How AI is used in Position Green's Carbon Management
For customers, auditors and partners.
1. Summary
Next-gen is the new generation of Position Green Carbon Management. This article explains where it uses AI, what each AI component does and does not do, and how you can trace any AI-assisted value back to its source.
One rule governs the design: AI proposes, code decides, a person approves. The architecture enforces it. It is not just a policy.
- AI never returns free text where a number is needed. Every agent is bound to a JSON schema at the API call. What comes back is a typed record, such as a classification, a candidate concept or a conversion factor. It is never prose that a parser has to interpret.
- AI never does arithmetic that ends up in a reported figure. When an agent produces something shaped like a quantity, it returns the inputs and platform code does the multiplication. The calculation engine, aggregation engine, unit conversion and factor ranking are ordinary code with unit tests.
- AI never picks the final emission factor. At most, an agent picks a concept, such as "parquet flooring". A ranking function then picks the actual factor, by region proximity, then specificity tier, then vintage year, then a fixed tie-break.
- Everything AI touched is labeled. Each record has a source field that says whether a value came from a registry, a predictive model, an AI agent or a user. Supplier and library records keep their history, so you can still read the previous value and how it was set after a change.
Where AI earns its place is the messy reading work. It reads a PDF invoice. It works out that "TRANSPORTKOSTNAD" is a freight fee, not a purchased good. It recognizes "Eik parkett 14mm" as oak parquet. Those are language problems. The carbon accounting itself stays arithmetic.
Four principles follow from the rule:
- A person stays in the loop. No AI output reaches a reported figure without passing a point where someone can accept, edit or reject it. Matches below the trusted confidence band go to review queues that show the proposal, its certainty and the source text. Every field of a generated decarbonization action is an editable form field. If you regenerate one step, your edits from earlier steps are kept.
- AI runs only when you ask. A user action triggers every customer-facing AI run. No background process re-classifies or re-values approved data on its own.
- Methodology lives in code, not in prompts. GHG Protocol scope and category placement, the specificity ladder, EN 15804 lifecycle handling and double-counting guards are all fixed library logic. Prompts are versioned database records. Only one version per agent can be published at a time, and a unique database index enforces that.
- Accuracy comes before coverage. Agents may answer "no match", and are told to. Concept assignment returns nothing rather than force a fit. The research agent reports that it found nothing rather than invent a factor. An empty answer goes to a person. A made-up answer would not.
2. Where AI is used
2.1 Reading invoices and documents
Invoice and document processing is where AI does the most work. When you upload a PDF, a chain of small agents processes it, each with one job, a strict output format and its own version. There is no single "read this invoice" prompt.
| Stage | Output |
|---|---|
| PDF OCR | Table-aware text for pages that are only images. PDFs with real text skip this step. |
| Product extraction | Supplier, invoice number, date, currency and document totals. Per line: quantity, unit, unit price, line total, VAT, source page and a confidence from 1 to 100. Plus a self-check that compares the sum of the extracted lines with the printed totals. |
| Product identity | What the line is: a normalized name, a candidate concept, the material type and the kind of line (product, service, fee, tax, financial flow or adjustment). |
| Quantity normalization | Conversion factors only: the effective unit, the value per item and the basis used. Platform code computes the actual quantity, not the model. |
Two behaviors matter for assurance. First, lines that are not goods stay out of activity data. Fee, tax and financial-flow lines are flagged to skip factor assignment, so a bank charge never becomes a purchased good. Second, estimated mass is kept apart. A mass guessed from a product description goes in a separate field marked basis: mass_estimate. Later steps treat it as a flagged estimate, never as a measured quantity.
Column mapping for spreadsheets and ERP data is not AI. It is rule-based code. So are currency normalization, incoterm handling, unit conversion and cleanup of organization numbers.
2.2 Classification
Suppliers. The platform tries several sources in order, and AI comes late: the official registry by organization number, then predictive sector classifiers (not generative AI), then third-party directories, then AI, then AI with web research. The row records which source won. So you can always answer the question "how do we know this supplier is in sector X?"
Supplier names. This is a privacy control that uses AI. Before a name becomes a supplier, an agent checks whether it is a real supplier, sensitive data, anonymized, or some other bad data. Employee names and private individuals in a purchase ledger are held back instead of slipping into a carbon inventory.
