It takes an AI tool four seconds to produce a Scope 3 figure that looks exactly like a real one: clean, specific, confident, 1,284,000 tonnes of CO2 equivalent. It will take you, or your successor, eighteen months to unwind it once an assurer pulls the thread and finds nothing underneath. That asymmetry, four seconds to create and eighteen months to defend, is the most important thing to understand about disclosure. An unsupported number is not a head start. It is a liability you have placed on the company's books and signed.

The Defining Asymmetry of Disclosure

Every field has its own physics. In disclosure, the governing law is an asymmetry: the report is easy to generate and hard to defend. Generating a sustainability report has never been easier. AI can draft the narrative, populate the tables, and produce a number for any cell you point it at. Defending that report has never been harder, because assurance (independent external checking of your sustainability data, the audit applied to your emissions and impact figures) now sits on top of it. Roughly 73% of large global companies obtain external assurance on at least some sustainability disclosures, a figure to verify against its source but unmistakable in direction. Every number you publish is a number someone is paid to test.

Hold those two facts together and the danger becomes obvious. The cost of producing a figure has collapsed toward zero. The cost of defending a figure has not moved. So the gap between "I can generate this" and "I can defend this" is the widest it has ever been, and AI widens it further every quarter. A professional who does not feel that gap in their gut will eventually fall into it.

This is why the central discipline of the whole program is not about prompting or tooling. It is about respecting the asymmetry. The easy half, generating the number, is a trap precisely because it is easy. The hard half, supporting the number, is the entire job.

Compare it to a field where the asymmetry runs the other way to feel why disclosure is different. In a brainstorming session, generating ideas is the bottleneck and evaluating them is cheap, so a tool that floods you with options is pure gain; the bad ideas cost nothing because you simply discard them. Disclosure is the mirror image. Here generation is cheap and evaluation, in the form of assurance, is expensive and external and adversarial. A tool that floods you with plausible numbers is not pure gain, because every number you keep, you must support, and every number you keep that you cannot support is a liability you have to carry. The same capability that is a gift in a brainstorm is a hazard in a disclosure, and the difference is entirely about which side of the work is hard.

What an Unsupported Number Actually Is

Let us be precise about terms, because the precision is the point. A supported number is a figure that traces to evidence: a measured reading, a source document, a named and dated emission factor, a disclosed and methodical estimate. An unsupported number is a figure that does not, no matter how plausible it looks. The crucial insight is that plausibility and support are completely independent. A number can look perfect and be supported by nothing. AI is extraordinarily good at producing exactly that: the plausible-but-unsupported figure.

In disclosure, an unsupported number has a name, and it is not "estimate" and it is not "efficiency." It is a misstatement: a figure in a public, regulated disclosure that is not supported by the evidence it implies. A misstatement is not a productivity gain you booked early. It is a defect in the report, and a material misstatement is the thing assurance exists to catch.

The distinction between an estimate and a misstatement deserves a hard line, because the line is exactly where careers are made or lost. An estimate is an honest answer to a hard question: where measured data is genuinely unavailable, you build a figure using a stated method, you label it as an estimate, you disclose its uncertainty, and you say why primary data was not available. It is supported, because the support is the method and the disclosure of its limits. A misstatement is a dishonest answer dressed as an honest one: the same gap, filled with a number that is presented as if it were measured, with no method, no label, and no disclosure. The arithmetic underneath the two can be identical. What differs is whether the figure tells the truth about what it is. An estimate says "this is our best supported approximation, here is how we built it." A misstatement says "this is a fact," when it is not. The whole of defensible disclosure lives in that difference.

An unsupported number is not a head start on the work. It is a misstatement you have not been caught for yet.

This reframing is the whole lesson. The instinct, encouraged by every AI demo, is to treat a generated number as an asset: something you now have that you did not have before, that moves the report forward. In disclosure the opposite is true. A number you cannot trace to evidence is a liability you have added to the report, and it grows more expensive the longer it sits there undetected, because more decisions get built on top of it.

Walking the Chain: From Plausible Output to Regulatory File

Abstract risk does not change behaviour. Watching the failure unfold does. So walk the chain that an unsupported AI number travels, link by link, and notice that at every link it would have been cheaper to stop than to continue.

