AI and the Circular Economy: turning product data into a practical project
Connecting product, material and maintenance information to build a useful, testable project.
A furniture manufacturer collects used chairs from a customer. To prepare them for a second life, its team needs to identify which upholstery to replace, which components to keep and which materials to separate. The information sits across technical specifications, orders, photographs and maintenance notes. Some records describe the original model; others reflect changes made during use.
In this illustrative scenario, AI could help read documents and connect information. A useful first output would be a verifiable record for each chair: what is known, what is missing and which intervention can be assessed. Building it requires establishing which object each source describes and whether it is current.
Our previous article explored how to choose a first AI use case for the Circular Economy. The next step is building an information base suited to the task. This article offers a practical approach for teams working with incomplete data.
Start with the activity, then identify the information
Asking what data is available can produce a long inventory. A more focused question helps: what information is needed to carry out this activity reliably?
Refurbishing a chair may require its product version, actual material composition, component condition and expected repair time. Finding an outlet for production waste may depend on composition, contaminants, available quantities and the receiving organisation’s specifications.
Make the connection between information and action explicit. Knowing the fibres in upholstery becomes useful when it helps identify a compatible treatment or discuss options with a recycler. Availability, logistics and acceptance of the material also need checking.
Four groups of data to connect
We suggest four groups for an initial review. This is a practical proposal that each project should adapt to its context.
Identity and composition. Product or batch code, version, components, materials and any replacements. A shared identifier connects sources without confusing similar objects.
Condition and use history. Inspections, photographs, maintenance, faults and modifications. Production records describe the original product; an up-to-date check helps establish what can be done with it now.
Available routes. Spare parts, repair capacity, potential users, recycler requirements, costs and logistical constraints. Technical possibilities need to be matched with workable destinations.
Intervention outcomes. Actual time and cost, replaced components, inspection results, final destination and, where available, the duration of subsequent use. These records help test whether the initial recommendations were useful.
In the furniture example, a photograph alone is unlikely to establish a fabric’s composition and treatments. Connecting it to the correct bill of materials and an inspection makes the information more usable. If the bill of materials is outdated, that limitation should remain visible.
Where AI can help
AI can be used to extract fields from documents, make maintenance notes searchable or suggest matches between descriptions and product codes. Vision models can support specific inspection tasks when evaluated under their intended operating conditions.
These applications need different controls. An assistant summarising a specification should allow users to find the original document and passage. A model flagging a defect needs comparison with a technical reference assessment. Failure to flag a problem does not prove that a component is safe.
Keep documented information, measurements and model estimates separate. If composition is missing, an AI-generated suggestion should be recorded as a hypothesis. Routing the material to a recovery process may require further checks or testing.
Include difficult cases in the test
An initial test should include the conditions the team actually encounters: missing labels, uncommon variants, photographs of different quality, old documents and unrecorded modifications.
The NIST AI Risk Management Framework 1.0 emphasises evaluating accuracy with realistic test sets representative of expected use, supported by a documented method. This principle matters when classification affects what happens to a product or material.
Treat errors separately. Rejecting a recoverable component can lose value; approving an unsuitable component can affect quality and safety. The process should define when human review is required and when assessment must stop because information is missing.
An example: does the bill of materials still describe the chair?
Suppose the original bill of materials specifies polyester upholstery, while a maintenance note records a later replacement without identifying the new fabric. An AI assistant could extract both entries and flag the conflict, linking each field to its source document.
The record should state that the current composition needs checking. The responsible person can look for the replacement order or arrange a check suited to the intended treatment. Once verified, they record the actual material and update the product record. If uncertainty remains, any route requiring known composition stays on hold.
AI’s contribution here is specific and testable: finding information, identifying inconsistencies and reducing reconstruction work. Technical confirmation remains a separate step. The same approach can apply to a replaced component, a surface treatment or a product version that differs from the recorded one.
A minimum record for the first test
For one product family, an initial record can include:
- product identifier and version;
- source and date of material information;
- observed condition and inspection method;
- missing or uncertain information;
- proposed route and required checks;
- person responsible for approval;
- completed intervention, cost and outcome.
The number of objects to assess depends on variability and the consequences of errors. The sample should support a credible comparison between the existing process and AI-assisted work. Recording only successful cases would distort the result.
Teams should also define information access, the terms governing the chosen services and responsibility for updates. Technical documents and photographs may contain confidential information. Tool selection needs to account for this too.
Measure results across the process
Useful measures may include time spent reconstructing a record, required corrections, cases put on hold and verification costs. Pair these with relevant circular outcomes: products actually returned to use, components recovered and accepted, and materials delivered to a verified destination.
Environmental benefits require an assessment appropriate to the situation. Transport, treatment, the duration of subsequent use and the alternative being replaced can affect the result. The quantity identified by a model alone does not demonstrate a reduction in environmental impact.
At the end of the test, the team should be able to answer three questions: is the information more reliable? Is the total workload, including verification, sustainable? Have the records enabled circular interventions that actually took place?
If the bottleneck remains a lack of spare parts or an available outlet, that issue needs attention before expanding the tool. If reconstructing information consumes much of the workload and AI measurably improves it, there is a stronger basis for extending the project.
Learn and apply the approach
The AI & Circular Economy Lab for Managers, developed with Federmanager Academy, provides an opportunity to explore the link between AI applications and circular opportunities.
For a specific business challenge, Tondo Lab can be a starting point to discuss the process, available information and a possible testing approach. Bringing one product family, a few documents and a concrete operational question makes the discussion more useful.
