AI agents in ESG data: why assessor sign-off still matters

The language of ESG software has shifted. Where providers once offered an AI assistant that answered questions, many now describe agents: software that maps data fields, flags issues, chases missing information, drafts narrative and moves a workflow on to its next step. Much of a property ESG programme is repetitive data handling, which is exactly what automation is for.

But asset and fund managers are not buying automation for its own sake. They are buying outputs that end up in front of regulators, lenders, investors and tenants: an ESOS notification, an EPC that shapes a letting decision, a carbon disclosure, a transition plan. Those outputs carry legal and reputational weight, and in most cases a named, qualified person has to stand behind them. The more an AI system does, the more important it becomes to be clear about where its work stops and a professional's judgement starts.

This article sets out the case for human-reviewed AI in ESG data. AI drafts, and a qualified assessor reviews and signs off. Numbers come from calculation engines, never from a language model. Every decision has an approval step and an audit log.

From copilots to agents: what has changed

A copilot waits to be asked. An agent takes a goal, breaks it into steps and works through them. In ESG data, that might mean reading a batch of utility invoices, matching each one to a meter, proposing an estimate for a missing month and drafting the paragraph that explains the result.

Across the market, agent designs fall into broadly three patterns. In the first, the agent proposes and a user reviews and approves each step, and the decisions are written to a log that can be downloaded and inspected. In the second, outputs are described as sourceable and traceable, so a reviewer can follow each figure back to its input. In the third, the agent is described as acting on a user's behalf with little or no human review, contacting people or changing records on its own.

For regulated UK property outputs, the third pattern is the one to be careful with. Autonomy is not the problem in itself. An agent that reconciles meter reads overnight and leaves a list of exceptions for a person to check in the morning is doing exactly what it should. The problem is autonomy at the point of decision: an output issued, a figure published or a third party contacted without anyone accountable having looked at it.

So the useful question to ask of any agent is not "how much can it do?" It is "at which points does it stop and wait for a person, and who is that person?"

Where the signature sits in UK property compliance

UK property compliance already answers the question of who signs. In each regime, the signature belongs to a person, not a system.

• ESOS. Under the Energy Savings Opportunity Scheme, an organisation's ESOS assessment must be signed off by an ESOS lead assessor, with limited exceptions set out in the Environment Agency's ESOS guidance. Lead assessors must be members of one of the approved professional body registers. A board-level director, or a person with equivalent management control, must also sign off both the assessment and the compliance notification. The phase 4 guidance sets out the detail.
• EPCs. A commercial EPC comes from a property assessment by a commercial energy assessor who belongs to an approved accreditation scheme.
• NABERS UK. A rating is carried out by a NABERS UK Licensed Assessor and certified through CIBSE Certification, the UK Scheme Administrator.
• Public environmental claims. The CMA's Green Claims Code says claims must be truthful and accurate, and must be substantiated. Whoever or whatever drafted the words, the business making the claim has to be able to back it up.
• Personal data. Where an AI workflow processes personal data, such as tenant contacts or occupancy information, the ICO's guidance on AI and data protection covers accountability, transparency, fairness and the safeguards around solely automated decision-making. The ICO notes that this guidance is under review following the Data (Use and Access) Act.

None of these regimes stops anyone using AI to prepare the work. What they make clear is that responsibility does not transfer to the software. An agent can help assemble the evidence an ESOS lead assessor reviews; it cannot be the lead assessor. That is the practical meaning of "AI drafts; a qualified assessor signs." AI is a production tool. The professional is accountable.

Three controls that make AI output signable

If a qualified person is going to put their name to an output that an AI system helped to produce, three controls make that reasonable rather than reckless.

1. Numbers come from calculation engines, never from a language model

Language models are good at reading, summarising and drafting. They are not a reliable way to add up consumption, apply emission factors or convert units, and they can produce a plausible figure that no calculation supports. In a well-designed workflow, every number (an energy total, an emissions figure, an intensity, a year-on-year change) is produced by a deterministic calculation engine from identified inputs and a stated factor set. The language model may describe the result; it never produces it. The test is simple: run the same inputs twice and you get the same answer, and you can show which factors and which version were used.

2. Every decision has an approval step

An agent might propose a field mapping, an estimate for a missing month, the classification of a sub-meter or a paragraph of narrative. Each of those is a decision. A decision that affects a reported figure should not take effect until a named person has approved it. Approval should be specific (this mapping, this estimate) rather than a blanket "accept all", and the reviewer should be able to see the evidence the proposal was based on. Where the output is regulated, the approval chain ends with the professional who signs it.

3. Every decision is written to an audit log

Boards, lenders and assessors tend to ask the same three questions. Where did this number come from? Who changed it? Can you reproduce it? An audit log answers them. It records what the agent proposed, what evidence it used, who approved or rejected it, when, and what changed as a result. The log should be readable by someone outside the software team, and it should be exportable, because the people who need it (an assessor, an auditor, a future software provider) may never be users of the system that wrote it.

These controls apply as much to a transition plan as to an ESOS file. Our piece on the transition plan data trail shows what they look like when every figure in a plan has to trace back to a meter, a certificate or a calculation. Ask the estate covers the related discipline of AI answers that cite their sources rather than paraphrasing them.

Questions to ask any ESG AI provider

These questions separate a controlled workflow from an impressive demo:

1. Which steps can the agent complete without a person approving them? Can you see that list, and can you change it?
2. Where do the numbers come from? If a figure appears in a draft, can the provider show the calculation, the inputs and the factor set behind it?
3. Who approved each output? Does the record show a named person, or only "the system"?
4. What happens when data is missing? Does the workflow flag the gap, estimate it with a clear label, or fill it silently?
5. Can you take the decision log and the underlying data with you? In a usable format, including if you change provider or your provider changes hands?
6. Where does personal data go? If the agent handles tenant or occupier information, where is the human review?
7. Will they show it on your data? A prepared demo proves the interface. Your own meters, invoices and certificates prove the controls.

Be wary of vocabulary that promises too much. No AI system is free of errors, and a claim that one is should prompt more questions, not fewer. Be equally cautious of any suggestion that an agent can stand in for the qualified assessor. In UK property compliance, it cannot.

What this article does not cover

This article is general guidance on the design of AI workflows for ESG data. It is not legal advice and does not determine compliance for any organisation. It does not set out the full requirements of the regimes mentioned; check the primary sources linked above. It does not claim that any AI system, including ours, replaces an ESOS lead assessor, a commercial energy assessor, a NABERS UK Licensed Assessor or any other qualified professional.