A sustainability manager at a mid-size construction firm described her Scope 3 data collection process as sending emails into the void. She had emailed 35 suppliers requesting emissions data. Eight replied. Three sent something usable. The other five sent PDFs ranging from a single unexplained number to a 60-page sustainability report with the relevant figure buried on page 43. That exercise consumed six months.
This story is not unusual, and it is the reason every carbon platform now has an AI story for Scope 3. The pitch is seductive: stop chasing suppliers, let the model fill the gaps. But the pitch quietly misidentifies the problem — and as of March 2026, a change to the GHG Protocol has made that misidentification expensive. AI has genuinely transformed parts of Scope 3 collection. It has transformed almost nothing about the part most people expect.
Who should read this
- Sustainability and ESG managers
- Procurement and category managers
- Carbon accounting leads
- CSRD and ESRS reporting teams
- Supply chain directors
- Anyone evaluating Scope 3 software
🔑 Key takeaways
- Scope 3 typically represents 70–90% of a corporate footprint, and Category 1 (purchased goods and services) usually dominates it — in chemical supply chains, 70–85% of a product footprint sits there.
- AI changed three things well: triage before engagement, product-level decomposition instead of spend-based estimation, and continuous monitoring instead of an annual snapshot.
- AI changed almost nothing about supplier response. A supplier that does not measure its emissions cannot report them, and a category manager with an existing relationship still outperforms an automated survey.
- The March 2026 GHG Protocol draft revisions raise the stakes: disclosing data by quality tier and verification status makes plausible-but-unverified AI estimates visibly weak rather than quietly acceptable.
- The winning model is hybrid: AI narrows five thousand suppliers to the two hundred that matter; humans convert those two hundred into primary data.
On this page
- Why Scope 3 collection is structurally hard
- The four methods, and what they cost you
- What changed in March 2026
- What AI actually changed
- What AI has not changed
- The estimation paradox
- The operating model that works now
- What this means for platform selection
- Common mistakes
- The bottom line
- Frequently asked questions
Why Scope 3 Collection Is Structurally Hard
Scope 1 and Scope 2 are measurement problems you own. You have the fuel invoices, the meter readings, the fleet records. The work is tedious but the data exists inside your organisation and someone can be told to produce it.
Scope 3 is a different category of problem, because the data lives inside companies that have no obligation to help you. Value-chain emissions typically account for 70 to 90% of a company’s total footprint, so the majority of your carbon number depends on information you cannot compel. Your leverage is commercial, not managerial — and for a supplier who represents 0.3% of your spend and for whom you represent 0.1% of theirs, that leverage rounds to zero.
Three compounding constraints make it worse. Most suppliers, particularly small and mid-size ones, have never calculated their own emissions and cannot answer even a well-designed request. Those that can often report on a different basis — different boundaries, different base year, different scope definitions — so the answers are not comparable. And the volume is punishing: a manufacturer with five thousand suppliers cannot run a meaningful conversation with each one, regardless of budget.
The Four Methods, and What They Cost You
The GHG Protocol recognises four calculation methods, in increasing order of accuracy, and recommends using the most accurate available for each category. Understanding what each one actually buys you is the prerequisite for judging any AI claim.
| Method | How it works | Practical use |
|---|---|---|
| Spend-based | Spend value × sector emission factor | Fast baseline for the whole supplier base; lowest accuracy |
| Activity-based | Physical quantities × activity factors | Better where you know volumes, weights or distances |
| Supplier-specific | Actual emissions reported by the supplier | Top suppliers by emissions exposure; reflects their real progress |
| Product-specific | Product carbon footprint for a specific item | Highest accuracy; needed for CBAM and product-level claims |
The uncertainty attached to each tier is what makes this hierarchy consequential rather than academic. Industry-average secondary data carries roughly 30–40% uncertainty for a specific product; generic emission factors from a database push that past 50% at product level. A footprint built entirely on generic factors is not wrong so much as it is uninformative — it cannot detect the improvement you are trying to demonstrate.
