Two people review a laptop and an exploded technical drawing among paper records, with abstract charts on monitors and an industrial site beyond the windows.
Illustrative procurement review, not a tested software interface, verified footprint or observed vendor result.
Diagram text description

A man and woman sit at a desk with a laptop, stacked files and paper charts. The woman holds an exploded technical drawing. Monitors show abstract green and blue charts, and industrial structures are visible through large windows. No readable software brand, source-record identifier, approval, precise emissions result or certification appears. The synthetic scene illustrates reviewing operational and product evidence; it does not establish a real customer, working integration, accurate calculation, achieved savings or assurance.

Choose carbon accounting software by testing the result your organisation needs to approve, not by counting dashboard features. Start with four connected questions: what result is required, what boundary defines it, which inputs support it, and can another person reproduce and review it? A polished demonstration is useful for learning the interface, but it is not evidence that your own data and controls will work.

Separate mandatory requirements from attractive extras. An annual corporate inventory, a product footprint, a disclosure submission and a target-tracking programme may share records while needing different calculation and approval workflows. Decide which must work first and who will own that outcome after implementation.

The practical shortlist test is a controlled evidence exercise: import representative records, correct a duplicate, change a factor version and reconstruct the exported result. The fictional protocol below gives concrete expected numbers and acceptance evidence. It proposes tests; it does not report that any vendor has passed them.

Select the right software category

Select the software category around the object being assessed. Corporate inventory follows the organisation's entities, operations and reporting period. Product assessment follows a defined product system and functional or declared unit. An ESG suite may collect broader disclosures without providing every calculation method needed for either task.

Primary needBoundary and organising objectDemonstration to request
Corporate inventoryEntities, operations, scopes and periodA reproducible site-to-group calculation with review controls
Product footprintProduct system, unit and lifecycle boundaryA versioned product model with allocation and supplier evidence
Broader ESG disclosureRequired indicators and reporting structureA traced output from approved source data to each required field

Table 1, shown one row at a time. Every field is retained.

Row 1

Primary need
Corporate inventory
Boundary and organising object
Entities, operations, scopes and period
Demonstration to request
A reproducible site-to-group calculation with review controls

Row 2

Primary need
Product footprint
Boundary and organising object
Product system, unit and lifecycle boundary
Demonstration to request
A versioned product model with allocation and supplier evidence

Row 3

Primary need
Broader ESG disclosure
Boundary and organising object
Required indicators and reporting structure
Demonstration to request
A traced output from approved source data to each required field

The categories can overlap, but a category label does not prove interoperability. Ask how a supplier record or product dataset transfers between workflows, including identifiers, units, versions and review status. A corporate total divided by units sold is not automatically an appropriate product footprint.

Where several tools remain necessary, test the handoff rather than buying the same advertised feature twice. Specify which system owns the source, calculation and approved output, and which one merely displays or receives it.

Turn reporting requirements into testable criteria

Write requirements as observable acceptance criteria. “Supports our reporting” is too broad. Name the output, included entities, reporting period, relevant scopes, calculation approach, factor policy, approvals and export format. Where spend data is used, specify currency, exchange-rate source and period. Separate those choices from activity-data unit conversion.

Make the first test follow the required report backwards to its inputs. For example, an emissions line must identify the facility and period, reconcile to its activity records, show the selected factor version and retain the approval that permits its use. Define what happens when evidence is missing or a reviewer rejects an estimate. An unresolved record should not quietly become zero.

Confirm the applicable standard edition and any reporting obligations with the responsible specialist. A provider's framework logo or a competitor's list of standards does not settle your organisation's legal applicability, assurance requirements or future rule changes. Software functionality and compliant application are different questions.

Classify mandatory criteria as pass, fail or unresolved before adding weighted preferences. Good visualisation should not compensate for an inability to export a reproducible calculation or restrict approvals to the proper role. This keeps the shortlist focused without pretending every desirable feature is equally important.

Test data ingestion and evidence provenance

Use representative messy records, not only a clean demonstration file. Include an invoice in an original unit, a corrected bill, a supplier spreadsheet with incomplete fields and a product revision. Ask the team to show the original evidence, parsed activity, conversion, factor selection and result together. Keep a human owner for each judgement.

Record identity needs more than a filename. A stable business key should distinguish tenant, supplier, site, meter or activity, reporting period and document revision where relevant. A file fingerprint can help detect an identical upload; it does not by itself establish whether two different-looking invoices describe the same consumption. Conversely, identical invoice numbers from different suppliers should not be merged blindly.

Test the exception route. Who confirms a proposed duplicate? Can a correction preserve the original values, actor, reason and time? Can a reviewer inspect the source without having permission to change an approved result? Check both normal access and a deliberately unauthorised approval attempt in a permitted test environment.

