What a $200 million manufacturer reveals about AI, Scope 3, and the evidence behind an emissions report ?

Direct answer: AI can make carbon accounting faster by extracting records, organizing supplier requests and finding candidate emission factors. A manufacturer still needs traceable activity data, appropriate factors, explicit boundaries, review controls and retained evidence before an emissions number is ready for customers or assurance.

OCEANS Sustainability is a carbon accounting platform from TheEvenity Pvt Limited for growing U.S. manufacturers. It connects Scope 1, Scope 2 and Scope 3 calculations with operational records, supplier information, calculation context and supporting evidence.

What this guide explains

• How AI can assist carbon accounting without hiding assumptions.
• Why manufacturers must assess relevant Scope 3 categories rather than treating missing data as zero.
• How electricity geography and emission-factor selection change reported results.
• What evidence another person needs to reproduce a carbon calculation.
• How OCEANS Sustainability structures emissions data, supplier evidence and reduction planning.

Imagine sending your biggest customer a carbon report on Friday.

On Monday, procurement replies: “Which emission factors did you use, and what did you exclude?”

If answering means rebuilding the calculation, the report was delivered before the work was finished.

That is the risk I want manufacturing leaders to examine as AI makes carbon accounting faster. A number can look complete while the decisions behind it remain unresolved.

Consider a synthetic automotive components supplier: $200 million in annual revenue, three plants in Michigan, Ohio, and Texas, and 40 direct suppliers. Two major automotive customers request emissions information through their supplier questionnaires.

This is a worked scenario, not a published benchmark of commercial tools. The company, transactions, customer exposure, and product assumptions below are illustrative. The EPA electricity factors are published data. Keeping those distinctions visible is part of the argument.

My background includes nine years in climate technology engineering, work with AERMOD atmospheric dispersion modeling, CPCB air quality backend systems, and GHG Protocol Scope 1, 2, and 3 methodology. That experience taught me to look behind a model’s output before trusting its precision.

AERMOD models air pollutant dispersion; it is not a corporate carbon accounting standard. The framework here is GHG Protocol, with EPA data where appropriate. Calling an inventory “EPA-grade” would blur that distinction. EPA’s description of AERMOD.

Where AI earns its place

There is useful work for AI throughout this manufacturer’s reporting process.

Supplier questionnaires arrive in different formats. AI can help extract questions, group repeated requests, and route them to procurement, finance, or plant operations. People still need to confirm that an answer actually addresses the question.

Utility bill extraction can turn pages of invoices into structured consumption records. Review can then focus on exceptions: an estimated meter reading, a duplicate bill, or a billing period crossing year end.

Emission factor search can become faster. The next step is still selecting the factor that fits the activity, geography, period, unit, and calculation boundary.

Spend classification can also accelerate an initial purchased goods estimate. Across 40 suppliers, this can help identify where requesting quantities or supplier emissions data would be most valuable.

These are useful applications to evaluate, not results from a tool benchmark. Their value depends on extraction accuracy, review controls, and preservation of source records.

An AI assisted workflow can preserve excellent evidence. A manual spreadsheet can lose it. The question is what the workflow lets another person verify.

Failure mode 1: An excluded category looks like zero

An empty category can conceal the largest assumption in the inventory.

Our supplier needs to assess purchased goods, upstream transportation, use of sold products, and end of life: Scope 3 Categories 1, 4, 11, and 12. Those four are the focus here, not a substitute for screening all 15 categories.

Category 11 requires particular care. A supplier of energy consuming motors and a supplier of passive brackets do not automatically have the same use phase boundary. GHG Protocol addresses direct use phase emissions from intermediate products themselves; it does not instruct a component supplier to claim the entire vehicle’s footprint. GHG Protocol’s intermediate product guidance.

For this scenario, assume the manufacturer sells 100,000 electric auxiliary motors in the reporting year. Each consumes an estimated 500 kWh over its life. Use an illustrative lifetime electricity factor of 0.30 kg CO₂e/kWh.

The calculation is 100,000 × 500 × 0.30 ÷ 1,000 = 15,000 metric tons CO₂e.

That is a sensitivity example, not a measured footprint. A real calculation needs supported product performance, lifetime, operating conditions, and electricity assumptions. Category 11 considers expected lifetime use of products sold during the reporting year. GHG Protocol Category 11 guidance.

Assume the rest of the screened inventory totals 25,000 metric tons. Omitting the motors would leave 15,000 out of 40,000 metric tons: 37.5% of the illustrative total.

