Two people in the building will decide whether your reporting AI program lives or dies, and neither of them is the CFO who approved the budget. The first is the senior analyst who has closed the GHG inventory by hand for six years and has just been told a model will "help." She hears one word: replacement. The second is the external assurer who will read every number under a limited-assurance engagement and who has been burned before by a beautiful report built on numbers no one could reconstruct. He hears one thing when you say AI: risk he now has to test. Win both and your program compounds. Lose either and the analyst quietly routes around the tool while the assurer widens the scope and raises the fee. This lesson is about the contract that earns both: AI assists, the human decides, the file proves it.

Two Audiences, One Contract

Most reporting AI programs are pitched to the wrong audience. They are sold upward, to the executive who wants the report cheaper and faster, and the pitch celebrates speed. But the two people whose trust actually determines whether the program produces a defensible disclosure sit at the working level and at the engagement table. The analyst is the person who has to use the tool and stand behind its output. The assurer is the person who has to test that output and sign a conclusion on it. If either withholds trust, the program fails in a specific, predictable way, and no amount of executive enthusiasm rescues it.

The analyst who does not trust the program does not sabotage it openly. She uses it for the trivial tasks, keeps her real work in the spreadsheet she controls, and treats the AI output as something to be redone rather than reviewed. You get the cost of the tool without the benefit. The assurer who does not trust the program does the opposite of ignoring it: he leans in. AI-generated content he cannot trace becomes an area of heightened risk, which means more testing, more sampling, more questions, a wider scope, and a larger fee, precisely the outcome the program was supposed to avoid. The uncomfortable truth is that a distrusted AI program is worse than no AI program, because it adds cost on both ends while adding no defensible speed.

The good news is that both audiences want the same thing, even though they describe it differently. The analyst wants to know her judgment still matters and that she will not be blamed for a machine's mistake. The assurer wants to know that every figure traces to evidence and that a named human decided it. Those are the same requirement seen from two sides. The contract that satisfies both is a single sentence the whole program can be built around, and it is the sentence this program returns to again and again: AI assists, the human decides, the file proves it.

A distrusted AI program is worse than no AI program: it adds cost on both ends and defensible speed on neither.

The Contract Decoded: Assist, Decide, Prove

The three clauses of the contract are not slogans; each is a design constraint that shapes how the workflow is built, and each answers a specific fear held by one or both audiences. Read them as engineering requirements rather than reassurances.

AI Assists: The Machine Is a Draft, Never a Decision

The first clause fixes the role of the model. AI produces drafts, candidates, extractions, and suggestions. It never produces a disclosed figure directly. A model can extract activity data from an invoice, propose an emission factor, cluster stakeholder inputs into materiality themes, or draft an ESRS narrative datapoint, but every one of those is an input to a human decision, not an output to the report. This is what protects the analyst: her judgment is not being automated away, it is being fed better inputs faster. The data wrangling that consumed her weeks is compressed, and the part that is actually her expertise, deciding whether a factor is right, whether a boundary is complete, whether a narrative is faithful, is exactly what remains hers. The clause reframes AI from a replacement for the analyst into a junior assistant who does the tedious first pass and whose work she checks.

The Human Decides: Accountability Never Transfers

The second clause fixes accountability. The disclosed number is chosen by a named person who can defend it. The model's suggestion is only a suggestion until a human accepts, edits, or rejects it, and that act of decision is what makes the number a disclosure the company stands behind. This is the cardinal rule of the entire program stated at the level of a single figure: "the model recommended it" is never a defense to an assurer or a regulator, because accountability for a public disclosure cannot be delegated to a tool. This clause protects the analyst a second way, because it means she can never be blamed for a machine's error that she was not the one to accept, and it protects the assurer, because it guarantees there is always a human to interview about any number he samples.

The File Proves It: Trust Is a Record, Not a Promise

The third clause is the one that turns the first two from a policy into something an assurer can rely on. It is not enough that AI assisted and a human decided; the file has to prove it happened. For every figure, the record shows what the AI produced, what the human did with it, who that human was, and what evidence the final number traces to. The provenance of an AI-assisted number is captured at the moment it is created, not reconstructed under pressure when the assurer asks. This is the clause the assurer cares about most, because assurance runs on evidence, not assertion. A program that can say "trust us, a human checked it" earns nothing. A program that can show the assurer the extraction, the human edit, the reviewer's name, and the source document behind any sampled figure earns the thing that lowers his assessed risk. Trust, to an assurer, is a synonym for a reconstructable file.

