A sustainability figure can be easy to calculate and difficult to defend. The source may cover the wrong period, the unit may be ambiguous, or the same supplier record may appear twice under different filenames. An AI system can turn that material into a polished table without resolving any of those problems.
For a CFO, controller or sustainability reporting owner, the priority is a traceable path from each reported figure to its source, method and review. AI can help collect and prepare the evidence. The reporting process still needs to determine what the evidence means and whether it belongs in the disclosure.
Start by building an evidence register for one material reporting line. That small step exposes the decisions a larger automation project must preserve.
Establish the reporting requirement first
The business case depends on the organisation's actual obligations and stakeholder requests. Singapore's timetable is differentiated by entity type and, for listed companies, the relevant tier. It should not be reduced to a statement that every business must immediately produce the same disclosures.
The Accounting and Corporate Regulatory Authority (ACRA) reporting and assurance requirements page, updated in July 2026, states that all Singapore Exchange-listed companies report Scope 1 and 2 greenhouse-gas emissions from FY2025. Scope 1 covers direct emissions from owned or controlled sources; Scope 2 covers emissions from purchased energy; Scope 3 covers other emissions across the value chain.
Straits Times Index (STI) constituents report Scope 3 from FY2026. Other requirements phase in according to the specified listed-company tiers. Large non-listed companies—those meeting both annual revenue of at least S$1 billion and total assets of at least S$500 million—report from FY2030 unless exempt. Scope 3 remains voluntary until further notice for non-STI listed companies and large non-listed companies.
Check the current schedule and the entity's position before setting a project deadline. A supplier may also receive information requests from a customer, lender or parent company without having the same direct reporting obligation. Record which requirement or commercial request the workflow serves.
External assurance is a separate responsibility. ACRA's schedule places limited assurance for Scope 1 and 2 from FY2029 for listed companies and FY2032 for large non-listed companies. Preparing an evidence trail supports that work; it does not make the preparer an assurance provider or turn an internal check into an assurance opinion.
Give every number a history
An evidence register should connect the reported value to the document or system record from which it came. It also needs the transformation between those points: unit conversion, allocation, aggregation, estimation or application of an emissions factor.
Keep the original source and the extracted value separately. If a reviewer corrects an extraction, retain both the correction and its reason. If a source document changes, the reporting owner should be able to identify which calculations and disclosures need reconsideration.
For value-chain emissions, the GHG Protocol Corporate Value Chain Standard provides a framework across 15 upstream and downstream categories. Choosing the category, organisational boundary and appropriate methodology remains part of the accounting work. The standard's landing page also cautions that it is designed for comparisons of an individual company's emissions over time, rather than unqualified comparisons between companies.
That matters when software proposes a supplier ranking. Different boundaries, methods and reporting periods can make two apparently comparable figures misleading. More complete extraction will not correct an invalid comparison.
Separate extraction, checking and accounting judgement
Use AI for a bounded task whose output a reviewer can inspect. For example, it can propose the supplier name, reporting period, quantity and unit from a document, with the relevant page or passage attached. It can also identify missing fields or prepare a request for clarification.
Apply explicit checks where possible. Does the period fall within the reporting window? Has this document already been processed? Does the unit match the expected measure? Are totals consistent with the component records? A model may help flag anomalies, but a repeatable rule is preferable when the rule is known.
Accounting judgements need an assigned owner. This includes choosing the reporting boundary, deciding whether an estimate is acceptable, selecting a factor and approving a correction to a material figure. Record those decisions as part of the evidence, rather than leaving them in an email exchange that the reporting team cannot later find.
A worked evidence record
Consider an illustrative electricity bill for a facility. The bill states consumption of 12,000 kWh for May. An AI extraction proposes that quantity, the facility identifier and the billing period, with a link to the source page. The workflow below stops at validated consumption; it does not invent an emissions factor or claim to calculate reportable emissions.
| Evidence field | Illustrative entry | Why it matters |
|---|---|---|
| Source | May electricity bill, retained original and page reference | Lets the reviewer inspect the underlying record |
| Entity and facility | Reporting entity and Facility A, matched to the approved facility list | Prevents an out-of-scope site entering the total |
| Activity period | 1–31 May | Separates consumption period from invoice or payment date |
| Extracted quantity | 12,000 kWh | Keeps the physical unit explicit |
| Validation | Period, facility, unit and duplicate checks completed | Records which checks were performed |
| Methodology handoff | Validated activity data passed to the approved calculation process | Separates extraction from factor selection and accounting judgement |
| Review record | Named reviewer, decision date and any correction reason | Makes approval accountable and changes traceable |
Suppose a second file arrives with a different name but the same bill reference and period. Counting both would double the consumption even if every field had been extracted accurately. The process needs a duplicate decision based on the source record, followed by an auditable exclusion or correction.
A similar principle applies to supplier questionnaires. A response may look complete while referring to a group-wide figure rather than the purchased product, business unit or period the request requires. The reporting owner needs a clear route to ask for clarification or apply an approved estimation method.
For example, a fictional supplier response gives a group-wide annual emissions figure when the buyer requested information for purchases from one subsidiary in another reporting period. The correct output is a clarification request and an unresolved boundary/method decision. Correct extraction does not justify allocating the group figure by spend without an approved basis.
The standalone evidence register extends the record through factor value and units, method approval, calculation version, allocation and output. Its fictional examples include an amended bill superseding the original. Where the approved factor is missing, activity data can be validated while emissions output remains held. That explicit state prevents a blank field from becoming zero in a seemingly complete total.
Make uncertainty visible in the workflow
An empty field should not silently become a model-generated estimate. Give records explicit states such as awaiting evidence, extracted, under review, approved and superseded. Show the owner of each unresolved item and the effect on the reporting deadline.
When an estimate is permitted by the chosen methodology, record its basis and distinguish it from measured activity. Preserve the version of any factor or reference dataset used, along with units and the reason it applies. The calculation should be reproducible after a tool, model or external webpage changes.
Access controls also matter. Supplier records may include personal, commercial or contract information that is unnecessary for the reporting task. Define which fields enter the AI service, who can see the source documents, and how long intermediate material is retained.
Test whether the work becomes easier to review
Evaluate the process on correctly extracted fields, missed material errors, duplicate handling and reviewer time. Include poor scans, inconsistent units, missing periods and amended source records. Fluency in the generated narrative is a weak acceptance measure for an evidence process.
Ask a reporting reviewer to reproduce one disclosed line using only the register and linked records. Any step that requires finding the original preparer reveals missing evidence or an undocumented judgement. Fix those gaps before increasing the volume of automated preparation; that is the foundation for a faster reporting process that remains explainable when someone challenges the number.
Read our editorial principles for evidence, contribution and correction standards.