Link One: The Plausible AI Number

An analyst needs a Scope 3 Category 11 figure (use of sold products, often one of the largest categories) and the underlying data is thin. They ask an AI tool, which returns 1,284,000 tCO2e with a confident one-line rationale. It is specific, it is in range, it does not look invented. The analyst, under deadline, pastes it into the inventory. Cost to stop here: one honest sentence, "we do not have the data to support this yet." Cost not paid.

Link Two: The Public Disclosure

The figure rolls up into the total, the total goes into the ESRS climate disclosure, the disclosure is filed and published. The number is now public. It has been read by investors, competitors, NGOs, and regulators. It is also, importantly, now part of a continuous record: next year's figure will be compared to it, and any change will need a reason. Cost to stop here: a footnote and a restatement before publication, embarrassing but private. Cost not paid.

Link Three: The Assurer Pulls the Thread

Under the assurance engagement, the assurer samples Category 11. They ask the simplest possible question: "Show me the basis for 1,284,000." There is no basis. There is no measured activity data, no documented method, no source. The trail ends at "the AI produced it," which is not evidence and never was. The assurer cannot give the figure a clean opinion. This is the moment the four-second number starts costing real money, and it is entirely outside the analyst's control now.

Link Four: The Restatement

The company must restate: formally correct a previously published figure. A restatement is not a quiet edit. It is a public admission that a published number was wrong, it reopens the prior period, it forces the re-assurance of the corrected figures, and it raises the obvious question of what else in the report was unsupported. One unsupported number rarely stays one problem.

Link Five: The Greenwashing Headline

A restated emissions figure is a story. Greenwashing (presenting a misleadingly favourable environmental picture) is exactly the accusation a downward restatement invites: the company claimed a footprint it could not support. The headline does not say "AI hallucinated a factor." It says the company's name and the word greenwashing. Reputational cost now dwarfs the time the AI number ever saved.

Link Six: The Regulatory File

In the post-Omnibus regime, the companies still in CSRD scope are the largest undertakings (more than 1,000 employees and more than EUR 450 million turnover), where a failed disclosure is a board-level event. A misstatement in a regulated filing can become a regulatory matter. Now the four-second number has a file with a regulator's name on it, legal involvement, and executive attention. Every link in this chain was avoidable at link one, for the price of one honest sentence.

Trace the cost curve along that chain and the lesson is stark. At link one, the cost of stopping is a sentence. At link two, it is an embarrassing pre-publication correction. At link three, it is a failed sample and an awkward conversation. By link four it is a public restatement and re-assurance. By links five and six it is reputational damage and regulatory exposure that no one on the original team can control anymore. The cost of the unsupported number does not rise gently. It steps up at every link, and the point of decision, the only point where the cost is trivial, is the very first one, before the number ever enters the inventory. Everything after that is more expensive, and the expense compounds.

This is why seasoned disclosure professionals develop what looks, from outside, like an allergy to convenient numbers. It is not pessimism. It is an accurate read of the cost curve. They have learned that the comfortable feeling of a filled cell at link one is borrowed against a much larger bill at link four, five, or six, and that the only cheap moment to refuse the loan is right at the start. The discipline is not to be slow. It is to recognize that an unsupported number is borrowing, and that the interest rate is brutal.

Why AI Sharpens the Asymmetry

AI did not invent the unsupported number; analysts under deadline have always been tempted to "fill in" a figure. What AI changes is the supply and the disguise. The supply is now infinite: you can generate a plausible number for any cell, instantly, at no cost. And the disguise is now excellent: a model's output carries no tell. A human guess often sounds tentative. An AI figure sounds authoritative, because fluency is what these models optimize for. The very confidence that makes the output feel usable is the property that makes it dangerous.

So the asymmetry that always existed, easy to generate, hard to defend, gets pushed to its extreme. The generating side approaches free and frictionless. The defending side stays exactly as hard as it has always been, because an assurer's standard of evidence has not changed. A tool that makes the easy half effortless while leaving the hard half untouched does not close the gap. It widens it, and hands you more rope.

There is one more subtlety that catches good analysts. Before AI, the friction of producing a number was itself a kind of accidental safeguard. If you had to go and find a figure, you usually ended up with its source attached, because the effort of finding it left a trail. AI removes that friction, and with it the accidental trail. A number that appears instantly appears without provenance unless you deliberately attach it. So the loss of friction is not only a loss of a natural pause; it is the loss of the byproduct that used to make numbers traceable. The discipline must now do on purpose what friction once did by accident.

