Watch a reporting function that bought AI but never redesigned. The carbon accountant still owns "the inventory" end to end, so AI just makes her personal spreadsheet faster and less traceable. The disclosure lead still owns "the narrative," so AI drafts prose that never touches the evidence file. Each silo pulls AI into its own private way of working, and the seams between silos, always where errors hid, now hide errors at machine speed. The team is faster and less assurable at the same time. The problem is not the tool. The problem is that the operating model was built for humans doing everything by hand, and nobody rebuilt it to fit the tool.

The Operating Model Must Fit the Tool

There is an old management truth: when you automate a task without redesigning the process around it, you get an expensive version of the old process, not a better one. In sustainability disclosure the stakes are sharper, because the old process was optimized for a world where every step was slow and human, and the assurance file was assembled at the end by whoever remembered what they did. AI breaks that world in two ways at once. It makes the throughput steps fast, and it makes the provenance steps fragile, because a fluent machine will happily produce a plausible number with no source unless the workflow forces it to trace one.

A task-silo organization cannot absorb this. In a silo model, work is divided by who owns which deliverable: the inventory, the materiality matrix, the CBAM declaration, the narrative. Each owner drops AI into their silo in their own way. There is no shared definition of when a number is "done," no shared provenance standard, no shared point where a human signs before the artifact advances. AI amplifies whatever the silo already does, including its bad habits. The redesign replaces the silo model with a workflow model: a single, explicit sequence of stages that every disclosure artifact passes through, where each stage is marked as AI-assisted or human-only, provenance is captured as work happens rather than reconstructed later, and sign-off gates sit at the points where a wrong number would become expensive.

The 2026 stakes make this more than a productivity question. The companies still in CSRD scope after the Omnibus, Directive (EU) 2026/470, are the largest undertakings, more than 1,000 employees and more than EUR 450M in turnover, where a failed disclosure is a board-level event. Roughly 73% of large global companies now obtain external assurance on at least some sustainability disclosures, up from 51% in 2019, so the report that a redesigned function produces is not just published, it is audited. And CBAM's definitive phase is live, ISSB is converging across dozens of jurisdictions, and the same fact base feeds all of them. A silo model that produces four private processes for four frameworks multiplies the seams, and each seam is a place an unsupported number can hide. The workflow model, by contrast, lets one traceable fact base flow into every framework's disclosure with the same provenance discipline, which is the only way to stay defensible across all of them at once without quadrupling the headcount.

The deeper reason silos fail under AI is that they were never really designed at all. They accreted. A person joined, took over a deliverable, and built a personal process around their own habits and tools. That was tolerable when everything moved at human speed and errors surfaced slowly enough to catch by eye. AI removes the slowness that used to be an accidental safety net. An unsupported factor that a human would have taken an afternoon to enter, and might have paused over, now enters in seconds, looks identical to a sourced one, and flows onward before anyone looks. The redesign is the deliberate act the function never performed: sitting down and defining, once, for everyone, how a disclosure artifact is made and how it is proven, so the speed AI provides runs on rails instead of running loose.

Automating a broken process gives you a faster broken process. In disclosure, a faster broken process is a faster route to a restatement.

Humans on Judgment, AI on Throughput

The organizing principle of the redesign is a clean division of labor: AI does throughput, humans do judgment. Throughput is the high-volume, low-ambiguity work where speed is the value and the answer is checkable against a source: extracting activity data from invoices and bills, looking up candidate emission factors, drafting the first version of a narrative datapoint, parsing supplier responses into structured fields, clustering stakeholder inputs into themes. Judgment is the low-volume, high-consequence work where being right matters more than being fast and the answer cannot be checked against a single source: deciding the reporting boundary, choosing whether a Scope 3 category is material, judging whether an estimate is honest or laundered, deciding whether a narrative faithfully represents a negative impact, signing that a figure is ready to be assured.

