You have done the hard part. The materiality workshop is over, the matrix is signed, and twelve topics made the cut. Now comes the writing: for every one of those twelve, you owe a rationale that explains why it is material, on impact, on finance, or on both, and that an external assurer can test line by line. That is roughly forty paragraphs of careful, evidence-anchored prose, due in ten days, and you have a GHG inventory to close in the same window. So you ask an AI to draft the rationales. In ninety seconds it returns twelve clean, confident, professional paragraphs. They read beautifully. And exactly one of them contains a sentence about a stakeholder concern that no stakeholder ever raised, asserted with the same calm authority as the eleven true ones. Finding that one sentence, before the assurer does, is the entire job of this lesson.

What a Materiality Rationale Has to Do

A materiality rationale is not a paragraph that sounds reasonable. It is the written explanation of why a topic crossed your threshold, and it has to carry a documented basis: the evidence behind the claim, the threshold you applied, and the record of who decided. Under the Corporate Sustainability Reporting Directive (CSRD), which survived the 2025 to 2026 Omnibus simplification and still binds the largest undertakings (more than 1,000 employees and more than EUR 450M turnover), this rationale is part of the disclosure, and an external assurer reads it. Roughly 73% of large global companies now obtain external assurance on at least some sustainability disclosures, up from 51% in 2019. So the rationale is an audited text. Every assertion in it is a claim the assurer can ask you to support, and "the AI wrote it" supports nothing.

Hold the two lenses precisely, because a complete rationale addresses both. Double materiality is the ESRS rule that a topic is material if it is significant from either direction, and you must assess both. Impact materiality is how your company affects people and the environment. Financial materiality is how a sustainability matter affects your company's own financial position, cash flows, access to finance, or cost of capital. The ESRS organise the underlying findings as IROs: impacts, risks, and opportunities, where impacts mostly drive impact materiality and risks and opportunities mostly drive financial materiality. Your rationale for a topic like climate change typically has to speak to both axes: the impact your emissions have on the climate, and the financial risk a carbon price or a physical disruption poses to you. A rationale that argues only one axis when both apply is an incomplete rationale, and an assurer will notice the silence.

The word that does the heavy lifting is basis. The basis-of-preparation is the documented foundation under a disclosure: for a materiality rationale, it is the set of evidence each assertion rests on, the materiality threshold you applied (the documented line, set in advance from severity and likelihood, that separates material from not material), and the governance record of who made the call. An assurer reading your rationale is really reading for the basis behind it. They are not grading prose. They are checking whether each sentence traces to something real.

What AI Is Genuinely Good At, and Where It Turns on You

Drafting prose is one of the things generative AI does best, and a materiality rationale is, on its surface, a prose task. Given your topic, your axis, your evidence, and your threshold logic, a model will assemble a clear, well-structured paragraph far faster than you can type one, and it will do it twelve times without tiring or drifting in tone. For the disclosure professional staring at forty paragraphs and a closing inventory, that speed is not a luxury. It is the difference between a calm review and a midnight scramble. Used well, AI turns the writing from the bottleneck into the easy part, freeing your hours for verification, which is where the risk actually lives.

But the same machine that drafts the true paragraph will, with identical fluency, draft a false one. This is the failure mode that ends careers in disclosure, and it wears three faces here. The first is the invented stakeholder concern: the model writes that "employees and local communities have raised concerns about water use," because that is the kind of sentence that belongs in a water rationale, even though your inputs show only a regulator and two customers raised it. The second is the invented target or commitment: asked to make the rationale sound robust, the model adds "the company has committed to a 30% reduction by 2030," a target your company never set. The third is the softened impact: the model, trained to be agreeable, downgrades "significant adverse impact on a water-stressed community" into "potential effects on local water resources," quietly editing a negative impact into something blander than the evidence supports. Each is fluent. Each is wrong. None announces itself.

AI drafts the prose; the analyst owns the judgment and the basis-of-preparation. A fluent sentence with no evidence under it is not a rationale. It is a misstatement waiting for an assurer to find it.