Concepts. The model does not choose among thousands of near-identical factors. The platform shows it a short list of concept cards. Each card has a name, a description, the units available and the number of variants. The agent picks one, or none, and gives a confidence and a written reason.
GHG scope and category. AI does not decide this. Placement follows the categories each library entry allows, your emission boundary, and explicit rules that block combinations that look plausible but are wrong (wood pellets under mobile combustion, for example). An AI suggestion cannot get past one of those rules.
2.3 Matching emission factors
This is the most important part of the article.
What AI does: it narrows the search. It reads "Eik parkett 14mm, 2,3 m²" and proposes the concept parquet flooring. A second agent then checks the product and factor pair on its own. That is a cheap guard against the most common failure: a confident match that is wrong.
What AI does not do: choose the number. Once the concept is set, a ranking function picks the factor within each lifecycle stage. It ranks by region proximity first (an exact region beats the country, the country beats a bloc, a bloc beats global), then by specificity tier (see section 4), then by vintage year (the most recent one not after the reporting year), then by a fixed tie-break. A lifecycle overlap guard then removes variants that would double-count a stage already covered, and records each removal.
The ranking function is plain, unit-tested code with no side effects. The same inputs give the same factor every time. No language model takes part. Weak matches are shown, not quietly accepted: no concept, a low confidence or a failed check sends the line to review.
2.4 Gaps and finding new factors
Gap handling in the inventory is rule-based. When no specific factor exists, the platform steps down the specificity ladder to a spend proxy and tags the figure with the tier it used. AI does not invent activity data. It never makes up a consumption figure, a spend amount or a supplier's reported emissions. Where source data is missing, you get a clearly labeled proxy or a clearly marked gap.
Spend-based sector factors are least precise for the long tail of purchased goods, where the library has no specific factor. From a spend-based line, you can ask the platform to look for a product-specific physical factor. The search runs in the background. When it finishes, you get a notification with the proposed factor, its source and its certainty, and you accept or deny it. You get the notification even when nothing was found, so a search with no result never looks like one still running. If nothing clears the threshold, the existing spend factor stays.
Certainty here is measured, not asserted. The score is built from signals anyone can check:
- Does the unit fit the product class?
- Does the cited URL resolve, and is its domain on the list of recognized factor databases?
- Is the figure actually on the page when it is fetched?
- How authoritative is the source?
- Do geography and year match the reporting context?
The model's own confidence counts for nothing in this score, and searches stay within the approved list of sources, not the open web. Accepting a result applies the factor to your company only. It does not change the shared library. Adding a factor to the shared library is a separate step that Position Green reviews, so one customer's accepted result never changes another customer's emissions.
2.5 Decarbonization planning
This is the second big area for AI. Agents propose scored reduction levers with written reasons and expand them into candidate actions. On request, about a dozen agents work together to turn a one-line idea into a modeled action. They cover the description, requirements, component quantities (with the reasoning saved), affected emission sources, remaining consumption, parameter sourcing and cost. Each run is saved as a record with its mode, status, requesting user, exact inputs and timestamps, and is locked so two runs cannot overwrite each other.
Parameters are sourced in a fixed order. One agent first tries to reuse your own historical parameters or a library parameter. Only if none fits does a research agent run, and its answer must carry a source URL and keep the search results it used. The scenario model itself is rule-based: once the parameters exist, the calculation engine computes the trajectory, the impact of each action and the totals. AI fills in the model. It does not run it.
AI is also used to summarize uploaded documents and to draft risk narratives in ESG due diligence.
3. What AI does not do
| Function | How it works instead |
|---|---|
| Calculating and adding up emissions | Rule-based engines. No emissions figure in the platform comes from a language model. |
| Choosing the emission factor (the number) | A ranking function: region proximity, then specificity tier, then vintage, then a fixed tie-break. |
| GHG scope and category | The categories the factor allows, limited by your emission boundary, with rules that block wrong combinations. |
| Column mapping on file and ERP import | Rule-based mapper. |
| Spotting anomalies, outliers, duplicates and variances | Thresholds, reconciliation against document totals and structural checks. Each can be explained to an auditor without reference to a model. |
| Emissions insights | Hotspots are found by ordinary totals and ranking, so the number behind a hotspot is the same number as in the inventory. |
| Report narrative and disclosure text | Written by your reporting team from platform data. |
The last row is a choice, not a gap. Disclosure text carries legal and assurance weight. A faster first draft is not worth the risk of an unsupported claim in a published report. Instead, the platform gives you the material for the narrative: the spread of specificity tiers, factor sources and vintages, the share of spend-based and activity-based emissions, and the number of unverified classifications. Your reporting team writes the words and owns them.