The mistake most functions make is letting AI drift across the line. AI drafts the materiality conclusion, not just the input clustering. AI selects the boundary, not just proposes options. AI writes the number into the disclosure, not just suggests a factor to verify. Every time throughput drifts into judgment, the function trades a defensible artifact for a fast one, and the drift is invisible until the assurer pulls the thread. The redesign makes the line explicit and enforced: each stage of the workflow is labeled, and the human-only stages cannot be skipped, because they are the stages where accountability lives. This is not a limitation on AI. It is what lets you use AI aggressively on throughput without inheriting the restatement, because the judgment that has to survive the assurer never got delegated to the machine.

The line is not arbitrary, and it is worth learning to place it precisely, because it is the single most consequential design decision in the whole redesign. The test is twofold. First, is the output checkable against a single authoritative source? Extracting a kilowatt-hour figure from a utility bill is checkable: the bill either says it or it does not. Deciding whether a Scope 3 category is material is not checkable against any single source, it is a judgment integrating evidence, framework, and context. Second, what is the cost of being wrong, and who bears it? A mislabeled draft caught at verification costs minutes. A boundary exclusion the assurer discovers costs a finding, possibly a restatement, and the CSO bears it. Work that is checkable and cheap to correct is throughput; work that is uncheckable and expensive to get wrong is judgment. When a stage sits ambiguously between the two, the safe default is to treat it as judgment and gate it, because the asymmetry of disclosure punishes the false confidence of treating judgment as throughput far more harshly than the mild inefficiency of over-gating.

There is a cultural dimension to this that a purely procedural redesign misses. Analysts under deadline pressure will, without meaning to, let AI creep across the line, because the machine's fluent output is seductive and the deadline is real. The workflow's labels are the guardrail, but the function also needs a shared understanding that using AI to do judgment is not a clever shortcut, it is the specific move that ends careers in this field. The best functions make this explicit in how they talk about the work: the machine is celebrated for throughput and firmly fenced out of judgment, and nobody is rewarded for a fast artifact that cannot be defended. Speed on throughput is the goal; speed bought by delegating judgment is the failure dressed as success.

The Redesigned Workflow: Stages, Labels, Gates, and Provenance

The redesigned function runs every disclosure artifact through the same explicit stages. The specifics differ by artifact, a Scope 3 figure and an ESRS narrative datapoint traverse different work, but the shape is constant: intake, AI-assisted throughput, human judgment, verification, sign-off gate, and file. Three properties make it a redesign rather than a relabel.

Explicit AI-Assisted and Human-Only Labels

Every stage carries a label the whole function shares. An AI-assisted stage means a machine did the work and its output is provisional until verified; a human-only stage means the decision is reserved to a named person and no AI output may substitute for it. The label is not decoration. It tells the analyst how much to trust the artifact in front of them, it tells the verifier what to check, and it tells the assurer where the machine ended and the human began. In the silo model this line lived in people's heads, differently in each head. In the redesign it is written on the workflow.

Sign-Off Gates at the Expensive Points

A sign-off gate is a point where a named human must approve before the artifact advances, and it is placed exactly where a wrong number would become expensive: after a factor is selected and before it enters the inventory; after an estimate is built and before it is labeled and disclosed; after a narrative is drafted and before it is tagged and filed. The gate is not bureaucracy for its own sake. It is the mechanism that keeps accountability human. Each gate records who approved, what they checked, and against what evidence, so the sign-off is itself an assurance artifact. A workflow with AI everywhere and no gates is fast and unaccountable. A workflow with gates at the expensive points is fast and defensible.

Provenance Built In, Not Bolted On

The deepest change is that provenance is captured as work happens, not reconstructed at the end. In the silo model, the assurance file was assembled in a panic after the report was drafted, from memory and scattered spreadsheets. In the redesign, every datapoint carries its source location, its primary-or-secondary status, its factor's named and dated origin, and its estimation method and uncertainty from the moment it is created, because the workflow will not let it advance otherwise. This is the single move that turns AI from an assurance liability into an assurance asset: when provenance is a property of the data rather than an afterthought, an AI-assisted number is assurable on exactly the same terms as a hand-built one, because the file reconstructs it end to end without the analyst in the room.