For integrations and AI-assisted extraction, retain the source and expose uncertain fields. Automation may reduce handling work, but it does not make an inferred unit, missing supplier response or ambiguous product match factual. An impressive import speed matters only if the resulting records remain inspectable and correctable.

Evaluate methodology and recalculation controls

A recalculation control should answer what changed, why, who authorised it and which outputs are affected. Freeze the earlier result before correcting an input or applying another factor version. Store the revised result alongside it, with enough information to reproduce both. “Latest factor” is not a sufficient version reference.

Distinguish data correction from a methodological or boundary change. Replacing an erroneous meter quantity should not silently change the factor too. Likewise, extending an organisational boundary needs a documented decision about comparable reporting and prior-period treatment, not an unexplained jump in a chart. EPA's inventory guidance places boundaries, data procedures, factor choice and inventory management within one controlled process; software should make those decisions visible.

The fictional protocol below illustrates two retained results for the same corrected activity: 1,400 kWh produces 560 kg CO2e with factor F-v1 and 490 kg with F-v2. They are alternative calculation versions, not two amounts to add. The 70 kg difference follows the changed coefficient; it is not evidence that the site physically reduced consumption.

Ask the reviewer to reconstruct the older result after the new version becomes current. If an update overwrites the old coefficient, activity revision or approval, a downloadable current total cannot recover the missing history.

The same fictional 1,400 kWh activity retains two calculation versions: 560 kg CO2e under F-v1 and 490 kg under F-v2. They are not added.
Fictional test pack P-17, May 2026. Illustration prepared November 2026. The two results use the same corrected activity and remain separate.
Diagram text description

One enclosing frame labelled 1,400 kWh contains two retained alternatives, not successive process stages or additive quantities. Factor F-v1 is 0.400 kg CO2e per kWh and gives 560 kg CO2e; F-v2 is 0.350 and gives 490 kg CO2e. These are fictional expected R2 and R3 results for test pack P-17, not observed vendor outcomes or actual grid coefficients. The input quantity is unchanged between these versions, so their 70 kg difference is a coefficient change, not physical energy savings. The acceptance test must preserve both factors, source revisions and calculation history.

Score implementation, ownership and total effort

Compare implementation and ongoing ownership across the same scope. Licence cost is only one component: data cleaning, connectors, user administration, calculation review, training, support and annual maintenance can remain substantial. Request written assumptions about what the vendor performs and what your team must supply.

WorkstreamNamed internal responsibilityEvidence for the comparison
Data preparation and integrationsData owner and system ownerRepresentative mapping effort and exception log
Calculation and reviewMethod owner and independent reviewerReproducible test output and approval history
Operation and supportProgramme owner and vendor contactService scope, response commitments and update process
Exit and portabilityData owner and procurement leadTested export package and documented restrictions

Table 2, shown one row at a time. Every field is retained.

Row 1

Workstream
Data preparation and integrations
Named internal responsibility
Data owner and system owner
Evidence for the comparison
Representative mapping effort and exception log

Row 2

Workstream
Calculation and review
Named internal responsibility
Method owner and independent reviewer
Evidence for the comparison
Reproducible test output and approval history

Row 3

Workstream
Operation and support
Named internal responsibility
Programme owner and vendor contact
Evidence for the comparison
Service scope, response commitments and update process

Row 4

Workstream
Exit and portability
Named internal responsibility
Data owner and procurement lead
Evidence for the comparison
Tested export package and documented restrictions

Security, retention, access controls and data location also require the appropriate technical and contractual review. A successful emissions calculation is not a security audit. Test user roles and agree who investigates failures before moving sensitive supplier or employee data into routine use.

Use the proof of concept to estimate your actual effort rather than inventing an industry-wide saving. Record what was measured, what remains a vendor estimate and what depends on changing your own process.

Run a controlled proof of concept

Fictional acceptance protocol: Alder Manufacturing is evaluating one site-electricity workflow. Test pack P-17 is for May 2026. All documents, people, factors and numbers here are invented. This is a proposed controlled test, not an observed vendor result or a real grid-factor recommendation. Data owner Mira prepares the pack; reviewer Leon checks outputs; only the designated approver can approve a revision.

The source register contains DOC-A-v1, invoice E-101 for meter M-01 at site S-01: 1 MWh, normalised to 1,000 kWh. DOC-B-v1, invoice E-102 for the same site and period, contributes 500 kWh of separate consumption. A second upload of DOC-A-v1 is an intentional duplicate, not another purchase. Later, corrected DOC-B-v2 replaces the 500 kWh value with 400 kWh. Preserve both documents and their relationship. Fictional factor F-v1 is 0.400 kg CO2e/kWh; F-v2 is 0.350 on the same assumed generation, period and gas-conversion basis.