Arithmetic is kept away from models for the same reason. A model's arithmetic can't be reproduced or audited, and there is no reason to accept that for a sum a computer already gets right every time.
4. Methodology and calculation integrity
4.1 The specificity ladder
Every emission factor applied has a specificity tier. The tier is both the ranking key when a factor is chosen and the data-quality signal in reporting.
| Rank | Tier | Type |
|---|---|---|
| 1 | Batch primary | Activity-based, most specific |
| 2 | Supplier product average | Activity-based |
| 3 | Jurisdictional activity average | Activity-based |
| 4 | Activity average | Activity-based |
| 5 | Supplier product spend proxy | Spend proxy |
| 6 | Activity spend proxy | Spend proxy |
| 7 | Supplier revenue proxy | Spend proxy |
| 8 | Sector/region spend proxy | Spend proxy, least specific |
Rank 5 is where activity-based ends and spend-based begins. Every spend-based total uses that cut, so you can answer "how much of our inventory is spend-based?" at any level. Tracking the share of emissions in each tier, period over period, gives you a data-quality story you can defend, straight from the platform.
AI helps you move up the ladder, but indirectly. Spend-based estimates exist mostly because reading invoices line by line is expensive. AI-assisted invoice reading turns a spend line ("NOK 42,000 to a building supplier") into an activity line ("2.3 m² of oak parquet"). That moves the line from rank 8 to rank 3 or 4, at a cost per document that manual work cannot match. Factor discovery does the same for the long tail. But AI cannot raise a tier by saying so. The tier follows from the factor actually applied. If only spend data exists, the line stays a spend proxy, however confident an agent is.
4.2 Emission factor governance
Position Green curates the library. AI does not generate it. The library is a hierarchy whose lowest level is concepts. Each concept holds variants with figures, units, region, vintage year, lifecycle stage and specificity tier.
Every change to the library goes into a log that can only be added to, and every change can be reverted. A row's current status comes from the log, not from overwriting the row. Library and supplier records keep their history: a change writes a new version and closes the old one's validity period. So you can always rebuild a past calculation with the factor version that applied at the time.
AI's role in the library is small and internal. One agent adds plain, searchable names to rows to improve matching. It does not create, change or approve factor values. When AI suggests concept links during an import, a curator must confirm them. They are never applied automatically.
4.3 Audit trail
For any AI-assisted value, the stored records show:
- The source data: the uploaded document with its upload record, or the source transaction.
- Which AI ran: the agent, its version and the model that version uses.
- What it was given: a snapshot of the inputs and the versioned prompt template.
- What it proposed: the proposal with its confidence, reasoning and cited sources.
- What the user decided: the verified flag, whether the value was set by AI or by a user, and the previous version with its validity period.
- Which factor was applied and why: its identity, version, region, vintage, tier and the ranking that picked it.
- When it ran and what it cost.
If an agent's prompt changes, you can still tell outputs from before and after apart by the version recorded on them.
5. Reviewing AI output
5.1 Three different signals
Uncertainty level. How much the platform trusts a supplier match or classification.
| Level | Meaning | In review queue? |
|---|---|---|
| Verified | Confirmed by a user or an official registry | No |
| Trusted | High-confidence automatic match | No |
| Confident | Plausible but not confirmed | Yes |
| Dubious | Weak signal | Yes |
These levels apply to suppliers and classifications. Invoice line matches in Activity Capture use their own scale: Confident (85% and up), Likely (70–85%) and Possible (below 70%). Both scales use the word Confident, but they measure different things. See the Activity Capture guide.
Model-reported confidence. A score the model gives its own output. It is a self-assessment and it is not calibrated: 0.9 does not mean 90% of such matches are right. Use it to decide what to review first, not as proof that a match is correct.
Composed certainty. The score on a discovered factor (section 2.4). It is built from checks against the outside world, and the model's own confidence counts for nothing. It is a measurement, not an opinion, and you can question each part of it.
Two structural signals are worth more than any score the model reports: whether an invoice's extracted lines add up to its printed totals, and whether the independent checking agent passed the product and factor pair.
