaranteed it would never be current enough or granular enough to inform an actual decision.
An agent that can do the collection and reconciliation on its own changes that calculus. When the data pipeline runs continuously instead of annually, the same records that feed a disclosure also surface the facility whose energy use jumped 20 percent after an equipment change, the sites where demand charges are dominating the bill, the fleet routes burning fuel against the trend. Roughly $2 trillion is spent inefficiently on energy every year. That waste was always visible in the documents, but nobody had the time to look – because “looking” used to mean a person, not a system that never stops.
This is the most underappreciated trend in the field right now: AI is turning reporting infrastructure into operating infrastructure. The companies treating emissions data as a byproduct of cost management, rather than the other way around, are getting both outcomes from one workflow.
Assurance will separate the serious tools from the demos
There is a second force at work, though, and it’s arriving on a schedule. Disclosure is shifting from something companies self-report to something a third party has to sign off on. Regulators are no longer satisfied with “trust us” – they want independent verification that the numbers are real, the same way a financial audit works. California’s climate disclosure program contemplates limited assurance over emissions data beginning with 2027 reporting, and European requirements under the Corporate Sustainability Reporting Directive already push companies toward audit-ready processes. Voluntary frameworks are moving the same direction. That shift – from self-reported to audited – is what will end up separating serious tools from demos.
Assurance changes what “good AI” means in this context. An auditor does not accept “the model said so.” Every number in an assured report needs a traceable lineage: this figure came from this document, extracted on this date, transformed by this method, reviewed by this person. A chatbot that produces a fluent, conversational answer has no such trail to offer – there’s no step-by-step record of work performed, because none was. An agent that actually did the extraction and reconciliation does leave that trail, because the trail is just a record of what it did. A hallucinated emission factor is not a rounding error; it is a defect that can invalidate a filing.
So, the industry is bifurcating. On one side are tools that use AI as a presentation layer, summarizing and rephrasing whatever data exists – chatbots wearing a sustainability skin. On the other are systems built for audits: AI doing the extraction and reconciliation itself, with confidence scoring, human review checkpoints, and an audit trail attached to every transformation. The first category demos alright. The second category survives an assurance engagement. Buyers are learning to tell the difference, often the hard way.
My advice to any team evaluating these tools is to ask one question early: show me the lineage of a single number, from source document to reported figure. The vendors who can answer in seconds are building agents that do the work. The ones who change the subject are still building chatbots for last year’s demo cycle.
The operational shift lands hardest, and most positively, on the teams themselves. The sustainability function has spent a decade absorbing more mandates without proportional headcount. Fragmented and unpredictable regulation, multiple overlapping frameworks, and growing customer data requests have turned many teams into full-time data janitors with a reporting deadline always in view.
When the work itself gets done by a system instead of a person, the composition of the job changes. Less time chasing a missing invoice; more time deciding which energy efficiency projects clear the hurdle rate. Less time reformatting the same data for a third framework; more time engaging suppliers whose numbers actually move the total. The skillset shifts accordingly: the valuable team members are the ones who can interrogate a model’s output, spot the anomaly that matters, and translate a data finding into a capital request the CFO will approve.
None of this requires believing any particular vendor’s story, including mine. The trend is structural. Document-heavy, rules-based, high-volume workflows get automated. Sustainability reporting is exactly that, so it is being automated. The only real questions for any individual company are sequencing and standards.
On sequencing: start with the data layer, not the report layer. Automating the final document while the inputs remain manual just produces polished versions of unreliable numbers. On standards: build for assurance now, even where it is not yet required, because retrofitting provenance into a workflow is far more expensive than designing for it.
The companies getting this right share a mindset more than a toolset. They treat the regulatory landscape as permanently unsettled and invest in capabilities that pay off under any version of the rules: clean data, traceable methods, and a workflow fast enough to inform decisions rather than merely document them. Agentic AI is what finally makes that posture affordable. The chatbot was never the story. What the technology can now actually go and do is.
The author is the co-founder and CTO of Gravity, where he builds AI systems that extract and structure carbon data for industrial businesses.
