From AI Hype to Working Carbon Accounting Prototype
Before
When we started, the company was operating in an EU market shaped by emissions-reporting requirements and corporate sustainability needs. The startup was building a carbon-accounting solution to help organizations work with incomplete emissions data.
At this stage, the company was still defining its core technical direction. There was no consensus on whether AI models could reliably infer missing emissions data from invoices and receipts and translate monetary patterns into estimated emissions.
Leadership (Founder, COO, and Product Directors) needed clarity: Which AI model(s) could be part of a robust carbon accounting product? Could AI meaningfully augment incomplete financial data? And if so, what would a working prototype look like?
Instead of answers, the company had uncertainty. Instead of a technical foundation, it had speculation about AI potential. Strategic decisions were on hold.
What We Did
We led a structured AI evaluation and rapid prototyping process to assess viability and technical direction, moving from hype to evidence.
Comparative AI Testing
We evaluated AI systems from OpenAI, Google, and Microsoft for understanding and extracting emissions-relevant data from semi-structured financial documents. We compared performance on:
- Inference quality and accuracy
- Consistency across document types
- Scalability and cost
- Integration readiness
This wasn't casual testing. It was structured, measurable evaluation against defined criteria.
Prototype Development
We designed and built a working prototype using the best-performing model (OpenAI GPT), with API-driven workflows for document ingestion and structured output. We tuned the pipeline for real invoice/receipt formats and emission estimation logic.
The prototype was functional, tested, and ready for integration into the product development roadmap.
Strategic Advisory
We assisted the COO with SQL and database structuring for early analytic use cases. We evaluated architectural options for dashboarding (including potential migration to Apache Superset). We guided prioritization of product features based on technical evidence rather than trend-driven decision-making.
This was not ad-hoc experimentation. It was a structured comparability framework with measurable criteria and a real prototype delivered.
After
Instead of AI hype and uncertainty, the company finished this phase with:
- A working AI prototype capable of extracting and structuring invoice data
- Evidence that one model performed best against the agreed evaluation criteria
- Defined API architecture ready for integration into product development
- A prioritized list of technical limitations and next steps
- A product roadmap that reflects real capability rather than theoretical AI promise
Leadership now had direction and confidence instead of speculation. They could see, tangibly, what was possible with AI, and what wasn't, at this stage of the product.
Instead of chasing every emerging model, the technical team focused on what worked. Instead of building on assumptions, product decisions were grounded in evidence. Instead of uncertainty, there was a validated path forward.
What This Enabled
For the first time, the startup could:
- Compare AI models side-by-side in a structured way, with measurable results
- Prototype emissions inference workflows with real data, not theoretical examples
- Align product vision with validated technical foundation instead of speculation
- Make strategic choices grounded in evidence rather than trend-chasing
The company moved from being caught between "should we use AI?" and "which AI should we use?" to a clear position: "This is the model we're building with, because we tested it against alternatives and it performs best for our specific use case."
The product roadmap could now reflect observed model performance and known limitations. The startup had moved from exploring AI as a capability to preparing to integrate AI into the product.
Executive Takeaway
This was a structured, product-aligned AI assessment and prototype engagement that delivered clarity, a working foundation, and a prioritized technical path forward. In six months, the company moved from uncertainty about AI capability to a tested prototype and architectural direction.
The engagement replaced assumptions with evidence for the next product decision.