Detection got cheap. Demand did not.
The NVIDIA account published on 10 November 2026 reads as a technology story and is better read as a cost curve. Greyparrot, in the United Kingdom, mounts sensors above conveyors to identify each object’s material and brand; it reports over 300 systems in more than 20 countries and upwards of a trillion detections, cutting the time to characterise 15 tonnes of waste from 375 hours to six.
AIRECO runs a pilot line in Tainan where recognition accuracy “tops 95%”; its founder’s formulation is exact — “The finer you sort, the more you can recycle.” EverestLabs reports roughly 150 robots across North America and Australia; reverse.fashion grades used garments at 3 metres a second.
These are company figures, not audited results. Read as a trend, they say one thing an operator can act on: the marginal cost of knowing what a material is has collapsed.
Set that against the value of the loss. The Circularity Gap Report 2026, from Circle Economy with Deloitte, puts avoidable value lost to linear practice at €25.4 trillion a year (±€4.7 trillion) against global GDP of €82.6 trillion — almost 31%. The composition deserves a second read. The largest pathway is end-of-life waste at €10.0 trillion, of which €6.5 trillion is residual material and product value rather than disposal cost; then energy losses at €8.7 trillion; then consumption of fixed capital — buildings, infrastructure and machinery deteriorating — at €5.2 trillion. Processing losses, the category closest to what a plant manager calls scrap, come to €904 billion: real money, and under four per cent of the total.
So the waste that matters most financially is not accumulating at the machine. It sits in material nobody recovered and assets nobody used hard enough. By value-chain stage the report puts €10.1 trillion in the use phase and €10.0 trillion downstream against €5.3 trillion upstream — the two stages where no single function owns the number.
Evidence that automation alone does not close it arrived five weeks before the NVIDIA post. The MIT Sustainable Supply Chain Lab published its seventh State of Supply Chain Sustainability on 15 October 2026, on 1,810 responses across 91 countries. Among advanced and fully automated warehouses using AI, 78% reported higher energy use since implementation and 76% reported higher waste — while only 24% monitored those consequences in real time. Meanwhile 86% of AI adopters reported reduced material waste. Both hold at once if the reduction is measured at the task and the increase accrues to the system — what happens when nobody instruments the consequence. Lead author Sreedevi Rajagopalan put it in nine words: “What comes next is harder: converting commitment into accountability.”
The camera found the material. The chart of accounts has no line for it.
The Matching Stack
Socradata separates any attempt to recover value from a residual stream into three layers, bought in the order below — the reverse of the order in which they bind.
What the system can see: material, polymer, grade, defect, brand, condition — object by object, at belt speed. The market has solved this layer and keeps discounting it. It generates classifications at very fine granularity, and obliges no downstream system to accept one of them.
What the system can name and therefore transact: a material master record, a grade specification, a quality tolerance, a unit of measure, a valuation class. Without it the stream has no identity in the ERP, so it cannot be received, costed, planned, substituted or sold — only disposed of under a cost code. Cheap in licence terms, expensive organisationally: opening a grade commits quality, engineering and finance to a tolerance someone must sign.
What the system will actually consume, on a clock: a planned order, an approved substitution ratio in a bill of materials, a contracted offtake, a redeployment work order. This is the demand side of the match, and where the money in the Circularity Gap Report sits. A stream with detection and specification but no committed demand becomes inventory that ages, then waste with better documentation. It is the only layer no vendor can supply.
The test is one sentence: a recovery programme performs at the granularity of its narrowest layer, not its widest. Detection at object level, specification at category level and no commitment is a stack whose effective resolution is zero, however good the cameras. Hence an operator can install an excellent sorting line and book no recovery — and hence the reverse works too: a specification and an offtake agreement with crude manual sorting beats a world-class sorter with neither.
Argentina is the cleanest illustration available, because there the detection layer is not the variable. Ecoplas, in its Dossier de Economía Circular de los Plásticos 2026, reports that the volume of recovered plastics fell 11% year on year in 2025, while the recycling chain — over 150 companies — runs at close to 50% idle capacity, and estimates more than 40% of discarded material could return to industry as feedstock. Capacity sitting half-unused while supply falls is not a sorting problem. It is an unmatched market: no specification obligation, no committed demand. Argentina remains, on Ecoplas’s account, the only one of the region’s five largest economies without post-consumer packaging responsibility rules or recycled-content targets, while Chile and Colombia have mandatory targets. At least four extended-producer-responsibility bills held parliamentary status as of 23 October 2026, after nearly two decades of debate.