The Two Fears, One Answer

It is worth laying the two audiences side by side, because seeing them together reveals that the contract is not two separate charm offensives but a single design that answers both fears at once. The analyst fears being reduced to an unaccountable button-pusher and then blamed for a machine's error. The assurer fears an invented number sailing into a public disclosure through a process he cannot see. Notice that the same three clauses defuse both. AI-assists means the analyst keeps the judgment and the assurer knows no figure was authored by a machine alone. Human-decides means the analyst is the accountable expert and the assurer always has a named person to interview. File-proves-it means the analyst can demonstrate she decided on a constrained draft and the assurer can reconstruct the number on demand. One contract, two fears, the same answer, which is exactly why building for one audience tends to serve the other rather than trading against it.

Winning the Analyst: From Fear of Replacement to Ownership

The analyst's fear is rational and should be met directly rather than waved away with "AI will not replace you." That reassurance is empty because it asks her to take a promise on faith. What actually earns her trust is a set of concrete moves that change her lived experience of the work, so she feels the upgrade rather than being told about it.

The first move is to give the tedious work to the machine and the judgment work to her, visibly. Show her that AI now does the invoice extraction, the first-pass factor lookup, the questionnaire triage, the narrative first draft, and that what is left on her desk is the review, the method choice, the boundary call, the sign-off. Her day shifts from data entry toward decision, which is both higher status and harder to automate. The second move is to make her the reviewer and the accountable owner, formally. She is the one who accepts or rejects the model's output, her name is on the sign-off, and the workflow is explicit that the AI cannot commit a number without her. This converts the source of her fear, the model's involvement, into a source of her authority, because the model works for her review.

The third move is to protect her from the machine's mistakes rather than exposing her to them. A badly designed program makes the analyst the last line of defense against a hallucination she had no way to catch, then blames her when it slips through. A program that earns trust builds the guardrails, grounding on the approved factor library so the model cannot invent a factor, provenance capture so nothing is unsourced, prior-period reconciliation so an implausible jump is flagged, and positions her as the informed reviewer of an already-constrained draft rather than the sole barrier against an unconstrained one. The fourth move is to invest in her skill, because the analyst who is trained to operate AI well becomes more valuable, not less, and a fast-growing green labor market rewards exactly that literacy. When the analyst can see that the program makes her the accountable expert who commands the machine, the fear of replacement is replaced by ownership, and an owner uses the tool properly instead of routing around it.

Winning the Assurer: From Suspicion to Lower Assessed Risk

The assurer's suspicion is also rational, and it is best understood in his own terms. Under a limited-assurance engagement he forms a conclusion based on procedures that are less extensive than a reasonable-assurance audit, and he does that by assessing where the risk of material misstatement is highest and concentrating his testing there. AI-generated content, if opaque, is a bright red flag on that risk map: content produced by a process he cannot see, that may have invented a number, is exactly the kind of thing that drives assessed risk up and testing up with it. The entire job of earning the assurer's trust is to move AI from the high-risk column to the well-controlled column on his assessment, because that is what reduces his procedures and his fee rather than inflating them.

The move that accomplishes this is transparency plus control, evidenced. The first concrete step is to disclose the use of AI to the assurer early, in the planning walkthrough, rather than letting him discover it. An assurer who learns mid-engagement that AI drafted the disclosures treats the whole file with fresh suspicion; an assurer who was told up front, and shown the controls, can plan around it. The second step is to show him the controls that constrain the AI: the grounding on approved factors, the human sign-off gate, the provenance capture, the labeling of primary versus secondary and of estimated versus measured. Controls are what convert a scary black box into a governed process he can rely on. The third step is to hand him a file that reconstructs any AI-assisted number on demand, from the model's draft through the human decision to the source evidence, so that when he samples a figure the answer is immediate and complete.

There is a fourth step that separates a program that merely survives assurance from one the assurer actively trusts: never launder an estimate as measured data, and never let AI do it silently. The single fastest way to destroy an assurer's trust permanently is for him to discover that a number presented as measured activity data was in fact an AI-generated industry average. Once he finds one, he assumes there are more, and every figure in the report inherits the doubt. A program that labels every estimate as an estimate, with its method and uncertainty, and reserves measured status for genuinely primary data, tells the assurer that the file is honest about its own weaknesses, which is the deepest form of credibility. Defensible estimation, disclosed, builds trust; fabrication laundered as measurement destroys it, and there is no faster path from suspicion to a widened scope.

A Worked Example: The Same Figure, Two Ways

Consider a single Scope 3 figure, purchased goods and services emissions from a mid-sized supplier, produced two ways, and watch what each does to the two audiences.

In the trust-destroying version, an analyst under deadline pastes the supplier's spend into an open chatbot and asks it to estimate the emissions. The model returns a confident number using an industry-average factor it names vaguely as "a standard dataset." The analyst, rushed, drops it into the reporting platform in the same field she uses for supplier-reported figures, with no note. The number looks clean and precise. Now trace its path through the two audiences. The analyst does not feel ownership; she feels she has been turned into a copy-paste conduit for a machine, and she trusts neither the number nor the process. Months later the assurer samples that supplier, asks for the basis, and discovers that a figure sitting in the measured-data field is an unsourced AI estimate. He does not just flag that number. He now assumes the same shortcut was taken elsewhere, expands his sample, raises the assessed risk for the whole Scope 3 category, and the engagement that was scoped as limited assurance turns into a scramble, a widened scope, and a hard conversation about restatement. The program that promised speed delivered a finding.