What the Asymmetry Asks of You in Practice

Understanding the asymmetry is not the same as living by it, so it is worth being concrete about the working posture it implies. First, treat every number that arrives without a source as incomplete, the way you would treat an invoice with no amount on it. It is not wrong yet, but it is not usable yet either, and it cannot enter the report until the source is attached. This single habit, refusing to let unsourced numbers move forward, stops most of the chain at link one without any heroics.

Second, separate the two halves of the work in your own mind and give the hard half the time it deserves. When AI saves you four hours on generation, the right response is not to declare the task done four hours early. It is to spend some of those four hours on support, on tracing, labelling, and documenting, because that is where the value and the protection actually live. A team that banks the AI time saving as pure speed and skips the support has not become more productive; it has become faster at producing liabilities.

Third, hold the line on language. Inside the team, never let an unsupported figure be called an estimate, because an estimate is a labelled, methodical, disclosed thing and an unsupported figure is not. Calling things by their right names, a misstatement a misstatement and an estimate an estimate, keeps everyone honest about what is in the file. The asymmetry is unforgiving, but it is also simple to respect once you accept that support, not speed, is the work.

Worked Example: The Same Cell, Two Ways

Watch one analyst handle one missing figure two different ways, and watch the liability appear or not appear.

The Liability Path

Category 1 purchased goods is missing a figure for a supplier representing 8% of spend. The analyst asks AI, gets 64,000 tCO2e, and enters it as if measured. The inventory now balances, the table looks complete, the deadline is met. On the surface this is a win: a hole filled, a report finished. Underneath, the analyst has booked a 64,000 tonne liability into a public, assured disclosure, indistinguishable from real data, waiting for the assurer to sample it. The "asset" is a misstatement with a timer on it.

The Asset Path

Same cell, same missing supplier. This time the analyst uses AI to build a transparent spend-based estimate: take the supplier's spend, apply a named, dated spend-based emission factor, and label the result clearly as a secondary estimate with its method and an uncertainty range. The figure entered is honest about what it is. In the file, it traces to its method and factor. When the assurer samples it, the answer is ready: "This is an estimate, here is the method, here is the factor, here is the uncertainty, and here is why primary data was unavailable." That figure is a genuine asset, because it can be defended. Same AI, same four seconds of generation, opposite outcome, and the only difference is whether the number was supported.

The contrast is the whole point. The liability path and the asset path both used AI, both filled the cell, both met the deadline. One produced a misstatement and one produced a defensible disclosure. Support is the entire difference between them.

Sit with how small that difference looks and how large its consequences are. The asset-path analyst typed a few more words of instruction and spent a few more minutes attaching a factor and a label. That is the whole delta in effort. Yet the liability-path figure is a 64,000 tonne misstatement waiting to detonate at the assurance engagement, and the asset-path figure is a clean, defensible estimate that the assurer can confirm and move past. A few minutes of support converted the same arithmetic from a liability into an asset. This is the most encouraging fact in the lesson: respecting the asymmetry is rarely expensive in the moment. It is a habit, not a heroic effort, and the cost of forming the habit is trivial compared to the cost of one trip down the chain.

Key Takeaways

• Disclosure runs on one defining asymmetry: the report is easy to generate and hard to defend, and AI widens that gap by making generation nearly free while defense stays exactly as hard.
• An unsupported number is a figure that does not trace to evidence, regardless of how plausible it looks. In a public, assured disclosure it has a name: a misstatement.
• A generated number is not an asset and not a head start. Until it is supported, it is a liability you have added to the report and signed.
• The chain is predictable: plausible AI number, public disclosure, assurer pulls the thread, restatement, greenwashing headline, regulatory file. Every link was avoidable at the first one.
• The cheapest moment to stop an unsupported number is before it enters the inventory, for the price of one honest sentence about what you do not have.
• AI sharpens the danger because its output is both infinite in supply and authoritative in tone; fluent confidence is exactly the property that makes an unsupported figure feel usable.
• A restatement is not a quiet edit. It reopens the period, forces re-assurance, invites a greenwashing accusation, and raises the question of what else was unsupported.
• The same AI and the same four seconds can produce a liability or an asset. The only difference is whether the number traces to evidence. Support is the entire job.