So what: a sorting line that identifies material your ERP cannot name produces a better-documented loss, not a recovery.
Three loops where supply was found and demand was never committed
Each pattern names the systems, the decision loop, the override path and the outcome. Outcomes are labelled illustrative where no public figure exists. The structural claim — that a stream without a specification and a demand element cannot be transacted — is checkable in any material master.
Argentine plastics conversion, Buenos Aires province — the grade that had no master record. Systems: SAP S/4HANA or equivalent material master and quality-inspection records, an MES carrying scrap and yield codes, WMS putaway of recovered stock, and the province’s RINE register for residuos industriales no especiales, which since September 2026 requires generators, transporters and operators to record what is produced, in what volume and composition, with a verifiable destination. Loop: disposition of post-industrial trim, purge and off-spec resin at the moment the line classifies it. The default is a scrap cost code, because the recovered grade has no material number, tolerance band or valuation class — so even correctly identified material cannot be received into stock, let alone planned. Override: the plant quality manager may release recovered grade into a named non-critical SKU up to an approved ceiling, above which it escalates to the technical director before the batch is committed. Outcome, illustrative: the gain is not better sorting but a stream that can be received, valued and consumed.
Southern Cone grocery and consumer retail — the return that expired inside its own decision window. Systems: OMS returns and credit memo, WMS returns disposition, the ERP for valuation and credit authorisation, and the planning engine that must accept the recovered unit as supply. Loop: restock, rework, donate, liquidate or scrap, against a clock on which the item loses value daily. The detection analogue here is condition grading, usually free text typed by a receiving clerk — so disposition is decided on an unreliable classification, by whoever is standing there, with no demand element reserved for any of the five outcomes. The Circularity Gap Report puts €594.6 billion of its €650.7 billion food figure in loss before the consumer — that is, inside the operation. Override: the category manager holds release authority to a stated cut-off with a margin floor, escalating to the commercial director — named per disposition class, not per shipment. Outcome, illustrative: with structured grading and demand pre-reserved for restock and rework, the recoverable share stops being decided by queue position.
Andean mining and agricultural fleet, Chile and Peru — the asset that depreciated without being used. Systems: EAM or CMMS condition records, ERP fixed-asset and depreciation ledgers, and the capital plan that authorises replacement. Loop: refurbish, redeploy, re-rate for lighter duty, or dispose — decided per unit as it comes off a contract or campaign. This pattern is almost never found in a sustainability programme, and it is a €5.2 trillion pathway in the Circularity Gap Report, with premature obsolescence of long-lived assets a further €6.5 trillion. Condition data usually exists; a specification for what a re-rated unit may do, and a receiving site willing to plan against it, do not. The default is replacement, because replacement has a budget line and redeployment has a conversation. Override: the asset manager may block a replacement requisition where a redeployable unit meets the re-rated specification, escalating to the CFO — the only party able to move capital between sites. Outcome, illustrative: utilisation, not emissions, is the number that moves first.
What to settle before the sensors are specified
The artefact is a disposition authority register, not a circularity strategy, and it takes three moves. First, open the specification before the detection business case is approved. For each residual stream create the material master, grade tolerance, unit of measure and valuation class, and record which function signed the tolerance. A stream without a signed specification is not a candidate for automation, because its best case is a well-measured write-off. Second, name the demand element that will consume it, with a quantity and a date — a substitution ratio in a bill of materials, a contracted offtake, a redeployment work order; something a planning run can see. Where none exists, the honest output is to sell the stream or stop collecting it — a decision rather than a drift. Third, put the disposition on a clock and instrument the net effect. MIT’s 78% against 24% is the warning: measure the system, not the task, or the efficiency gain will be real and the net result worse.
So what: sustainability is what efficiency looks like when nothing is left in between. Specify the stream, name the buyer, then switch on the camera — KPIs before APIs.