Two frameworks push in the same direction. Under CSRD’s ESRS E1, in-scope companies must disclose what proportion of their Scope 3 inventory is primary data versus secondary. And the SBTi requires near-term targets to cover at least 67% of Scope 3 emissions — a threshold reachable through a combination of reduction targets and supplier engagement targets. Neither expects perfection. Both expect a planned, visible migration from averages toward primary data.
What Changed in March 2026
On 31 March 2026 the GHG Protocol published its Scope 3 Phase 1 Progress Update — the first substantial overhaul of Scope 3 guidance since 2011, following 42 technical working group meetings between September 2024 and the end of 2025. It is a draft, not a final rule, with final standards expected in 2027. The direction, however, is unambiguous, and four proposed revisions bear directly on how AI-generated data will be judged.
| Revision | What it proposes | Why it matters for AI |
|---|---|---|
| A1 | Disaggregate Scope 3 data by quality tier | Spend-based proxies get labelled as the lowest tier — visibly |
| A2 | Disclose whether data is verified, partly verified, or not verified | Unverified model output must be declared as such |
| A8 | Restrict company-level allocation to single-industry suppliers | Diversified suppliers need product- or site-level data instead |
| B1 | Set a 95% minimum coverage floor with justified exclusions | Tail spend can no longer be quietly left out |
Read A1 and A2 together and the implication is sharp. Today, a Scope 3 total built largely on modelled estimates looks much like one built on supplier data — a single number in a report. Under the proposed revisions, the composition becomes visible: how much sits in each quality tier, and how much has been verified. An AI that produces confident estimates without improving provenance does not make your disclosure stronger. It makes the weakness legible.
What AI Actually Changed
Across the supply chain ESG and carbon platforms we review, three genuine shifts have taken hold. Each is narrower than the marketing suggests, and each is real.
Triage before engagement
Screening thousands of suppliers to find the few hundred that actually drive the footprint — before a single email is sent.
Decomposition, not estimation
Breaking a purchased good into its materials and manufacturing processes rather than multiplying spend by a sector average.
Continuous signal
Monitoring supplier risk continuously from external sources instead of relying on one annual questionnaire cycle.
1. Triage — knowing where to look
The old sequence was to email everyone and analyse whatever came back. The new sequence inverts it: analyse first, engage second. EcoVadis layers AI risk screening over its network of more than 175,000 rated companies, profiling millions more beyond the formally rated set so a buyer can risk-classify an entire base before commissioning a single assessment. IntegrityNext operates across more than two million suppliers in 190-plus countries, scoring risk from self-assessment data enriched with third-party financial records and live news monitoring.
The practical effect is a reallocation of scarce human attention. If your team can run twenty serious supplier conversations a quarter, the question that matters is which twenty — and AI answers that question far better and far cheaper than a spreadsheet sorted by spend.
2. Decomposition — better than spend × factor
This is the most technically interesting shift. Watershed runs on a library exceeding 500,000 emission factors including CEDA, and its Product Footprints capability uses AI to decompose purchased goods into constituent materials and manufacturing processes — moving Category 1 past spend-based estimation without waiting for the supplier to respond.
The distinction matters. Spend-based estimation says “you spent €2M with a chemicals supplier, sector average implies X tonnes.” Decomposition says “this product is 40% polymer A, 30% process B, shipped this distance” and builds up from there. The second is still an estimate, but it is a structured, defensible one that responds to product design changes — which means it can actually detect a reduction.
3. Continuous signal — beyond the annual snapshot
Annual questionnaires produce a photograph of a moving object. AI-driven monitoring across news, regulatory and financial sources turns supplier risk into something closer to a live feed, and engagement scoring — rating suppliers on responsiveness, completeness and data quality tier — turns a shapeless chase into a managed pipeline where every supplier has a status and every request a deadline.
The amber step is where AI earns its place: deciding which suppliers deserve a human conversation. Everything downstream of that decision is better because the decision was better.
What AI Has Not Changed
Three constraints have survived the AI wave intact, and any vendor conversation that glosses over them is worth interrogating.