The commercial exposure deserves a separate number. Suppose the two requesting customers account for 30% of revenue. That is $60 million in annual business connected to those customer relationships. It is not a forecast of lost sales, a regulatory penalty, or a claim of financial materiality under an accounting standard.

It tells the CFO why an unexplained exclusion deserves attention before the questionnaire goes out.

Failure mode 2: The factor has a value, but no address

“US electricity” can erase a 1,086-ton difference.

Assume the Michigan and Texas plants each purchase 10,000 MWh in the year and their service locations map to RFCM and ERCT respectively. Confirm actual subregions using location and electricity provider; state names alone are insufficient.

EPA eGRID2023 Revision 2 reports total output CO₂e rates of 975.978 lb/MWh for RFCM and 736.629 lb/MWh for ERCT. EPA eGRID summary data.

Using 0.45359237 kilograms per pound, the location based calculations are:

Plant assumption	Annual electricity	Calculated emissions
Michigan, RFCM	10,000 MWh	4,427 metric tons CO₂e
Texas, ERCT	10,000 MWh	3,341 metric tons CO₂e
Difference	Same consumption	1,086 metric tons CO₂e

The RFCM result is approximately 32.5% higher. Applying the ERCT factor to both plants would understate the Michigan result by approximately 24.5%.

The arithmetic is simple. Selecting and retaining the right factor is the substantive work. The Ohio plant needs its own mapping. Contractual electricity claims also need a separate assessment; this example covers location based accounting only.

The same provenance problem appears in purchased goods. EPA’s Supply Chain GHG Emission Factors v1.3 use kg CO₂e per 2022 US dollar. A spend calculation must retain the selected industry code, factor variant, and dollar basis. A dollar denominated factor cannot simply be multiplied by kilograms of steel. EPA dataset documentation.

Spend estimates can be legitimate inventory inputs, including within an assured inventory when appropriate. Their limitations need to remain visible.

Failure mode 3: “95% confident” does not explain 95% of anything

Confidence needs a definition before it deserves a percentage.

Imagine a Category 1 result marked “95% confidence.”

Does that describe invoice extraction accuracy? Confidence in the supplier’s industry classification? A statistical interval around the emissions estimate? Those are different claims.

A classifier can be very certain that an invoice concerns steel while the selected factor poorly represents the steel’s production route.

I would ask to open one calculation record and see the invoice, purchased quantity, factor identifier and version, unit conversion, boundary decision, uncertainty description, and exclusions.

For data quality, assess technological, temporal, and geographical fit, alongside completeness and reliability. GHG Protocol explicitly identifies these dimensions. GHG Protocol factor selection guidance.

A useful note might read: “Purchased mass reconciled to receiving records; emissions calculated using supplier production data; product allocation unverified; moderate confidence in representativeness.”

That gives a reviewer something to investigate. An unexplained percentage gives them a decoration.

Failure mode 4: The PDF becomes the stopping point

A polished report can still contain an unreproducible number.

Take one total from the report. Can a colleague recover the underlying activity, applied factor, conversion, and inclusion decision without asking the original preparer to reconstruct their reasoning?

If the factor changes next year, can that colleague explain whether the reported reduction came from operations, purchasing, or a database update?

If a supplier sends a corrected file, can they identify every affected calculation?

These are practical review questions. Reproducibility supports assurance; it does not itself constitute an assurance opinion. A third party still evaluates the inventory against the applicable criteria and engagement requirements.

The PDF is a communication format. The calculation records carry the explanation.

One purchased steel calculation, with its evidence attached

Here is how I would structure one Category 1 line for the synthetic manufacturer.

Start with 1,000 metric tons of purchased steel, or 1,000,000 kg, documented by supplier invoices and matched receiving records for the reporting year. Reconcile returns, duplicates, and unit conversions before calculating emissions.

Next, select the method. GHG Protocol provides supplier specific, hybrid, average data, and spend based approaches for purchased goods. Its supplier specific approach uses purchased quantities and supplier product emissions data covering upstream production. GHG Protocol Category 1 calculation guidance.

For this teaching example, assume a supplier product footprint of 1.80 kg CO₂e/kg steel, covering raw material extraction through the supplier gate. This factor is synthetic. It is not an EPA factor or a quoted commercial database value.

The factor source in the example is a supplier declaration, document STEEL-PCF-2025, revision 1.0, stored in the manufacturer’s factor register. The document identifier is also illustrative. No external database version is implied.