5.2 Working through the review
When you accept a suggestion, the row gets a verified flag but keeps AI as its source. So accepting a value and creating it stay distinguishable. When you edit, the platform writes a new version with you as the source and editor. When you regenerate, it creates a new run with its own input snapshot and keeps your earlier edits.
The review queue is a work list. It does not block anything. An unverified Confident match still applies its factor to your emissions. Review is built into the workflow, with no approval gate that stops data. If you need a separate maker and checker, add that as a process control on top of the queues: the platform records who verified what, and when. Work through the queue by emissions size, not by confidence score, and check each item against a source rather than just clearing the list.
For assurance, the rule of thumb is this. The platform records what the AI assumed. You record why that assumption is acceptable. The first happens automatically. The second does not. You need both to defend a figure.
6. Data governance and security
6.1 What data AI may process
| Feature | Data sent to a model |
|---|---|
| Invoice processing | The uploaded invoice PDF (the whole document) |
| Identity, quantity, concept assignment, validation | Extracted line text and quantities, plus candidate concept cards from Position Green's library |
| Emission factor discovery | The product description, its unit and quantity, and the reporting context |
| Supplier classification and validation | Supplier names, organization numbers, country codes |
| Decarbonization | Company context, emission sources in scope, the action description and existing parameters |
| Document summary and ESG due diligence | Uploaded document contents; supplier identifiers and public risk information |
Never sent: user credentials, authentication tokens, access-control data, billing data, or other customers' emissions results. Each prompt is built from the requesting customer's own data.
One caveat: invoice processing sends the whole document. If a PDF holds more than carbon accounting needs (personal names, bank details, contract terms in the footer), that is sent too. Where this matters, upload a structured extract instead of the full PDF, and replace personal names of suppliers with pseudonyms before export.
6.2 Providers and processing regions
| Provider | Models | Used for | Region |
|---|---|---|---|
| Microsoft Azure OpenAI (Position Green's own resource) | GPT-4o mini, GPT-4.1 nano/mini, GPT-5 nano/mini, GPT-5.1 | Classification, extraction, validation, generation | Sweden Central, Data Zone |
| Microsoft Azure OpenAI | text-embedding-3-small | Semantic search embeddings | Sweden Central |
| Microsoft Azure AI Foundry (Position Green's own tenant) | OpenAI-compatible deployments | Emission factor discovery | Norway East |
| Google Vertex AI | Gemini 2.5 Flash Lite / Flash / Pro, and search-grounded variants | Document and PDF processing, web research for decarbonization parameters | europe-west1 |
We reach both providers under Position Green's own cloud contracts, in European regions. No customer data goes to a consumer AI service or a shared public API. Web research runs on search-grounded Gemini, and its inputs are research questions about technologies and factors, not customer transactions. Which provider an agent uses is a setting, not code: each prompt version names its model. So the provider mix can be audited and changed per agent without a new release.
Chat and completion models run on Azure Data Zone capacity, which is Azure's EU-scoped deployment type. Microsoft's terms state that for a Data Zone deployment in a resource located in an EU member nation, "prompts and responses may be processed in that or any other European Union Member Nation," and that data at rest stays in the customer-designated geography. The Foundry endpoint used for factor discovery is in Norway East, which is in the EEA but not the EU. If you require EU-only processing, assess that endpoint separately.
6.3 No training on customer data
This is a contractual commitment from both providers.
Microsoft. Prompts, completions and embeddings "are NOT available to OpenAI or other providers of Models sold by Azure," "are NOT used by providers of Models sold by Azure to improve their models or services," and "are NOT used to train any generative AI foundation models without your permission or instruction." The models are stateless ("no prompts or completions are stored in the model") and run in Microsoft's own Azure environment, where they "do NOT interact with any services operated by providers of Models sold by Azure, for example, OpenAI (e.g. ChatGPT, or the OpenAI API)."
Google. The Google Cloud Service Specific Terms state: "Google will not use Customer Data to train or fine-tune any AI/ML models without Customer's prior permission or instruction." On retention, the same terms commit that "absent Customer's prior permission or instruction, Google will not store outside Customer's Account (i) Customer Data prompted to a Generative AI Service for longer than is reasonably necessary to create the Generated Output, or (ii) the Generated Output."