In the trust-building version, the same analyst uses the grounded workflow. The AI extracts the supplier's spend from the invoice and proposes a spend-based estimate, but it is required to return the named, dated factor source from the approved library and to label the output as a secondary, spend-based estimate. The analyst reviews it, confirms the method is appropriate for a supplier with no primary data, edits the uncertainty note, and signs off. The record captures the AI draft, her edit, her name, the factor source, and the estimate label, all at the moment of creation. Her experience is that of an expert who used a fast assistant and made the call, so she owns the number. When the assurer samples the same supplier, he receives, on demand, the whole chain: draft, decision, reviewer, source, and an honest label that says estimate rather than measurement. He tests it, finds it reconstructable and honestly labeled, and files it in the well-controlled column. The identical figure, produced under the contract, lowered his assessed risk instead of raising it. Same number, opposite outcome, and the only difference was the contract: AI assisted, the human decided, the file proved it.

Building the Contract Into the Culture, Not Just the Policy

A contract written into a policy document earns nothing; it earns trust only when it is visible in how the team actually works and in how the program is introduced. The strategist's job is to make the three clauses the default behavior, and there are a few moves that do this reliably. Introduce AI to the analysts as a tool that makes them the accountable expert, not as an efficiency that makes them cheaper, and back that framing with the concrete role of reviewer and signer rather than leaving it as rhetoric. Introduce AI to the assurer before he finds it, as a governed process with controls you will show him, rather than as a surprise he uncovers. Make the provenance capture and the sign-off non-optional parts of the workflow, so that doing it the trustworthy way is the only way, not the virtuous exception. And measure the program on assurance findings and analyst adoption, not only on cycle time, so that the metrics reward trust rather than punishing it.

There is a sequencing lesson here that separates programs that earn trust from programs that assume it. Trust is not declared at a kickoff meeting; it is accumulated through small, repeated demonstrations that the contract holds under pressure. The first time an analyst catches a hallucinated factor because the grounding flagged it, she trusts the guardrail a little more. The first time an assurer asks for the basis of a sampled figure and receives the full chain in seconds rather than a week of scrambling, he moves that part of the file toward the well-controlled column. Every clean reconstruction is a deposit, and every laundered estimate or missing sign-off is a withdrawal that costs far more than a deposit earns. The strategist's discipline is to protect the balance: make the trustworthy path the only path, so that the deposits accumulate automatically and the withdrawals become structurally hard to make. A program that treats trust as a standing balance to be defended, rather than a speech to be given, is the one that still has both audiences a year later.

The deepest point is that the analyst and the assurer are not obstacles to be managed around; they are the two people whose trust is the actual product of a reporting AI program. A program that earns both has captured the rare win the whole discipline is built on: faster and more defensible at once, because the same discipline that makes an AI output traceable makes it both trustworthy to the person who signs it and reliable to the person who tests it. The contract is not a compliance overhead layered on top of the speed. It is the thing that makes the speed safe to keep.

Key Takeaways

• Two people decide whether reporting AI succeeds: the analyst who must use and stand behind the output, and the assurer who must test it. Win both or the program adds cost on both ends while adding no defensible speed.
• The contract that earns both is one sentence: AI assists, the human decides, the file proves it. Each clause is a design constraint, not a slogan, and each answers a specific fear.
• A distrusted AI program is worse than no program. The analyst routes around it and keeps her real work in her own spreadsheet; the assurer treats opaque AI content as heightened risk and widens the scope and the fee.
• Win the analyst with concrete moves: give the tedious work to the machine and judgment to her, make her the accountable reviewer and signer, protect her from the machine's mistakes with guardrails, and invest in her AI literacy so she becomes more valuable, not less.
• Win the assurer by moving AI from his high-risk column to his well-controlled column: disclose AI use early in the planning walkthrough, show him the controls, and hand him a file that reconstructs any AI-assisted number on demand.
• The fastest way to destroy an assurer's trust is to let an estimate be laundered as measured data. Once he finds one, he doubts them all. Label every estimate with its method and uncertainty; reserve measured status for genuine primary data.
• Provenance must be captured at the moment a number is created, not reconstructed under pressure. To an assurer, trust is a synonym for a reconstructable file, not a promise that a human checked it.
• The analyst and the assurer are not obstacles to manage around; their trust is the actual product of the program. Earning both is what makes the speed safe to keep.