Draft From the Evidence, Not From the Topic

The single move that tames all three failure modes is to change what you feed the model. If you prompt "write a materiality rationale for water," the model writes from its training data: a generic, plausible water paragraph studded with the kind of claims water rationales usually contain, several of which will not be true for you. If instead you prompt "draft a rationale for water using only these inputs," and you paste your actual evidence, your stakeholder list, your severity-and-likelihood scoring, your threshold, and your risk-register entry, then the model has to build the paragraph from your facts. It can still drift, so you still verify, but you have closed the widest door through which invention enters: the empty prompt that invites the model to fill the gap with plausible fiction. Ground the draft in the evidence and the model becomes a writer working from your file rather than an author inventing one.

You can tighten the grounding further with three instructions that cost nothing and prevent a great deal. First, tell the model to use only the supplied evidence and to add nothing from general knowledge, so it cannot reach for the generic claim. Second, tell it to leave any gap explicit rather than fill it: if your inputs do not establish a financial effect, you want the draft to say so, not to invent one. A model that writes "no financial risk was identified in the supplied inputs" has handed you something honest and actionable; a model that writes "moderate financial considerations may apply" has handed you a plausible sentence with nothing behind it. Third, ask the model to flag, separately from the prose, any assertion it is unsure it can support from the inputs. None of this makes the output trustworthy on its own, but it shifts the model from confident invention toward visible uncertainty, and visible uncertainty is far easier to verify than buried fiction.

It is worth naming why the empty prompt is so seductive, because understanding the pull helps you resist it. Under deadline, pasting a topic name and getting back a complete paragraph feels like enormous leverage, while assembling the evidence packet to paste in feels like the slow part you are trying to avoid. But the evidence packet is not overhead; it is the rationale. The prose is the cheap part, the part the model can do. The evidence, the threshold, and the decision are the expensive part, the part that makes the paragraph a disclosure rather than an essay, and they have to exist whether or not you use AI to write the words around them. So the grounded prompt is not extra work. It is the work, arranged so the model does the typing and you do the thinking.

Verify Every Assertion Before It Ships

Grounding reduces invention; it does not eliminate it. So the load-bearing discipline is verification, and it is concrete. Read the draft assertion by assertion and, for each one, ask a single question: where is the evidence for this, and does it match. A stakeholder claim must match your actual input list. A number, a target, or a commitment must match a source document you can name. A characterisation of an impact ("significant," "adverse," "high severity") must match your severity scoring, not be softer or harder than the evidence. A statement that the topic clears the threshold must match the threshold you actually applied and recorded. Any assertion that fails this test gets struck or rewritten, not negotiated. The standard is simple and unforgiving: no sentence ships unless the file can support it.

Two assertions deserve special suspicion because they are the ones the model most loves to invent and the assurer most loves to test. The first is any quantified target or commitment, because a fabricated target is among the gravest disclosure errors, a public promise the company never made. Trace every target to the board minute, the strategy document, or the public commitment that established it, and if you cannot find one, the target does not exist and the sentence comes out. The second is any claim about who raised a concern, because that is the link back to your stakeholder evidence and the easiest place for the model to launder a plausible guess into a stated fact. Match every "stakeholders have raised" to the actual inputs, by source, and rewrite anything you cannot match.

A useful habit is to read the draft twice with different eyes. On the first pass, read it as a writer would, checking that it is clear, complete on both axes, and well structured. On the second pass, read it as an assurer would, ignoring the prose entirely and treating each sentence as a claim to be supported: underline every assertion of fact, every number, every named stakeholder, every characterisation of severity, and for each one ask where in the file it is supported. The two passes catch different things. The writer's pass catches a missing axis or a muddled sentence; the assurer's pass catches the invented target sitting in an otherwise flawless paragraph. It is the second pass that protects you, and it is the one the deadline tempts you to skip, which is exactly why you build it into the workflow as a required step rather than a good intention.

One more discipline closes the loop: capture the verification itself, not just the corrected prose. As you confirm each assertion, note where its support lives, the input ID, the register entry, the workshop minute. That note is not busywork; it is the basis-of-preparation, assembled as a by-product of verification rather than reconstructed painfully later. When the assurer reads a sentence and asks "show me," you do not re-derive the support from memory; you open the note and point. The team that captures verification as it goes walks into the engagement with the file already built. The team that verifies in its head and keeps only the clean prose walks in with a beautiful document and no basis, which is the same position as having done no verification at all, because an unrecorded check is one you cannot prove you made.

Before and After: A Drafted Rationale, Turned Defensible

Watch one rationale move from a fluent draft to an assurable text. The setting is the same mid-cap food manufacturer, writing the rationale for water after the workshop concluded water is material on both axes.