Test actionExpected reproducible outputAcceptance evidence
Import the two distinct invoices1,000 + 500 = 1,500 kWh; R1 = 600 kg CO2e with F-v1Original units, conversion and DOC-A/B references
Introduce the duplicate A uploadTwo unique consumption records; current total stays 600 kgDuplicate identified, excluded copy retained and decision logged
Approve corrected invoice B-v21,000 + 400 = 1,400 kWh; R2 = 560 kg with F-v1B-v1 retained, correction reason, actor and approval
Apply F-v2 in a separate revisionR3 = 1,400 × 0.350 = 490 kg CO2eBoth factor versions, change reason and retained R1/R2
Export and independently rebuildLeon reproduces 600, 560 and 490 kg from their respective versionsActivity, evidence, coefficients, approvals and manifest remain linked

Table 3, shown one row at a time. Every field is retained.

Row 1

Test action
Import the two distinct invoices
Expected reproducible output
1,000 + 500 = 1,500 kWh; R1 = 600 kg CO2e with F-v1
Acceptance evidence
Original units, conversion and DOC-A/B references

Row 2

Test action
Introduce the duplicate A upload
Expected reproducible output
Two unique consumption records; current total stays 600 kg
Acceptance evidence
Duplicate identified, excluded copy retained and decision logged

Row 3

Test action
Approve corrected invoice B-v2
Expected reproducible output
1,000 + 400 = 1,400 kWh; R2 = 560 kg with F-v1
Acceptance evidence
B-v1 retained, correction reason, actor and approval

Row 4

Test action
Apply F-v2 in a separate revision
Expected reproducible output
R3 = 1,400 × 0.350 = 490 kg CO2e
Acceptance evidence
Both factor versions, change reason and retained R1/R2

Row 5

Test action
Export and independently rebuild
Expected reproducible output
Leon reproduces 600, 560 and 490 kg from their respective versions
Acceptance evidence
Activity, evidence, coefficients, approvals and manifest remain linked

Every test's observed status is Not tested. The table defines expected outcomes, not a completed acceptance log. For the duplicate, an uncorrected staging sum would be 2,500 × 0.400 = 1,000 kg; the correction must leave 600 kg from unique consumption and preserve the excluded upload. Whether the application quarantines it immediately or requires a reviewed correction, it must not present the duplicate-inclusive amount as the approved result. Do not erase the evidence to make the total look right.

Require an export with stable activity IDs, document IDs and available source files, original and normalised units, site/period, both invoice revisions, factor identifiers and permitted coefficient evidence, calculation versions R1–R3, actors, reasons, approvals and unresolved exceptions. A manifest should identify the actual exported files and their checksums. No source files or hashes were created for this fictional pack; those are acceptance requirements for the real exercise. Leon should reconstruct each result outside the vendor interface using the delivered evidence, not a screenshot or unexplained total.

Add a permissions test: Mira may propose the correction but cannot impersonate Leon's review or the final approver. Preserve the rejected action and confirm that no approved total changes. Record limitations in licensed factor access, retention or external reconstruction explicitly rather than assuming a CSV contains everything.

Open decision: demonstrate whether the proposed system can export historical calculation versions and the permitted source and factor evidence needed for independent reconstruction. Consequence: leave this mandatory portability criterion unresolved and do not approve selection until the capability is demonstrated or the accountable buyer explicitly changes the requirement and accepts the documented limitation.

Connect product evidence to the appropriate calculation

If product data is the bottleneck, separate evidence preparation from the calculation it will support. A versioned SKU and bill of materials can identify components, material quantities, supplier records and review status. Their references should survive transfer into the chosen product model or organisational workflow, with units and revision dates intact.

For example, a revised component quantity must point to its own source and product revision rather than silently replacing every historical record using that SKU. The receiving calculation still needs an appropriate boundary, allocation method and review. A complete-looking BOM is not by itself a finished product footprint, a corporate inventory or assurance.

Ask any proposed provider to demonstrate the specific preparation and handoff capabilities required. Keep unsupported capabilities out of the acceptance record. This is a neutral description of a product-evidence workflow, not a claim that a named platform already performs it or replaces corporate accounting, LCA expertise, standards or independent assurance.

Conclusion: buy the operating model, not the demo

Buy the operating model that survives your controlled test. Require the mandatory result, source trail, correction process, reviewer roles and export to work together on representative data. Keep proposed expectations distinct from observed passes and unresolved requirements. The same discipline should govern implementation changes after procurement.

At the decision meeting, review failed and untested cases alongside the successful demonstration. Assign an owner and consequence to every unresolved item, and put important delivery commitments and limitations into the agreed scope. Do not replace a missing capability test with confidence that a future connector or consultant will somehow solve it.

The best shortlist outcome may be a narrower initial deployment or a clearer division between tools and expert work. Choose it deliberately. A system is useful when your organisation can govern and reproduce its results repeatedly, not merely when the opening dashboard looks convincing.