In practice, obtain the actual declaration and examine production route, site, reporting period, allocation, included processes, and the basis used to convert different gases into CO₂e. If using a secondary database instead, retain its exact dataset identifier, release, geography, unit, and system boundary.

The calculation is 1,000,000 kg × 1.80 kg CO₂e/kg = 1,800,000 kg CO₂e, or 1,800 metric tons CO₂e.

Keep the boundary explicit. This line includes upstream steel production. Supplier gate to our plant transportation is tracked separately in Category 4. Our plant’s own fuel and purchased electricity sit in Scopes 1 and 2. Other materials need their own records.

Before adding transport, inspect whether the supplied footprint already includes delivery. Otherwise, one freight movement can appear twice.

Here is the evidence record I would want a reviewer to receive:

Field	Illustrative entry
Record	CAT1-STEEL-2025-001
Activity source	Invoice and receiving record references; reconciliation retained
Quantity	1,000,000 kg purchased steel
Factor source	Synthetic supplier declaration STEEL-PCF-2025
Version and unit	Revision 1.0; 1.80 kg CO₂e/kg
Boundary	Raw material extraction through supplier gate
Calculation	1,000,000 × 1.80 ÷ 1,000 = 1,800 tCO₂e
Data status	Quantity documented; factor modeled; emissions calculated
Uncertainty	Supplier allocation and process coverage require review; no statistical interval established
Exclusion log	Delivery freight recorded in Category 4; other goods handled separately
Review history	Preparer, reviewer, dates, source files, and change reason retained

“Documented,” “calculated,” and “estimated” describe different parts of the record. They are not substitutes for a quantitative uncertainty analysis or formal assurance classifications.

For an illustrative sensitivity check, varying the factor from 1.50 to 2.10 yields 1,500 to 2,100 metric tons. That range is an assumed scenario, not a 95% confidence interval.

This line is not the whole Category 1 inventory. Across 40 suppliers, repeat the exercise, reconcile purchasing coverage, identify services and missing materials, and document how remaining gaps are estimated. Improving one steel calculation cannot justify leaving the rest invisible.

The customer relationship brings the issue home

SB 253 requires US based entities with more than $1 billion in annual revenue that do business in California to report Scope 1 and 2 beginning in 2026, and Scope 3 beginning in 2027. CARB set August 10, 2026 as the first year reporting deadline. CARB’s implementation announcement.

Our standalone $200 million manufacturer falls below that revenue threshold. Corporate group circumstances need separate assessment. Its immediate pressure in this scenario comes from customers requesting supplier data for their own reporting and procurement processes.

A customer’s questionnaire may ask for a corporate inventory, a product footprint, or both. Those deliverables have different boundaries. Clarifying the request should happen before anyone exports a total.

For leadership, a useful test is to choose one material emissions line and ask someone outside the preparation team to reproduce it. Record what they cannot find. That becomes the next improvement list.

Speed matters when teams are assembling thousands of records. It matters even more when the same records must support next year’s explanation.

Your carbon number is only as credible as the evidence behind it.

Frequently asked questions

Can AI calculate carbon emissions for manufacturers?

AI can help extract, classify and organize the underlying data. The resulting emissions calculation still requires an appropriate methodology, reviewed activity data, applicable emission factors, documented boundaries and human oversight.

What is evidence-linked carbon accounting?

Evidence-linked carbon accounting keeps each reported result connected to its source activity data, emission factor, calculation method, assumptions, exclusions and review history. This makes a result easier to reproduce, explain and improve.

Why is Scope 3 difficult for manufacturers?

Scope 3 often depends on purchasing records, supplier-specific information, logistics data, product-use assumptions and secondary factors held across different systems and organizations. Data quality and calculation boundaries can vary materially by category.

What is OCEANS Sustainability?

OCEANS Sustainability is a carbon accounting platform from TheEvenity Pvt Limited for growing U.S. manufacturers. It connects activity data, Scope 1–3 calculations, supplier information, emission factors and supporting evidence so teams can understand emissions hotspots and evaluate reduction options.

This is the gap OCEANS Sustainability was built to close: connecting activity data, emission factors and the evidence chain within a clear Scope 1, Scope 2 and Scope 3 structure. Explore the OCEANS Sustainability platform, review its carbon accounting methodology, or learn about the Founding Design Partner Program.

By Praneeth, Founder of OCEANS Sustainability at TheEvenity Pvt Limited