A supplier that does not measure cannot report
No model can extract primary data from a company that has never calculated its emissions. AI can estimate that supplier’s footprint — usefully, even accurately — but the output is still an estimate, and under the proposed A2 revision it will have to be labelled as unverified. The only route to primary data runs through the supplier actually doing the measurement, which usually means helping them.
Response rates are a relationship problem
The most consistent finding in supplier engagement practice is unglamorous: a category manager making a direct request through an existing commercial relationship reliably outperforms an automated survey from an unfamiliar platform. Short supplier workshops explaining what data is needed and why yield materially higher response rates than email alone. Programmes that work assign responsibility to category managers who already talk to these suppliers — not to one person in the sustainability team with a distribution list.
Migration takes reporting cycles, not weeks
Moving the top ten to twenty suppliers from spend-based to activity or supplier-specific data typically takes one to two full reporting cycles. That timeline reflects procurement calendars, supplier capability building and internal data plumbing — none of which a model accelerates. Any plan that assumes primary data across a major category within a single cycle is a plan that will slip.
The Estimation Paradox
Here is the tension that most Scope 3 AI marketing does not resolve. Better estimation makes your number more accurate. It does not make your disclosure stronger — and under the proposed GHG Protocol revisions it may make the weakness more visible.
Consider two companies with identical reported Scope 3 totals. Company A used a sophisticated AI model to estimate 90% of its footprint from spend and product data. Company B collected supplier-specific data covering 40% of its footprint and estimated the rest. Today, their reports look similar. Under revisions requiring disclosure by quality tier and verification status, Company B’s disclosure is visibly stronger, and the gap will be legible to any assurance provider, investor or SBTi validator who reads the composition rather than the headline.
The resolution is not to use less AI. It is to use AI for the jobs that improve provenance rather than substitute for it: finding the suppliers whose data would move the most, decomposing products into structured components that can later be validated against real supplier figures, and maintaining the engagement pipeline that converts estimates into primary data over time. Estimation is the fallback, not the destination.
The Operating Model That Works Now
The pattern that produces credible Scope 3 numbers in year two — rather than a defensible-looking year one that stalls — follows a consistent sequence.
- Build the spend-based baseline first. It is the map, not the destination. Without it you cannot identify which suppliers matter, and every later step depends on that ranking.
- Let AI rank the base by emissions exposure, not by spend. The two lists differ substantially — a modest chemicals or logistics spend can outweigh a large services contract.
- Pick the engagement cohort you can actually serve. Twenty to two hundred suppliers, depending on team size. A shorter list handled properly beats a long list handled by autoresponder.
- Route requests through category managers who already own those relationships, with the sustainability team supplying the technical content rather than sending the emails.
- Use AI decomposition on the long tail so the remainder is structured and product-linked rather than a flat sector average.
- Track and disclose by quality tier from the first cycle. Building the tier breakdown now costs little and positions you for revisions A1 and A2 rather than retrofitting under deadline.
What This Means for Platform Selection
The capability question to ask a vendor is not “do you use AI for Scope 3?” — everyone says yes. It is which of the three shifts they actually deliver, and whether they improve provenance or merely fill gaps.
| Your dominant problem | Capability to test | Platforms built around it |
|---|---|---|
| Too many suppliers to triage | Risk screening across a large network | EcoVadis, IntegrityNext |
| Category 1 dominates the footprint | Product-level decomposition, factor depth | Watershed |
| Calculation must be independently verified | Verified methodology, documented lineage | Normative |
| Financed emissions dominate | PCAF attribution and data quality scoring | Persefoni |
| Data scattered across entities | Multi-entity collection governance | Sweep, KEY ESG |
| Disclosure governance and assurance | Audit trail, framework mapping | Novisto |
One question separates serious vendors from the rest: ask how their platform will represent your data under a quality-tier disclosure requirement. A vendor who has thought about revisions A1 and A2 will answer immediately. A vendor selling gap-filling will change the subject.
Common Mistakes
Process mistakes
- Emailing the entire supplier base before ranking it by emissions exposure.
- Assigning supplier engagement to one sustainability team member instead of category managers.
- Treating a spend-based baseline as the deliverable rather than the map.