From Silos to Flow: What Changes in Practice

In the silo model the carbon accountant, the disclosure lead, and the value-chain data lead each ran a private end-to-end process. Handoffs between them were informal and undocumented, which is precisely where errors and untraceable numbers accumulated. The workflow model does not abolish those people; it threads them onto one flow with explicit handoffs. The value-chain data lead owns the intake and supplier-data stages; the carbon accountant owns the boundary and the factor-and-method gate; the disclosure lead owns the narrative and the faithfulness gate; the AI-Reporting Lead owns the AI-assisted stages and the verification; the Assurance Liaison owns the file that assembles continuously across all of it. The handoffs that used to live in email now live in the workflow, each with a label and, where it matters, a gate.

This is also how you make AI adoption honest. A silo can claim it "uses AI" while quietly letting the machine make judgment calls, and no one sees it until the assurer does. A shared workflow with labeled stages makes drift visible: if a materiality conclusion arrived through an AI-assisted stage with no human-only gate behind it, the workflow shows it and the artifact does not advance. The redesign is therefore both an efficiency change and a control: it captures the throughput speed the board wants and it enforces the judgment discipline the assurer requires, in one operating model rather than two competing ones.

Worked Example: Redesigning the Scope 3 Flow

Take the artifact where the redesign pays most: a single Scope 3 category figure, purchased goods and services, which alone can be a large slice of a 75%-of-footprint Scope 3 inventory built on data that 79% of reporters say they struggle to get.

Before, in the silo. The carbon accountant owns "Scope 3." She uses AI to spend-map suppliers to a category, has AI suggest emission factors, and pastes the results into her workbook. The narrative that describes the method is written separately by the disclosure lead, weeks later, from a verbal summary. Supplier non-responses are filled with averages inside the workbook with no flag. When the assurer asks how a given supplier's figure was derived, the accountant reconstructs it from memory, the factor's source is a half-remembered database, and the narrative in the report describes a method the workbook does not quite match. Fast to produce, impossible to defend.

After, in the workflow. Intake is a human-only boundary decision: the accountant documents which suppliers and spend are in scope and why. The value-chain data lead runs the AI-assisted supplier-data stage, and every response is tagged primary while every non-response is flagged as a disclosed gap, never averaged silently. Factor lookup is AI-assisted, and each candidate factor arrives with a named, dated source. Then the gate: the accountant signs off the factor and method, and the sign-off records what she checked against which evidence. The narrative is AI-drafted grounded on this exact evidence file, then passes the disclosure lead's faithfulness gate, which links each claim to its support and confirms the described method matches the data. Verification runs the checklist. The Assurance Liaison's file has been assembling the whole time, so when the assurer asks how that supplier's figure was derived, the answer is one click: source, factor with provenance, primary-or-estimated status, uncertainty, and the sign-off that approved it. The inventory closed in weeks and it reconstructs end to end. Same people, same AI, redesigned flow, opposite assurance outcome.

Key Takeaways

• Buying AI without redesigning the operating model gives you a faster version of the old process, and in disclosure a faster broken process is a faster route to a restatement.
• Task silos cannot absorb AI safely: each silo drops AI into its own private way of working, and the undocumented seams between silos hide errors at machine speed.
• The redesign replaces silos with a single explicit workflow every artifact passes through, with stages, labels, sign-off gates, and provenance built in.
• The organizing principle is humans on judgment, AI on throughput: AI does high-volume checkable work, humans own low-volume high-consequence decisions like boundary, materiality, estimation honesty, and sign-off.
• Drift is the danger: every time AI crosses from throughput into judgment, the function trades a defensible artifact for a fast one, invisibly, until the assurer pulls the thread.
• Explicit AI-assisted and human-only labels tell the analyst how much to trust an artifact, the verifier what to check, and the assurer where the machine ended and the human began.
• Sign-off gates sit at the expensive points and record who approved, what they checked, and against what evidence, so the sign-off is itself an assurance artifact that keeps accountability human.
• Provenance captured as work happens, not reconstructed at the end, is the single move that makes an AI-assisted number assurable on the same terms as a hand-built one, reconstructable without the analyst in the room.