When we first looked at Greyparrot, the pitch was positioned as sustainability technology. Computer vision on waste streams, classification of material types, compliance reporting for circular economy regulations. All of that is real — but it framed the company in a way that undersold the actual opportunity. The interesting thing about Greyparrot is not the sustainability angle. It's that waste stream data, if collected systematically and at the right points in a supply chain, is a leading indicator of operational performance that most supply chain software doesn't have access to.
Let me explain why that's a more interesting claim than the ESG compliance angle.
What waste data actually tells you
In a manufacturing or distribution context, waste generation is not random. It's a direct output of what's flowing through the facility. If you know what materials are coming out of a production line as waste — packaging types, product offcuts, damaged goods classifications — you know things about the production run that the ERP system may not tell you until the next day's reconciliation cycle. Packaging waste volumes correlate with production throughput. Damage rates in outbound handling correlate with operational quality at the pack station. Material composition of returns waste tells you something about the quality of goods coming back from the field.
This sounds obvious stated plainly. The reason it hasn't been systematically captured until recently is that waste streams have historically been invisible to software. Nobody was putting sensors on skips. The data existed in physical form — you could look at the skip and draw inferences — but it wasn't ingested into any operational intelligence layer. Computer vision at sufficient accuracy and throughput now makes this instrumentation economically viable for the first time.
The data flywheel in waste analytics
Greyparrot's technical approach is to deploy fixed camera rigs at material recovery facilities and waste sorting lines, and to run a continuously improved classification model against the video feed in real time. The model classifies materials at the item level — not just "plastic" but the specific resin type and likely source product category. That granularity is what creates the supply chain intelligence signal.
Here's the flywheel: more facilities instrumented means more labeled training data. Better training data improves the classification model. A more accurate classification model makes the data useful for more applications — from basic compliance reporting through to operational optimization loops. The classification model is already at the point where it outperforms manual sampling audits in accuracy, and runs continuously rather than periodically. That's the gap that creates a durable data advantage.
For supply chain operators specifically, the value proposition is becoming clearer as the data volume grows. A FMCG distributor who can see in near-real-time what percentage of their outbound packaging is ending up as single-use waste at the recovery facility downstream — and which product categories are contributing disproportionately — has an operational signal that informs both procurement decisions and packaging design choices. Neither of those decisions has historically had access to end-of-life data from the physical supply chain.
Regulatory pressure as tailwind, not thesis driver
The EU's packaging and packaging waste regulation (PPWR), the Extended Producer Responsibility frameworks being implemented across EU member states, and the Corporate Sustainability Reporting Directive (CSRD) collectively create a compliance demand for waste stream data that didn't exist at scale three years ago. Large manufacturers and distributors now have reporting obligations that require documented evidence of what's happening to their packaging materials downstream.
This regulatory tailwind is real and it accelerates adoption. We're not saying the compliance use case isn't valuable — it is. But we want to be clear about how we think about it: regulatory compliance is the wedge that gets the cameras deployed and the data flowing. The operational intelligence use case — using waste stream data as a real-time signal about supply chain performance — is the layer that creates defensible long-term value. Compliance software commoditizes over time as the regulations standardize and the reporting requirements become checkbox exercises. Operational intelligence that's embedded in a company's production optimization loop does not commoditize in the same way.
What made us comfortable backing this at seed
Our diligence on Greyparrot focused on three questions. First, is the computer vision model genuinely differentiated or is it a fine-tuned general-purpose vision model that a competitor could replicate quickly? The answer was that the classification taxonomy — which took two years of labeling work across diverse material recovery facilities — is a genuine moat. Not the model architecture, but the labeled dataset that the model is trained on. Second, what is the path from compliance reporting to operational intelligence? The team had a clear thesis about this, grounded in conversations with facility operators who were already asking for the operational data. Third, how does the company avoid being disintermediated by the facility operators themselves, who could theoretically deploy their own cameras once the use case is proven?
The third question led to the most interesting diligence thread. The answer is that the classification model and the cross-facility benchmarking — comparing your waste stream composition against an anonymized network of comparable facilities — requires a level of continuous model improvement and data aggregation that individual facilities cannot self-provide. The value is in the network, not in any single deployment. A single facility with cameras gets compliance data. A hundred facilities sharing anonymized waste stream composition data gets a real operational benchmark that tells you where you stand relative to comparable operations. Individual operators can't build that benchmark themselves.
The broader pattern: physical world data becoming supply chain signal
Greyparrot is part of a broader pattern we're tracking: the instrumentation of physical supply chain events that have historically been data dark. Waste streams are one example. We've also looked at computer vision for loading dock utilization, vibration sensors for cold chain equipment health, and acoustic monitoring for conveyor line anomaly detection. In all of these cases, the same logic applies: a physical event that has always occurred but was never captured digitally becomes a new input to operational intelligence systems when sensor cost and model accuracy reach the threshold that makes instrumentation economically viable.
The companies getting this right are the ones who understand that the sensor data itself is not the product — the intelligence derived from it is. Greyparrot's cameras are infrastructure for the classification model. The classification model is infrastructure for the operational benchmarking. The benchmarking is what creates customer lock-in. Getting the layer stack right, and not confusing the infrastructure for the product, is the difference between building a sensor hardware company (thin margins, high capex, fast commoditization) and building a data intelligence company (compounding data advantage, high switching cost, network effects as the dataset grows).