A disclosure lead is reviewing a Scope 3 category her team has just closed, and on paper it looks finished. Every supplier row has a number. The total is clean. The category reconciles. Then she asks one question: of the two hundred suppliers in this category, how many actually responded? The analyst checks, and the answer is one hundred and twenty. Eighty suppliers never replied, and yet there are two hundred numbers. Somewhere in the pipeline, eighty holes were quietly filled with an average and folded into the total, and now the category reads as complete when it is 40% estimated from non-responders nobody flagged. This is the most dangerous outcome in value-chain data work, more dangerous than a wrong number, because a wrong number can be found and corrected while a hidden gap is invisible by design. The whole point of this lesson is to make the holes visible: to use AI to flag non-responses, partial responses, and implausible values so that every gap is disclosed and handled with a labelled estimate, never erased into a tidy average. A disclosed gap beats a hidden fabrication, every time.

The Gap Is Information, Not an Embarrassment

The instinct that creates hidden gaps is the belief that a hole in the data is a failure to be concealed before anyone important sees it. That instinct is exactly backwards. In an assured disclosure, the gap is information the assurer needs and is entitled to. Knowing that eighty of two hundred suppliers did not respond tells the reader precisely how much of the category rests on estimation rather than primary data, which is a core input to how much they should trust the figure. Hiding the gap does not make the data better; it removes the reader's ability to judge the data at all, and it converts an honest limitation into a concealment that, when discovered, reads as fabrication.

This is the asymmetry at the heart of the lesson. A disclosed gap is a managed limitation: you state that eighty suppliers did not respond, you cover them with a labelled estimate whose method and uncertainty are stated, and the assurer can see exactly what is measured, what is estimated, and how complete the coverage is. A hidden gap is a fabrication waiting to be found: the same eighty suppliers are silently averaged in, the category looks complete, and when the assurer samples a "supplier figure" that turns out to be an undisclosed fill, you have a finding that questions the integrity of the whole inventory. The number can be identical. One is good practice; the other ends the engagement badly. The entire difference is whether the hole was shown or buried.

So the goal is not to eliminate gaps, which is impossible when 79% of reporters cannot reliably get supplier data. The goal is to make every gap visible, classify it, and handle it openly. AI is genuinely useful here because finding and classifying gaps across a large dataset is a detection task, and detection is something the technology does well, provided you ask it to surface the holes rather than to fill them.

Three Kinds of Gap AI Can Flag

Not all gaps are the same, and treating them as one undifferentiated "missing data" problem hides important distinctions. There are three kinds, and AI can flag each.

The Non-Response

The first and simplest is the outright non-response: a supplier who was asked and never replied. AI flags these by reconciling the list of suppliers contacted against the list of usable responses received, and the difference is your non-response set. This sounds trivial, but it is the gap most often lost, because a non-responder produces nothing to look at, so it disappears unless something actively counts the absence. The flag turns silence into a named, counted set: these eighty suppliers, representing this much spend or activity, did not respond. That set is now a visible object you can handle, rather than a void the pipeline fills on its own.

The Partial Response

The second is the partial response: a supplier replied, but the response is incomplete. They gave a value with no unit, a number with no period, a figure with no statement of whether it was measured or estimated, or they answered some questions and skipped others. AI flags partials by checking each response against the required fields and marking which are present and which are missing. The partial is the most recoverable gap, because the supplier is engaged and a follow-up often closes it, which is exactly why surfacing partials, rather than silently accepting the half-answer, is high-value: it routes a chaseable supplier to the analyst instead of letting an incomplete record settle into the inventory as if it were whole.

The Implausible Value

The third is subtler: the response that arrived complete but is probably wrong. A figure an order of magnitude off the supplier's prior year. An emissions intensity far outside the plausible range for that sector. A unit that implies an impossible quantity. These are not missing data, but they are gaps in reliability, and an implausible value accepted at face value corrupts the inventory as surely as a missing one. AI flags these by comparing each value against prior-period figures, sector benchmarks, and internal consistency, and surfacing the outliers for human investigation. The flag does not declare the value wrong; it declares the value worth a second look before it is trusted, which is the cheapest fraud-and-error check available.