The one exception: abuse monitoring. Azure checks prompts and completions for harmful content and scores usage patterns. Content sampled for automatic review "is not stored by the abuse monitoring system or used to train the AI model or other systems." Content flagged as harmful, or as part of a possibly abusive pattern, may be stored separately. In some cases authorized Microsoft employees may see it, through secure workstations with just-in-time approval. For EEA deployments, Microsoft states those reviewers are located in the EEA. Microsoft also offers modified abuse monitoring, an option for approved customers that removes the storage and human review step entirely. Whether it is on can be checked directly on the Azure resource. Google's terms have no such retention exception.
6.4 Retention, credentials and access
Stored in the platform (customer database, hosted in the EU, separated per customer, under the standard retention policy): AI outputs, run details (agent version, tokens, cost, run time, requesting user, timestamps, workflow input) and version history.
Not stored: the raw prompt sent to the provider and the raw completion returned. What stays is the parsed, typed result and the run details. Because prompt templates are versioned, a prompt can be rebuilt from its template and the recorded inputs.
Credentials and access. No API key sits in source code, environment files or container images. Keys live in Azure Key Vault, and the backend fetches them at runtime with its managed identity. All model traffic uses TLS. Every AI run belongs to one company, and its prompts use that company's data only. AI outputs are ordinary platform records, with the same customer separation and role-based permissions as everything else: a user who cannot see a supplier cannot see its AI classification. Only Position Green back-office roles can manage agents and prompts. Customers cannot write or change prompts.
6.5 Customer control
AI comes bundled with the module it serves. Modules are switched on per company and enforced on the server, not just hidden in the interface. The stronger control is how AI is used: every customer-facing AI run starts with a user action. If you don't use AI-assisted invoice upload, no model processes your invoice data. A contractual AI opt-out is set up per customer by agreement with Position Green. It is not a product setting.
Your settings also limit what the platform accepts from AI. Your emission boundary decides which GHG categories are allowed, so a suggestion outside it cannot be placed. You can override a supplier's region, sector, GHG category or material category, and your choice wins over the global classification. The reporting year decides which factor vintage is picked. The region decides proximity ranking.
7. A worked example
Input: a two-page PDF invoice from a Norwegian building supplier, 14 lines, with page 2 scanned.
Reading the invoice. Page 2 is detected as image-only and read as a table. Supplier "Byggmakker AS", invoice 2024-4471, NOK, total excl. VAT 41,880. Self-check: the extracted lines add up to 41,880, which matches the printed total. Document confidence is 94. Line 11, from the scanned page, scores 61.
Identity. Line 3, "Eik parkett 14mm 2,3m²", becomes the concept parquet flooring, material wood, kind product. Line 13, "Frakt", is a fee and line 14, "Miljøgebyr", is a tax. Both skip factor assignment.
Quantity. Effective unit m², value per item 2.3, basis explicit_unit. Platform code computes 1 × 2.3 = 2.3 m².
Match and check. Matched to the parquet card at confidence 0.88, with the reason: "explicit 'parkett' plus oak species; m² denominator matches the invoice unit, avoiding a mass estimate." A second agent confirms the pair.
Factor choice (no AI). The Norwegian variant (region distance 0) beats the Nordic bloc (1) and the global average (3). The activity-average tier beats the spend proxy. The 2023 vintage is picked for 2024 reporting. Result: 17.9 kg CO₂e/m², modules A1–A3.
Calculation (no AI). 2.3 × 17.9 = 41.17 kg CO₂e, placed in GHG category 3.1 by the allowed categories and the boundary. No blocking rule applies.
Review. Line 11 shows up with confidence 61. You open page 2, correct the quantity from 4 to 6 and save. The value is stored as set by you, and the AI's original value stays in the history.
AI decided that line 3 is parquet and line 13 is freight. Code decided the quantity, the factor and the emissions.
8. Limitations
AI makes wrong suggestions, and will keep doing so. Expect these failures:
- Misread multi-column or scanned layouts, with lines merged or quantities taken from the wrong column.
- Concept matches that look right but are wrong.
- Errors in units and pack sizes.
- Wrong sector classification for suppliers with generic names, which then leads to the wrong spend-based factor.
- Decimal separator mix-ups in documents from different locales ("0,207" versus "0.207"). That is an error of 1,000 times, and it is exactly why reconciliation against document totals matters.