Before (the raw AI draft, from a topic-only prompt): "Water is a material topic for the company. Employees, local communities, and investors have consistently raised concerns about the company's water consumption across its operations. The company recognises water as a critical shared resource and has committed to reducing its water intensity by 30% by 2030. Climate-related water stress poses potential effects on local resources and may present moderate financial considerations. Water is therefore disclosed under ESRS E3." It is fluent, confident, and board-ready. It is also riddled. The stakeholder list is invented (your inputs show a regulator and two customers, not employees and investors). The 30% target was never set. "Significant adverse impact on a water-stressed community" has been softened to "potential effects on local resources." And "moderate financial considerations" papers over a specific, live abstraction-permit risk.

After (verified against the file, drafted from evidence): "Water is material on both impact and financial axes. On impact, the company withdraws water at three plants located in basins classified as high water stress, with a significant adverse impact on a vulnerable downstream community; this was raised by a regulator and two large customers and scored high on severity (irreversible depletion, vulnerable population) and certain on likelihood (withdrawal is ongoing), clearing the impact threshold. On finance, two of those three sites face tightening abstraction permits (risk-register entry RR-2026-014), which could force capital spend or curtail production, a financial effect assessed as material in magnitude and probable in likelihood. The materiality conclusion was confirmed at the 14 February materiality workshop. Water is disclosed under ESRS E3." Every assertion now traces: the stakeholders to the input list, the impact characterisation to the severity scoring, the financial risk to a named register entry, the decision to a minuted workshop. There is no target, because no target was set, and inventing one to sound robust would have been the worst error on the page.

The AI still wrote the first version in seconds and saved you the typing. What it could not do was check its own stakeholder list against your inputs, refuse to invent a target, hold the impact at the severity your evidence supports, or anchor the financial claim to a register entry. That was you. The model drafted the prose. You owned the judgment and the basis, and that is the only division of labour that survives an assurance engagement.

How to Draft Rationales Without Getting Burned

A few rules make AI a safe accelerator here. Draft from the evidence, never from the topic: paste your inputs, your scoring, your threshold, and your register entries, and instruct the model to use only those. Verify every assertion against the file, treating any unsupported sentence as a defect to be struck or rewritten, not a draft to be polished. Hunt targets and stakeholder claims first, because those are the model's favourite inventions and the assurer's favourite tests. Hold each impact at the severity your evidence supports, watching for the agreeable model that softens a negative impact into something blander. Make sure every rationale speaks to both axes where both apply, and document the threshold and the decision alongside the prose, so the basis travels with the text. And keep the link from each rationale back to its evidence, so when the assurer reads a sentence and asks "show me," you can.

Do that, and AI gives you the thing it is genuinely good for: forty paragraphs drafted in minutes, your hours redirected to verification, and a set of rationales where every assertion traces to evidence. You walk into the assurance meeting not defending prose you cannot explain, but handing over a file where the writing is fast and the basis is solid. The model drafted the words. You owned the judgment. That is the rationale that holds.

Key Takeaways

• A materiality rationale is not reasonable-sounding prose; it is the written explanation of why a topic crossed your threshold, carrying a documented basis of evidence, threshold, and who decided, that an assurer tests line by line.
• A complete rationale addresses both lenses where both apply: impact materiality (your effect on people and planet) and financial materiality (the effect on your finances), framed as IROs against a documented threshold.
• AI drafts the prose fast and well, freeing your hours for verification; that is its genuine value, and verification is where the disclosure risk actually lives.
• The same fluency that drafts a true paragraph drafts a false one: watch for the invented stakeholder concern, the invented target or commitment, and the softened negative impact, each asserted with calm authority.
• Draft from the evidence, not from the topic: a topic-only prompt invites the model to fill gaps with plausible fiction, while pasting your actual inputs and scoring forces the paragraph from your facts.
• Verify every assertion against the file, asking where the evidence is and whether it matches; no sentence ships unless the file can support it, and unsupported sentences are struck, not polished.
• Give targets and stakeholder claims special suspicion: trace every target to the minute or commitment that set it, and match every "stakeholders raised" to the actual inputs by source.
• The model drafts the words and the analyst owns the judgment and the basis-of-preparation; used well, AI makes the writing fast and the file solid at the same time.