- Planning for primary data coverage within a single reporting cycle.
Data and disclosure mistakes
- Reporting one Scope 3 total without recording the quality tier composition behind it.
- Accepting AI estimates as a permanent answer rather than an interim position.
- Excluding tail spend quietly — the draft 95% coverage floor closes that door.
- Restating the base year each cycle so improvements become impossible to demonstrate.
The Bottom Line
AI has genuinely changed supply chain ESG, but in a narrower and more useful way than the pitch decks claim. It made triage cheap, decomposition possible, and monitoring continuous. Those are real gains, and a programme that ignores them is leaving significant capacity on the table.
What it did not do is remove the human work at the centre. The suppliers who matter still have to be asked, by someone they know, for data many of them must first learn to produce — and the March 2026 draft revisions to the GHG Protocol will make it increasingly visible which companies did that work and which automated around it.
AI has made it dramatically cheaper to know which two hundred suppliers matter. It has not made it any easier to get them to answer. The first is a software problem, now largely solved. The second is a relationship problem — and no model has solved that one yet.
Frequently Asked Questions
Can AI replace supplier data collection for Scope 3?
No. AI can estimate a supplier’s emissions from spend, product composition and sector data, and it can do so far better than a flat sector average. But an estimate is not primary data, and a supplier that has never measured its emissions cannot supply primary data no matter how good the model is. Under the GHG Protocol’s proposed March 2026 revisions, unverified estimates would have to be disclosed as such — so AI supplements collection rather than replacing it.
What did the GHG Protocol change for Scope 3 in 2026?
On 31 March 2026 the GHG Protocol published its Scope 3 Phase 1 Progress Update, the first substantial overhaul since 2011, following 42 technical working group meetings. Four draft revisions matter most: A1 requires disaggregating data by quality tier, A2 requires disclosing whether data is verified, A8 restricts company-level allocation to single-industry suppliers, and B1 sets a 95% minimum coverage floor with justified exclusions. It is a draft, with final standards expected in 2027.
How much of a company’s footprint is Scope 3?
Typically 70 to 90% of total emissions, though it varies sharply by sector. Category 1 — purchased goods and services — usually dominates; in chemical supply chains, 70 to 85% of a product carbon footprint commonly sits in that single category. This concentration is why triage works: a relatively small number of suppliers usually accounts for the large majority of the footprint, and identifying them is the highest-leverage step in the whole process.
Do we need primary data from every supplier for CSRD or SBTi?
No, and neither framework asks for it. Under CSRD’s ESRS E1, companies must disclose what proportion of their Scope 3 inventory is primary versus secondary data — a transparency requirement, not a completeness one. The SBTi requires near-term targets covering at least 67% of Scope 3 emissions, achievable through a combination of reduction targets and supplier engagement targets. Both expect visible progress from averages toward primary data, not immediate perfection.
How long does it take to move suppliers from spend-based to primary data?
For the top ten to twenty suppliers by emissions exposure, typically one to two full reporting cycles. The constraint is rarely software — it is procurement calendars, supplier capability building, and internal data integration. Programmes that plan for a single cycle usually slip; programmes that treat year one as baseline and triage, and year two as the first primary-data cycle, tend to deliver credible numbers on schedule.
What actually improves supplier response rates?
Relationships and specificity, more than technology. A category manager making a direct request through an existing commercial relationship consistently outperforms an automated survey from a platform the supplier does not recognise. Framing the request around your own targets rather than regulatory obligation helps, as does keeping the initial ask small and concrete. Short supplier workshops explaining what data is needed and how to produce it yield materially better results than email alone.
Where to Go Next
Explore the platforms built for each layer of this problem in the Supply Chain ESG and carbon accounting categories — including Watershed, Normative, EcoVadis and IntegrityNext. For the wider technology shift, see how AI is transforming carbon accounting in 2026; for the regulatory backdrop, our CSRD guide. To turn your inventory into a costed pathway, use the free Net Zero Roadmap Toolkit, and see how every platform is scored in our methodology. Primary sources: the GHG Protocol and the Science Based Targets initiative.