AI does not replace expert judgment. You still decide the organizational and operational boundary, the consolidation approach, which categories are material, whether a proxy is acceptable, whether a reduction assumption is realistic, and what gets published. The platform records decisions. It does not make them.
AI does not create verified supplier data. A factor that AI researched or found, with a source URL, is published information that the model found. It is not supplier-reported data. It does not equal a product carbon footprint from the supplier, a verified EPD or an assured GHG inventory, and it never gets the batch-primary tier.
AI does not guarantee compliance. The platform supports workflows aligned with the GHG Protocol and preparation for CSRD/ESRS reporting. Neither the platform nor any AI feature certifies compliance. The reporting organization is responsible for every published figure.
The biggest remaining risk is not technical. The built-in checks catch a good share of errors: reconciliation against totals, the independent checking agent, the blocking rules, source checks, the approved source list, and thresholds that send weak results to review. No check prevents an assumption nobody wrote down, or a verification ticked without looking. The platform records who verified what and when. Whether the verification really happened is a matter of process, and worth spot-checking.
9. Frequently asked questions
Does AI calculate my emissions?
No. No language model computes an emissions figure. Calculation is rule-based, unit-tested code. AI prepares typed inputs, such as classifications, extracted invoice lines and conversion factors, and the calculation engine does the arithmetic.
Can I review or override AI suggestions?
Yes, always. Low-confidence outputs go to review queues, and you can edit every generated record. An override is saved as a new version with your name on it, and the AI's original proposal stays in the history.
Is customer data shared with external AI providers?
Two providers process data, Microsoft Azure and Google Cloud, under Position Green's own infrastructure and contracts, in European regions. Both commit by contract not to use customer prompts and completions to train their models. Google also commits not to store prompts or outputs longer than needed to produce the response. The one place a provider may keep data is Azure's abuse monitoring, described in section 6.3 together with the option that removes it.
Can AI generate supplier-specific emission factors?
No. AI can find published information and record it with its source URL. That is researched public data, not data reported by the supplier, and it never gets the primary-data tier. An accepted factor also applies to your company only. Adding it to the shared library is a separate step that Position Green reviews, so one customer's accepted result never changes another customer's emissions.
How does AI support audits?
AI outputs are structured records, not free text. So every AI-influenced value carries its origin, confidence, reasoning and version history. An AI-assisted classification is often easier to trace than a manual one, because a manual one rarely records why the person chose it.
10. Appendix: where AI is used, at a glance
| Area | Agents | User review |
|---|---|---|
| Reading documents | PDF OCR; product extraction with self-check; product identity; quantity normalization | Yes: low-confidence lines and estimated quantities |
| Classification | Supplier classification; supplier name check; supplier alias matching; concept assignment | Yes: anything below the trusted band |
| Emission factors | Factor check; factor discovery; adding search names to the library (Position Green curation only) | Yes: you accept or deny each discovered factor |
| Decarbonization | Lever suggestions; action suggestions; action generation (about 12 agents working together) | Yes: every generated field can be edited |
| ESG due diligence | Supplier assessment (5 agents); risk summaries (3 agents); document summary | Yes: AI-generated content is marked in the interface |
| Search | Semantic search embeddings | No: retrieval only |
Every agent above records the version that produced its output, its token count, cost and run time, linked to your company.
Standards the platform's methodology refers to: GHG Protocol Corporate Standard; GHG Protocol Corporate Value Chain (Scope 3) Standard and Technical Guidance; GHG Protocol Scope 2 Guidance; EN 15804 lifecycle module definitions; ESRS E1; PCAF Global GHG Accounting and Reporting Standard; ISO 14064-1.
Provider terms quoted in section 6.3: Microsoft, Data, privacy, and security for Foundry Models sold by Azure, and Foundry Models sold by Azure abuse monitoring; Microsoft Products and Services Data Protection Addendum; Google Cloud Platform Service Specific Terms, §18 (Training Restriction) and §20 (Generative AI Services).
More detail is available from Position Green on request: emission factor library governance, specificity tier definitions, the supplier classification cascade, agent and prompt version control, data processing agreements, and a full worked audit trail for one reported figure.
Position Green · Carbon Management (Next-gen) · August 2026
Was this article helpful?
That’s Great!
Thank you for your feedback
Sorry! We couldn't be helpful
Thank you for your feedback
Feedback sent
We appreciate your effort and will try to fix the article