A disclosed gap is a managed limitation an assurer respects. A hidden gap is a fabrication waiting to be found. The number can be identical; the only difference is whether you showed the hole or buried it.

Flag, Then Handle: Never Fill Silently

Flagging is only half the discipline. Once a gap is visible, it has to be handled, and the way it is handled is what separates good practice from the laundered fill. The cardinal rule is that a gap is never filled silently. There are exactly two honest ways to handle a flagged gap, and a third dishonest one to forbid.

The first honest handling is to close the gap: chase the non-responder, follow up on the partial, query the implausible value, and obtain the real data. This is always the preferred move, and surfacing the gap is what makes it possible, because you cannot chase a supplier you never noticed was missing. The second honest handling is to cover the gap with a labelled estimate: where the real data genuinely cannot be obtained in time, build a transparent estimate, label it explicitly as an estimate, state its method and uncertainty, and disclose that it covers a non-responding portion. This is defensible because it tells the truth about itself, exactly as the estimation lesson taught. The forbidden third handling is the silent fill: dropping a plausible average into the missing rows, unlabelled, so the category looks complete and measured when it is neither. The silent fill is fabrication, and it is the precise thing the flag exists to prevent. The flag makes the gap visible so that you are forced to choose one of the two honest handlings, rather than letting the pipeline default to the dishonest one.

This is where AI is most dangerous if misdirected. Asked to "complete the category," a generative model performs the silent fill by default, fluently and invisibly, because completing the table is what it was told to do. Asked instead to "flag every supplier that did not respond, every incomplete response, and every value that looks implausible, and do not fill anything," the same model becomes the instrument that makes the gaps visible. The difference is entirely in the instruction. Never ask the model to complete or fill; ask it to flag, count, and surface, and handle the flagged gaps yourself.

A Worked Example: One Category, Two Closings

A company is closing Scope 3 Category 1 for two hundred suppliers. One hundred and twenty responded usably; eighty did not. The eighty non-responders represent a known portion of purchased-goods spend. Watch two closings.

Before (the hidden gap, what looked finished): The analyst asks AI to "produce the complete Category 1 emissions for all two hundred suppliers." The model returns a single clean total, with the eighty non-responders silently filled using a generic sector average and folded invisibly into the figure. Every row has a number. The category reconciles and looks complete. There is no flag that eighty suppliers never replied, no line distinguishing the measured portion from the filled portion, no method note, no uncertainty. The disclosure lead, reviewing the tidy output, has no reason to suspect anything until she asks how many actually responded. When the assurer later samples a non-responder's "figure," it is revealed as an undisclosed average, and the finding is not "one number is off" but "this category concealed that 40% of it was estimated," which casts doubt on the entire inventory. The clean total was the most dangerous artifact in the file.

After (the disclosed gap, the defensible closing): The analyst instructs AI to reconcile contacted suppliers against usable responses and flag every non-response, partial, and implausible value, filling nothing. The model returns: one hundred and twenty usable responses recorded with their provenance and tiers; eighty named non-responders flagged as a counted set with their associated spend; a handful of partials flagged for follow-up; and three implausible values flagged for investigation. The analyst chases what she can, recovering some non-responders and closing several partials. For the residual non-responders she genuinely cannot reach, she builds a labelled spend-based estimate, states the method and uncertainty, and discloses that it covers the non-responding portion. The closed category now shows the split between measured and estimated, names the coverage gap, and states how the gap was handled. When the assurer asks how complete the category is and how much is estimated, the analyst answers in one sentence and points to the labelled portion and the named gap. The total might be nearly the same as the hidden version. The difference is that this one told the truth about its own holes, and survives.

Same two hundred suppliers, same eighty non-responders, possibly the same final tonnage. One closing buried the gap and produced a finding; the other flagged the gap and produced a defensible disclosure. The flag was the hinge.

Not All Gaps Are Equal: Weighting by Materiality

Once the gaps are flagged, a second judgment matters: not every hole deserves the same effort, because not every hole moves the number by the same amount. A non-responder representing a few thousand euros of office-supply spend and a non-responder representing a quarter of your purchased-goods footprint are both gaps, but they are not equally important to close, and treating them identically wastes scarce chasing effort on the trivial while the material one waits. This is why naming the spend or activity behind each gap is not bookkeeping but prioritisation: it tells you which holes actually threaten the figure and the assurer's view of completeness, and which are rounding error.

