Skip to content
6 min read

AI Ethics in ESG Projects: How AI Supports Your ESG Goals

Container ship beneath a port crane

AI already touches most ESG workstreams: it reads meter data for a carbon footprint, screens job applicants for bias, and flags irregular transactions long before an auditor would. Used well, it moves ESG goals from a reporting exercise into daily practice. Used carelessly, it introduces exactly the risks ESG programmes exist to manage: biased outcomes, opaque decisions, and data handled without proper safeguards. The EU AI Act, in force since August 2024 and applying in stages (prohibited practices and AI-literacy duties since February 2025, obligations for general-purpose AI models since August 2025), sets the legal floor. This article works through where AI genuinely supports ESG goals, where it works against them, and what needs to be in place before AI touches data that ends up in an ESG or CSRD report.

How AI supports ESG goals: use cases per pillar

AI earns its place in an ESG programme when it is tied to one pillar and one measurable target, not added as a general efficiency layer. The table below maps three common use cases to the goal they support.

ESG pillar AI use case Target it supports
Environmental Predictive control of heating, cooling or production load Energy-efficiency or scope 1/2 reduction target
Social Bias screening in recruitment, accessibility tools (speech-to-text, screen readers) Diversity or accessibility target in the workforce section of the report
Governance Anomaly detection in transactions, automated compliance monitoring Fraud-loss ratio or compliance-incident count

DeepMind reported that its AI cut the energy used to cool Google's data centres by up to 40%, the kind of result a facility can track against its own electricity or scope 2 target. On the social side, accessibility applications such as speech-to-text and screen readers, together with bias-screening tools in recruitment, follow the same logic: a named gap, accessibility or representation, matched to a named control. On governance, banks and payment providers run anomaly detection to flag transactions that a person then reviews, feeding straight into the incident figures a governance section reports. In each case, a person still verifies the result before it is reported.

Where AI works against ESG goals: bias, opacity, privacy and energy

Each use case above carries a matching risk, and transparency, fairness and data protection, the principles that used to open this article as a separate section, are really just the lens for spotting it early. Algorithmic bias is the best documented of the three: facial-recognition training sets have been shown to be over 75% male and over 80% white, pushing error rates for dark-skinned women above 20%, in one study above 34% (Gender Shades, 2018). A hiring model trained on years of resumes can learn the same imbalance.

AI use case Main ethical risk Control
Predictive energy or production control Errors feed unnoticed into reported emissions figures Threshold alerts, periodic human review of deviations
Recruitment and accessibility screening Bias reproduces past discrimination Bias testing (AIF360, Fairlearn), diverse and refreshed training data
Transaction and compliance monitoring Decisions are not explainable, breaking the audit trail Documented logic and data sources, human sign-off before action

Data protection adds a fourth layer: AI trained on employee, customer or partner data can retain or infer more than the original ESG purpose needs, which is a direct GDPR exposure. And the energy problem is structural, not incidental: running the AI that cuts a facility's energy use itself consumes energy, so the saving only counts once the AI's own compute is netted against it, not just the process it optimises.

AI in ESG reporting under CSRD and ESRS

Since the Omnibus I reform, full CSRD reporting duties apply only to companies with more than 1,000 employees and more than €450 million in net revenue; companies newly in scope report from financial year 2027 onward. Below that threshold, the Voluntary Standard (VS), also known as VSME, covers similar ground in a lighter form, including the basic module's energy and emissions data points: total energy consumption in MWh, and scope 1 and scope 2 (location-based) emissions. AI is increasingly the tool that pulls these figures out of meter readings, invoices and ERP exports, doing the unit conversion and reconciliation work a person used to do by hand.

Whether AI use itself counts as a material topic is a double-materiality question, not a yes-or-no rule. It is an impact topic if the AI has a real effect on people, for example job applicants or employees it rates, and a financial topic if AI-related compliance costs or incidents are significant enough to matter to a reader of the figures. A company using AI only for internal checks on its own scope 1 and 2 data is unlikely to clear that bar; one using AI to score ESG-linked supplier risk is closer to it.

In practice, that assessment is worth documenting once and reusing: list where AI touches reported sustainability data, assess impact and financial materiality for each use, and carry only the material ones into the materiality analysis.

Governing AI that touches ESG data: EU AI Act and controls

The EU AI Act has been in force since August 2024 and applies in stages: prohibited practices and AI-literacy duties since February 2025, obligations for general-purpose AI models since August 2025. ESG and CSRD teams using AI, whether for scoring, classification or drafting, need to check which of their tools fall into a regulated category and keep the paperwork that proves it, rather than assume a tool is exempt because it only touches internal data.

Before AI touches data that ends up in a report, four controls should already be in place:

  • A named owner for each AI use that feeds into reported data.
  • Documented data sources and model logic, kept current when either changes.
  • A bias or accuracy test before go-live, repeated on a schedule, not only once.
  • A human sign-off step before AI output enters a report or a decision.

A workable sequence for a first AI use case: map which reported figures it touches, run the bias or accuracy test, document the result, assign the sign-off, then repeat the test at the next reporting cycle. The NIST AI RMF and ISO 42001 formalise this and are a reasonable structure to borrow from. For a fuller walk-through of risk classes and deployer duties under the Act, see this site's piece on the EU AI Act for climate tech and ESG tools.

FAQs

1. What are practical examples of AI used to reach ESG goals?

Predictive energy management for heating or cooling, bias screening and accessibility tools in recruitment, and anomaly detection in financial transactions are the three most common uses, one per ESG pillar. Each works best tied to a single measurable target rather than used as a general efficiency layer.

2. Is AI use a material topic under double materiality?

Not automatically. It is an impact topic if the AI has a real effect on people, such as job applicants or rated employees, and a financial topic if AI-related compliance costs or incidents are significant. Document the assessment per use case rather than deciding once for the whole company.

3. Which ESG tasks can AI safely automate, and where does a human need to review?

AI can safely pull and reconcile structured data, such as energy or emissions figures from meters and invoices. Anything that scores or judges people, or that feeds a number into a published report, needs a named human sign-off before it goes out.

Johannes Fiegenbaum

Johannes Fiegenbaum

ESG and sustainability consultant based in Hamburg, specialised in VSME reporting, carbon accounting and climate risk analysis. Works with mid-sized companies, investors and financial institutions, and developed the VSME reporting software VSEasy together with clients.

More about