On September 30, we co-hosted an executive breakfast with Inrate in Zurich. The discussion kept coming back to one idea. Double materiality is no longer held back by a lack of data. It is held back by the judgment needed to make sense of data that is vast, contradictory, and often late. We asked Saurabh Srivastava, Managing Director at Inrate, and Sylvain Forté, CEO of SESAMm, to pick up where they left off. How does a controversy become a rating? How fast should that happen? And what can AI do along the way, and what can it not?
Saurabh, you split the data problem into three: contradiction, fragmentation and contested data. What does each look like for a rating agency?
Saurabh Srivastava, Inrate: All three come up every day. Contradiction is when the same emissions figure differs between a company's sustainability report, its CDP response and its local filings. Fragmentation is subtler: financed emissions, for example, are spread across asset classes, so no single source gives you the full picture. Contested data is the hardest of all. A supplier dispute is reported differently by each party, and a court may only settle it years later.
Each of these needs a judgment call, and each judgment has to be traceable and auditable. That's why we use AI for triangulated quality assurance, checking every data point against several independent sources, while an analyst makes the final call and documents it.
Sylvain, which of those can AI actually solve today, and which still need a human?
Sylvain Forté, SESAMm:
Fragmentation is where AI helps most, because it's a problem of scale. Pulling scattered information together across millions of companies and around 100 languages is not something a team can do by hand, and no one should be writing summaries by hand anyway.
On contradiction, AI gives you an independent view. Companies often report much later than what is already visible in the public domain, so what journalists and NGOs are saying is a useful check on what a company discloses.
Contested data is where the story keeps moving. Take a local NGO accusing a factory of polluting a river. The story may only appear in the local press, and it isn't always clear which company the factory belongs to. AI finds it, links it to the right company and keeps following it as it develops: whether a fine is imposed, whether a court rules, how many people are affected. All of that feeds a first severity score.
Throughout this process, humans stay in the loop at two levels. Every day, our ESG analysts conduct quality control and use what they find to keep refining the algorithms. Reviewing everything by hand would be impossible, and that's the whole point of AI at scale. And the final judgment call stays with the people using the data. A secondaries investor that inherits 500 companies can use AI to narrow the list down to five, then make the call on those five itself. Our principle is AI for scale, humans for validation.
Saurabh, Inrate sources its controversy data from SESAMm. What happens once a signal reaches you?
Saurabh: What we receive from SESAMm is the events flow: controversies picked up across a huge range of sources, often in the local language, matched to the right company, classified by risk type, and given a first severity score. Our job is to translate those signals into what they mean for our rating, and that's where our own controversy framework comes in. We look at how severe and credible an event is, how long-lasting it is, and whether the company has taken corrective action.
A controversy is also never assessed in isolation. It's one of three pillars in our rating, alongside our business activity analysis and our disclosure analysis. Controversies matter because they show what's happening outside of what a company reports. You can trust disclosures, but you need to verify them. Put simply, SESAMm tells us what is happening, and we decide what it means for the company's impact and risk.
Sylvain, controversy signals tend to move before ratings do. Is that a problem for ratings?
Sylvain: The evidence does point that way. Several studies by our quantitative clients found that controversy signals tend to anticipate ESG rating downgrades, with statistically significant results.
But I wouldn't read that as ratings being slow. Ratings are built to be stable because investors use them to allocate capital, and some facts stay contested for years. The value is in having both: an early signal for monitoring, and a stable rating you can build a portfolio on. What matters is that a serious controversy reaches the people making the rating quickly, so they can decide whether it warrants a review.
Saurabh, you called the falling cost of assembling impact evidence "the most interesting thing happening for us today." Why?
Saurabh: Because of the contrast with financial materiality. There, we were already well served by AI and machine learning: the information was structured, and there was always a strong P&L reason to use it. Impact is different. No company publishes its externalities, and none ever will, so you have to piece that picture together from local-language filings, trade documents and many other sources. Not long ago, much of that was collected and analyzed by hand. Now a lot of it reaches us already indexed and ready for an analyst to use, so the cost of assembling that evidence has come right down. Add growing access to geospatial data, and we can back an impact claim with far more evidence, and far more confidence, than we could a few years ago.
Sylvain, where does local-language data add the most to that picture?
Sylvain: Inrate's ratings start from what companies do, not only from what they disclose, and that's where local-language data adds the most. A dispute at a supplier or a protest against a project is usually reported first in the local press, and sometimes only there.
It matters even more where there's no disclosure at all. Private companies and infrastructure projects often publish no sustainability report. For due diligence on them, outside-in data is usually enough, even if it isn't enough for a full score. Infrastructure is a good example of demand running ahead of supply: insurers in particular want to understand underwriting risk on projects that traditional coverage barely reaches.
Saurabh, disclosure is shrinking, with the EU Omnibus narrowing reporting and some Swiss large caps disclosing less. Can providers close the gap?
Saurabh: Partly. The duty is shifting to providers, and better estimation, for example, on scope 3 categories, and growing access to geospatial data make up for part of what's being lost. But there are limits. Biodiversity and physical risk are local by nature, and an estimate can't fully replace information about a specific site.
The further down the value chain you go, the darker it gets. Tier-3 and tier-4 suppliers are still largely opaque, for investors and often for the companies themselves.
Sylvain, can AI shed light on those deeper tiers?
Sylvain: It's the only realistic way. A company with 1,000 suppliers already has over a million relationships by tier 3, and risk travels through those chains, from private suppliers into public portfolios. No team can map that by hand. It takes AI-assisted coverage and navigation.
And it's happening at an odd moment. Corporate sustainability reporting is being scaled back, while ESG rating providers now face new authorization and transparency rules. Providers are being asked to fill more gaps and, at the same time, to be more open about how they fill them. That makes the combination we've described, AI for scale with accountable human judgment on top, more important, not less.
About SESAMm. SESAMm is an AI company specializing in controversy monitoring and ESG risk. It analyzes over 35 billion documents in around 100 languages to cover 5 million companies and infrastructure projects across public and private markets. sesamm.com
About Inrate. Inrate is an independent Swiss sustainability rating agency. Its impact ratings, screening, regulatory and thematic datasets, and stewardship services start from what companies do, not only what they disclose. inrate.com
.avif)
.png)