The practical discipline is to rank the flagged gaps by their materiality to the category total, then spend your effort top-down. The material non-responders get chased hardest, because a real number from a large supplier is worth far more than any estimate. Where a material gap cannot be closed, it gets the most careful labelled estimate, with the tightest method and the most honest uncertainty, because that is the portion the assurer will probe most closely. The immaterial gaps can be covered more coarsely, with a clearly labelled estimate, because their effect on the total is small and the cost of perfect precision is not worth it. Materiality does not change the rule that nothing is filled silently; every gap, large or small, is still flagged and handled openly. What materiality changes is where the human spends the hours that flagging freed up, concentrating judgment where it actually protects the number rather than spreading it thin across holes that do not matter.

This ranking is also what makes the gap report legible to an assurer rather than a wall of equal-looking flags. A report that says "here are the three material non-responders, representing 30% of the category, each chased and then covered by a tightly bounded estimate, and here are the forty immaterial ones covered by a coarse labelled average" tells the assurer exactly where the risk sits and that you concentrated your effort there. A report that lists forty-three undifferentiated gaps with no sense of which matter tells the assurer you did not think about materiality at all, which is itself a weakness. The flag makes the hole visible; the materiality weighting makes the handling intelligent, and the two together turn a pile of missing data into a prioritised, defensible plan the assurer can follow line by line.

Working Rules for Flagging and Handling Gaps

A handful of rules keep gaps visible and honestly handled. Treat the gap as information the assurer needs, not an embarrassment to conceal, because a disclosed gap is a managed limitation and a hidden one is a fabrication waiting to be found. Use AI to flag the three kinds of gap, the non-response, the partial, and the implausible value, by reconciling contacted against received, checking responses against required fields, and comparing values against prior periods and benchmarks. Never ask a model to complete or fill a category; ask it to flag, count, and surface, and forbid silent filling outright. Handle every flagged gap one of the two honest ways: close it by chasing the real data, which surfacing made possible, or cover it with a labelled estimate whose method and uncertainty are disclosed. Forbid the silent fill absolutely, because dropping an unlabelled average into missing rows is the exact fabrication the flag exists to prevent. Count and name non-responders with their associated spend or activity, so the coverage gap is a visible, quantified object rather than a void. Surface partials for follow-up rather than accepting half-answers, because the partial is the most recoverable gap and the supplier is already engaged. Flag implausible values for a second look before trusting them, since an outlier accepted at face value corrupts the inventory as surely as a missing one. Follow these and AI turns gap-handling into the part of the pipeline an assurer respects most. Ignore them and AI turns your holes into invisible fabrications that surface at the worst possible moment.

Key Takeaways

• A hidden gap is more dangerous than a wrong number, because a wrong number can be found and corrected while a gap silently filled with an average is invisible by design until an assurer pulls the thread.
• The gap is information the assurer needs and is entitled to, not an embarrassment to conceal; hiding it removes the reader's ability to judge the data and converts an honest limitation into a concealment that reads as fabrication.
• A disclosed gap is a managed limitation an assurer respects; a hidden gap is a fabrication waiting to be found, and the number can be identical, so the only difference is whether you showed the hole or buried it.
• AI can flag three distinct kinds of gap: the non-response (asked, never replied), the partial response (replied but incomplete), and the implausible value (complete but probably wrong).
• Non-responses are the gap most often lost, because a non-responder produces nothing to look at, so they disappear unless something actively reconciles contacted suppliers against usable responses and counts the absence.
• A flagged gap must be handled one of two honest ways: closed by chasing the real data, which flagging makes possible, or covered with a labelled estimate whose method and uncertainty are disclosed.
• The silent fill, dropping an unlabelled average into missing rows so the category looks complete, is fabrication and the precise thing the flag exists to prevent; never ask a model to complete or fill, ask it to flag and surface.
• Count and name non-responders with their associated spend, surface partials for follow-up, and flag implausible values for a second look, so every hole is a visible, quantified object handled openly rather than a void the pipeline fills on its own.