Governance: the disposition authority register
One entry per residual stream class, enforced at the disposition credential in the ERP or WMS rather than in a policy document: a class with no signed specification and no named demand element blocks and escalates rather than defaulting to a scrap code. Each entry records the stream class, the material master and grade tolerance, the function that signed it, the approved substitution ratio and ceiling, the consuming demand element, the release authority above that ceiling, and the review date. Reviewed quarterly, because a ratio set at one virgin-resin price is wrong at the next. Owner: the COO or operations director — never the sustainability function, which cannot change a bill of materials, and never the vendor. Where a partner controls the grading model, that becomes a procurement clause: a classification only one party can inspect cannot govern a disposition both pay for.
KPIs: five numbers, each with a baseline
Commitment coverage — share of identified recovered tonnage holding both a signed specification and a named consuming demand element. Baseline 0%, because it has never been counted; target 100% of the top value decile before detection capex is released. Substitution rate — recovered or re-rated units consumed as a share of those specified; what separates recovery from relabelled storage. Disposition latency — median hours from classification to committed disposition; target under 24 hours for condition-sensitive classes, 72 for industrial residues. Net material yield — output accepted at specification over total input, at the system boundary rather than the workstation. Net energy and waste per unit after automation — the MIT counter-check, reported beside the task-level gain; if the task improves and this does not, the programme is moving the loss, not removing it.
Roadmap: twelve months from detection to offtake
Q1 — Count the streams, not the tonnes. List every residual, off-spec, returned and end-of-contract stream; mark which has a material master, a signed tolerance, a demand element. The empty columns are the programme. Q2 — Specify the top value decile and sign the tolerances. Quality, engineering and finance countersign; replace the scrap default with an escalation default at the disposition credential. Q3 — Commit demand. Substitution ratios into bills of materials, offtake where internal demand is absent, disposition latency on the operating review. Only now does detection capex earn its case. Q4 — Move it into contracts and registers. Specifications into supplier and customer agreements; traceability obligations mapped to the systems that carry them. For anyone shipping into the European Union, the Digital Product Passport registry has been operational since 20 July 2026 and the passport becomes mandatory for certain battery categories on 18 February 2027, with steel, textiles, aluminium and tyres following under ESPR delegated acts. Interoperability or it doesn’t scale applies to the material master as much as to the API.
The constraint moved from the belt to the material master.
For a decade the answer to why so little was recovered was that nobody could cheaply tell what anything was. That has expired. When a sensor reads a belt at four metres a second and replaces a manual sample of under one per cent, perception is no longer the binding constraint. What remains is older and less photogenic: the systems that run operations were built to plan, cost and commit what a business buys, and were never given an identity for what it recovers. A material master exists for purchased resin, not for resin recovered from its own purge. The loss is therefore largest exactly where ownership is weakest — ten trillion euros in the use phase, ten downstream, five upstream.
This is why “sustainability means efficiency” is a structural statement rather than a slogan. An efficient system connects supply to demand with nothing left in between; scrap and idle capacity are the residue of a failed match, paid for twice — once as material written off, once as a cycle that runs longer than it needed to. Read that way, the regional picture stops being a moral argument. More than one hundred and fifty Argentine recyclers at half capacity while recovered volume falls eleven per cent is a demand-signal failure any operations director would recognise in any other commodity. The governance evidence sits in the same place: MIT found sustainability built into daily decisions by fifty-seven per cent of companies where a cross-functional team exists, against thirteen per cent where it sits inside one function. A forty-four point gap is not a technology finding. POC theater here has a specific shape: a pilot that proves the camera works and never asks who will buy the output. From pilot to policy means writing the specification and the offtake into the operating model before the second camera is ordered.
Operational decision intelligence sits at this seam — between the model that classifies a material, the system of record that must name it to transact it, and the plan that commits someone to consume it. Vendors are now excellent at the first. The second and third are the operator’s own work, and decide whether recovery produces margin or a better-instrumented loss. Socradata transforms ERP, WMS and supply-chain data into predictive intelligence and governed operational decision systems.
Find out which of your recovered streams has a buyer
Every Wednesday, The Operational AI Dispatch translates one consequential AI signal into an operating model, a KPI set and an action plan for leaders running enterprise operations, ERP, WMS, supply chains and public systems.