A senior partner at a mid-market private equity firm recently told me they had stopped reading the ESG sections of their portfolio companies’ quarterly reports.
The numbers had stopped meaning anything. Every score was generated by a vendor’s AI system, and every score moved a few points each quarter for reasons no one in the firm could explain.
Across private equity, AI-generated ESG outputs are flowing into LP reports, fundraising decks and regulatory filings with little scrutiny. AI models scrape public data, run sentiment analysis and output aggregate scores that are then reported unchecked. GPs who fix this before the 2027 vintage cycle will raise on better terms. GPs who do not are carrying unpriced model risk on the disclosure that LPs scrutinize most closely.
Over two years advising private equity firms on AI deployment and governance, I have watched the most technically impressive systems carry the highest governance risk. Their outputs look authoritative and arrive faster than anyone can verify them.
The way forward is AI that surfaces questions for human investigation rather than issuing verdicts.
The score trap
ESG scores had a credibility problem long before AI entered the room. MIT Sloan researchers labeled it “aggregate confusion:” ESG ratings from six major providers, including MSCI, S&P Global and Sustainalytics, correlate at 0.54 on average. Put plainly, pick any company and any two of the big raters will score them roughly the same only about half the time. In contrast, credit ratings across major agencies correlate at 0.92, meaning they almost always agree. ESG scoring is closer to art criticism than to accounting.
AI accelerates the problem rather than solving it. A typical agentic ESG monitoring system ingests public filings, news flow, NGO reports and portfolio submissions, runs them through a vendor’s taxonomy and produces a portfolio-level score. The system looks impressive in a meeting but is difficult to audit afterward.
Regulators have policed this kind of gap before, when people rather than models produced the numbers. The Securities and Exchange Commission fined Goldman Sachs Asset Management and BNY Mellon over ESG misstatements in 2022, and German prosecutors raided DWS, Deutsche Bank’s asset management arm, in the same year. None of those cases turned on AI. They turned on funds marketed as ESG-screened when the process behind the label did not hold up. The inputs have changed since. US federal enforcement has cooled, and in September 2024 the SEC disbanded its Climate and ESG Task Force. The scrutiny that remains is moving to the people who write the checks.
The LPs are the auditors now
Limited partners have built the machinery for this. More than 500 GPs and LPs now report a common set of ESG metrics through the ESG Data Convergence Initiative, the standardization effort Carlyle and CalPERS started in 2021. It exists because LPs were burying GPs under a growing volume of their own ESG data requests, and it collects operational metrics (Scope 1 and 2 emissions, board diversity, work-related injuries, net new hires) rather than producing another score. On the diligence side, the ILPA due diligence questionnaire has become the standard most funds answer, now paired with a climate module built with the PRI.
What few LPs ask yet are the AI-specific questions: which of the numbers a model produced, and who checked it. That gap is where the next round of fundraising friction will open.
AI that asks questions
The firms getting this right have inverted the workflow. Instead of asking AI to generate a score, they ask it to surface anomalies that warrant human attention.
For example, one European GP I work alongside flags any portfolio company whose monthly energy consumption deviates more than 15% from its 12-month baseline, adjusted for production volume. The system does not assign a sustainability rating. It produces a question: Why is energy intensity rising at this site? An operating partner with sector experience then investigates. Sometimes the answer is a faulty meter. Sometimes it is genuine operational drift. The human always makes the judgment.
This is the logic the strongest public disclosures already follow. EQT, the listed Swedish private equity group, ties the pricing on some of its financing to a short list of audited operational KPIs at portfolio companies, measured quarterly. Pick the few metrics that matter and watch them closely, the way a good investor watches cash. An AI system earns its place by flagging movement in those metrics rather than stacking a score on top of them.
Governance for AI ESG outputs
Private equity firms already interrogate AI-generated financial outputs. An AI-derived EBITDA adjustment in an investment committee memo gets challenged in the room. An AI-generated ESG score in an LP report often does not. The same discipline should apply to both. Here are four practices that can close that distance, drawn from how I advise firms:
- Match oversight to consequence. A deal-screening note for an analyst can run on light review. ESG metrics heading into an LP report should carry dual sign-off, the way a quarterly valuation does.
- Require the system to separate what is measured directly from what it inferred from incomplete data – and to say so on the page.
- Test the reviewers as much as the model. Inject known errors into ESG outputs in sandboxed review and track how many the team catches. Aviation and medical diagnostics have done this for decades.
- Budget for the oversight AI creates. If a sustainability team can suddenly monitor 200 portfolio companies instead of 60, the firm has taken on 140 new review obligations. Most firms have not staffed for them.
The fundraising edge
The friction starts small. During a re-up review, an LP’s operational due diligence team asks who checked the carbon figures. More than once the answer has been that an AI system produced them and no one signed off. The figure is usually fine. The missing reviewer is the problem, and it turns a routine re-up into weeks of follow-up. I have watched the same thin answer push an allocator to wait for the next fund.
LPs hold cards they did not hold 18 months ago. Fundraising is slower, capital has concentrated in fewer managers, and they are using that position to press harder on data quality. An LP choosing between two similar funds can make AI governance the tiebreaker. They should press three questions:
- Which ESG metrics in your last LP report were generated or modified by AI, and who reviewed them before it went out?
- What is the dual sign-off process for AI-generated outputs that reach limited partners?
- When an AI system flags an ESG anomaly, what is the average time to human resolution, and what share of flags are dismissed without investigation?
GPs that can answer cleanly will raise faster and on better terms. GPs who cannot will field sharper versions of these questions every cycle, until AI governance becomes a condition of the commitment rather than a differentiator. I expect that around the 2027 vintage.
So read your own last LP report the way an allocator now will. Find every ESG number a model generated or touched, and ask who checked it before it left the building. If you cannot answer that, you have found your first project.
Dr. Leigh Coney is founder and principal consultant at WorkWise Solutions.
Disclosure: WorkWise Solutions advises private equity firms on AI deployment and governance and builds purpose-built AI systems for the sector.
Guest posts on ImpactAlpha represent the opinions of their authors and do not necessarily reflect the views of ImpactAlpha.