From Research to Stewardship: The Next Step for AI in Responsible Investment

We asked a room of institutional investors at one of the world’s largest annual responsible investment gatherings, PRI in Person, where they saw the greatest potential for AI and alternative data. Stewardship and engagement took 71%, ahead of portfolio risk analysis at 55% and screening at 52%.

I did not expect that. Most of the discussion about AI in responsible investment has centred on research, screening and portfolio construction, and those are where the tooling is furthest along. Active ownership came top anyway, which points to where the headroom is. That fits the mood of the week in Amsterdam, where this year’s edition of the PRI annual conference gathered nearly 2,000 participants.

Investors are under growing pressure to show what their engagement achieves rather than how many meetings they held, and that is a measurement problem AI is reasonably well suited to.

The capabilities for that already exist. We built them to read disclosures and scan news coverage, because that is where the data was and where the first demand sat. What investors are asking for now is that we take them the rest of the way, so the process runs end to end rather than stopping at the point of analysis.

What it takes AI to do this job right is what we explored during the session I joined at PRI in Person: “The next generation of investment intelligence: putting alternative data and AI analysis to work.” I shared the panel with James Phare of Neural Alpha, Professor Peter Schwendner of ZHAW and PanGeo Nature Risk, and Tim Verheyden of Goldman Sachs Asset Management.

What unstructured data adds to an investment process

Preparing for the session, I tried to set out what unstructured data potentially can add to an investment process. I ended up with six benefits that, conveniently, spell MOSAIC.

• Multiplicity gives you perspectives that complement or challenge management’s account.
• Originality gives you signals beyond the evidence everyone already uses.
• Specificity gives you detail on assets and activities that company-level metrics miss.
• Availability fills gaps where disclosure is thin, which matters more now that the CSRD scope has narrowed.
• Immediacy brings evidence forward outside of the annual cycle, extracting data from quarterly calls or satellite.
• Context explains why a metric moved and whether the issue recurs.

Which of these you need obviously depends on the investment question you started with and what unstructured or alternative dataset you are working with. These are potential benefits, rather than a checklist that every dataset will deliver. For example, satellite imagery may add immediate detail about a particular asset, while news coverage may help identify emerging controversies that differ from the management’s storyline. The test is whether the source helps answer the investment question.

What unstructured data can tell investors

Two pieces of our own research give a sense of what this looks like in practice.

In the first, we used AI to pull text and quantitative data from corporate disclosures across 900 firms in high-emitting sectors reporting green CapEx. Companies with targets aligned to 2°C or lower put an average of 27% of CapEx into green projects, against 25% for those with no stated target. Target ambition on its own tells you less about where capital is going than you might assume.

In the second, we used natural language processing to identify sustainability controversies in news coverage, then examined more than 12,000 incidents. Estimated six-month underperformance against comparable companies ran from around 2% for lower-severity incidents to 5% for high-severity ones.

The studies answer different investment questions. The CapEx analysis tests whether climate ambitions are reflected in capital allocation; the controversies study examines the relationship between sustainability incidents and subsequent share-price performance.

Both show how unstructured data can help investors test assumptions and identify questions to put to companies. These studies sit on the research and screening side, because that is where the work has been done. The step investors want now is to carry that standard of evidence into the conversations they have with companies.

That is what our Climate Intelligence solution, developed by our new colleagues from Glass Lewis, is built for. It combines disclosures, AI-supported analysis and expert methodology to assess transition strategy and execution, turning findings into specific questions to put to management.

AI design and governance in responsible investment: what it takes

First and foremost, start with a clear investment purpose: what decision are you trying to inform, or what do you want to achieve? That purpose should guide the choice of data and tools. Unfortunately, I still see lots of players getting excited by AI and starting with the tool. No surprise so many AI projects seem to fail to deliver on their investments!

As importantly, delivering on the goal requires three key capabilities working together:

• Good quality datasets that are relevant to the question, with sufficient coverage and reliable sources.
• Sustainability subject matter expertise to establish what matters and interpret the evidence in context.
• AI and technology competencies to build, test and monitor the tools used to extract and analyse that evidence.

Finally, the outcome works as long as governance carries the same weight, and I would group that under four pillars:

1. Integrity covers how data is sourced and secured, along with copyright, reliability, and bias.
2. Accuracy asks whether the model extracted the right thing, and whether it matches the definitions, scopes, and time periods you are working with.
3. Transparency means the sources, methods, and assumptions are visible, including where a number is an estimate.
4. Accountability makes clear who is responsible for individual analytical choices, who oversees the process, and who is accountable for the system as a whole. Those responsibilities may be spread across investment, sustainability, and technology teams, as well as external providers. They need to be explicit, including who can challenge an output or change the system.

That division of responsibilities also applies when a provider supplies extracted evidence, analytical assessment, or both. The provider is responsible for the services it delivers, while the investor remains accountable for how those outputs are used and for the resulting investment or stewardship decisions. The investor should always own the decision.

Build or buy AI capability? Increasingly both

Prof. Dr. Peter Schwendner’s research shows how far open data now reaches. His team screened tailings dams across a portfolio using public global datasets for elevation, river flow paths and population, with no company disclosure at all, and found that the size of a dam predicts the consequences of a breach far less well than the number of people living downstream.

Investors are taking on this kind of question in-house as well, including the financial implications of physical climate risk and adaptation. Extraction and modelling have stopped being the exclusive domain of providers. Even so, investors at the conference shared their empirical evidence that something like 80% to 90% of these insights still reach investors through a provider, and specialised partners clearly retain a role.

Where an organisation draws that line will vary. What should not vary is the procurement process behind it, which needs to check the data, the expertise and the AI governance of whatever you buy.

Where AI in responsible investment goes next

The speed benefit is the easy one to see: familiar work, done faster and cheaper. I am more interested in the second one, seeing what we would otherwise miss through wider coverage and connections across evidence. Processing will keep getting easier, while missing evidence and uncertainty stay where they are, so judgement and interpretation only become more valuable.

The capability is largely there. It has mostly been aimed at the parts of the process that were easiest to automate. Point it at stewardship, and the evidence has to meet the same standard in a different form. It gets discussed with the company it describes, so it needs to be specific enough to ask about, and it needs to stay on the record, so last year’s commitment can sit next to this year’s.

That is a harder product to build and a harder one to evaluate. Judging by the audience poll in Amsterdam, it is the one investors are asking for.