Grounded AI for ESG company profiling

Challenge to overcome

Sustainable investing has never had more information available to it and never had a harder time turning that material into conviction.

Greater emphasis on corporate transparency has expanded the information available to investors. Driven by the EU's Corporate Sustainability Reporting Directive and the evolving standards of the ISSB, a single company's sustainability reporting now routinely exceeds a hundred pages. But volume is not clarity. Jurisdictions move at different speeds, and sustainability reporting requirements vary across markets. Investors with a global scope therefore need a consistent analytical framework that can scale across these different reporting environments.

At the same time, the incentive to overstate sustainability accomplishments has grown. Greenwashing today is rarely outright misrepresentation. It is subtler: cherry-picked data, vague long-term goals unsupported by credible transition plans, prominence given to immaterial initiatives. Each shifts the burden of proof onto the analyst.

The result is an analytical bottleneck. Analysts spend their scarcest hours locating and transcribing information rather than interpreting it, and the risk of relying on superficial metrics rises exactly when deep, material analysis matters most.

The industry knows what is holding it back. In Swiss Sustainable Finance's own market study, the most cited barrier to adopting AI in sustainable finance is the reliability of outputs, identified by 76% of respondents, ahead of system integration, data quality, cost, ethics and regulatory uncertainty. Building confidence in AI-generated analysis is therefore a central challenge for adoption.

That framed the problem we set out to solve: not how to make an analyst faster, but how to make a machine's contribution to an investment view something a professional can verify, defend and sign their name to.

AI application / use case

Bank J. Safra Sarasin has developed and deployed an in-house AI-assisted ESG profiling tool. For any company in our research universe, it produces a structured SWOT analysis - strengths, weaknesses, opportunities and threats - for each material ESG key issue, from board independence to human capital development to product safety. It has been in production use by our sustainable investment analysts for around a year.

The workflow runs in five steps.

1. The analysis is scoped. Each company is assessed against the ESG key issues our own methodology identifies as material for its sub-industry, so the questions asked of a semiconductor manufacturer are not the questions asked of a regional utility.
2. The system assembles the evidence. It draws on two sources at once. The first is our own proprietary view: our internal ESG scores, how the company sits against its sub-industry peers, our reputational risk assessments, and the record of our own stewardship engagement and our analysts' meetings with company management. The second is what the company itself publishes: its sustainability and annual reporting, its filings, and whatever other material is relevant to the question at hand.
3. It drafts the analysis, with a source for every statement. The system works as a research assistant rather than an author: it builds its analysis only from the evidence in front of it, and each point it produces ends with a reference either to our internal data or to a specific document and page. An analyst can move from a conclusion to the sentence that supports it in seconds.
4. The draft is checked before the analyst sees it. A separate verification step reviews every individual point and assesses whether the claim is genuinely supported, whether the cited source actually contains it, and whether it is relevant to the issue at hand. Where the result falls short of our standard, the system makes another drafting attempt. Every draft undergoes verification before reaching an analyst, who remains responsible for accepting or challenging its conclusions.
5. The analyst validates and decides. The output is treated strictly as a draft. The analyst confirms the facts and challenges the narrative. The judgement calls stay human: is the target credible, is governance sound, and what does the evidence mean against the wider market and geopolitical backdrop?

The individual assessments come together into a company-level SWOT, which analysts refine in conversation with the tool and export directly into our ESG profile format, for use in investment discussions and client reporting.

Each analysis is also retained. The evidence gathered, the reasoning applied and the analyst's conclusions are retained in a form the firm can build on, so that understanding of a company deepens across reporting cycles rather than being assembled from scratch each time.

Figure 1: From manual effort to strategic insight - the reallocation of analyst time from data collection and drafting toward validation, insight and review. Proportions are illustrative.

Use case key beneficiaries

☐ Relationship Managers

☒ Portfolio Managers

☒ Research teams, macroeconomists

☐ Control functions

☐ Support functions (HR, CFO, …)

☐ Other: Clients

Benefits of the AI use case for the financial services sector

A sourced starting point for analysis. Our analysts estimate the tool removes most of the time previously spent producing a first draft. Starting from structured, referenced material leaves more time to challenge company disclosures and assess their investment implications.

Broader company coverage. Reducing the effort required to collect and organise information can make in-depth research viable for more companies, including smaller businesses and those in less well-covered markets.

Traceable evidence. Source references help analysts test claims against the underlying disclosures, including whether an ambitious target is supported by a credible transition plan. Easier verification helps expose subtle overstatement.

Proprietary data stays proprietary. Within this workflow, client data, internal methodology and licensed research remain on the bank’s infrastructure. This controlled environment allows analysts to combine company disclosures with proprietary ESG scores, research and engagement insights without sending sensitive information to an external AI service.

The competitive edge shifts from data to process. As sustainability disclosures become more widely available, competitive advantage increasingly depends on how a firm combines them with proprietary research and analyst judgement to form a defensible investment view.

Supporting technology

The system runs on the bank’s own AI infrastructure, with language models accessed through a secure internal gateway.

Hosting models internally allows company disclosures to be analysed alongside proprietary ESG research without sending that research to an external AI service. The bank retains control over how this information is processed and which models are used, allowing the workflow to reflect its own analytical methodology and governance requirements.

The workflow described above is supported by task-specific AI agents and configurable search tools. JSS’s proprietary methodology determines the material ESG issues for each sub-industry, while analysts can refine issue definitions, search terms and selected evidence to reflect company-specific circumstances and their professional judgement. The agents use the same language model, with dedicated instructions for each task. The main technical components are:

• Evidence retrieval. Semantic and keyword searches identify information relevant to the company’s material ESG issues. Industry- and company-specific vocabulary helps locate evidence across differently worded disclosures, while analysts can direct attention to particular topics or sources.
• Specialised analysis agents. The agents interpret internal ESG signals, extract company-disclosed evidence and synthesise the findings. Each has a defined task and structured output requirements, including source references and instructions to acknowledge missing evidence.
• Automated verification. A verification agent assesses evidential support, citation correctness and relevance, complemented by deterministic checks of supporting quotations and numerical values. Factual statements and analytical interpretations receive different checks. Results below the verification threshold trigger another drafting attempt, within a fixed limit, and verification findings remain available for analyst review.
• Profile synthesis and refinement. A synthesis agent consolidates the issue-level findings into a company-level SWOT draft. Through the tool, analysts can refine its emphasis, classification and wording in conversation with the system, then export it into the bank’s ESG profile format for review and use in investment discussions.