AI Product Rankings Depend on the Criteria Teams Choose

An AI system can produce a clear winner without resolving the hardest question in the room: what should count as winning? In a product design review, a sealed device might rank above a repairable one because the assessment rewards manufacturing cost and reliability. The result may be internally consistent, yet still leave out repair costs, product lifetime and environmental effects.

That gap matters because a confident ranking can make a contested decision feel settled. For sustainability professionals, the task is to examine the choices behind the recommendation while colleagues may be eager to proceed, under deadline pressure or invested in the analysis. The useful question isn’t simply whether the model reached the right answer. It’s whether the business gave it the right problem to solve.

The priorities are part of the model

In the hypothetical design review, repairability affects the ranking only if the assessment accounts for relevant consequences after the product is sold. Even then, the calculation depends on assumptions: whether customers will repair the device, whether replacement parts will be available and whether a longer product life lowers environmental impacts across the full lifecycle.

Different teams may bring legitimate but competing priorities. Engineering may focus on reliability; procurement may need continuity of supply; service teams may consider what they can maintain. Sustainability may draw attention to effects beyond the immediate budget or planning period. The model can compare the criteria it receives, but its output can’t decide how much each function’s priorities should count.

That work happens when people define the problem. Samir Passi and Solon Barocas examined this process in their 2019 paper, Problem Formulation and Fairness. Based on six months of fieldwork with a corporate data-science team, they described how turning business objectives into modelled problems involves negotiation and discretion. Choices about target variables and proxies help determine what the system can represent.

For a product team, some requirements may need to remain firm constraints, while other priorities need an agreed basis for comparison. Making that distinction explicit gives colleagues something concrete to review. A well-presented explanation of an AI ranking can clarify how a result was produced; it can’t resolve a disagreement over which outcomes matter.

An explanation can win agreement without improving judgment

There’s a second risk once a recommendation is on the table: people may mistake understanding the system’s rationale for having good grounds to accept it. Research by Gagan Bansal and colleagues, presented at CHI 2021, tested AI-assisted decisions across three experimental tasks involving review sentiment and logical reasoning. Their study found that explanations increased acceptance of AI recommendations, including incorrect ones.

The researchers also found that explanations didn’t significantly improve human-AI team accuracy compared with simply showing the AI’s confidence. For decision-makers, that’s a reason to look beyond whether an output sounds persuasive. Ask what evidence supports its assumptions, what the system leaves out and what would make the recommendation change.

Separate broad interest from product evidence

A project from my own work shows why evidence needs to match the decision. We proposed using recycled material in selected product components to reduce reliance on virgin material. Customer research showed broad interest in recycled content, but it didn’t validate this particular change. Questions about technical suitability and product specifications also needed further work.

The proposal was postponed. The rationale for using recycled material was clear, but the available evidence couldn’t show that the material would meet the product’s requirements or that customers would respond positively to the specific change. A general preference didn’t answer a product-level question.

AI’s role in this example is hypothetical. An assessment could help organise the evidence, but it would need to keep the limits visible. Broad customer interest shouldn’t be presented as validation for a particular component. Technical suitability still needs to be established. A polished business case can make gaps easier to overlook without filling them.

For sustainability professionals, postponement can feel like a setback to an important commitment. Yet a colleague’s technical concern can point to the next question the team needs to answer. If the concern is dismissed as resistance, the evidence gap remains. Naming the specific uncertainty makes it possible to agree on a test, research question or other next step.

Make it safe to question the recommendation

The people in a product review may be carrying different responsibilities. An engineering lead could be accountable for failures in use; a procurement colleague might be managing a launch deadline; a sustainability professional may be concerned about impacts that the assessment doesn’t capture. Those possibilities aren’t evidence of anyone’s motives. They’re a prompt to ask what each person would need to know before supporting a different choice.

Questioning an assessment can also feel personal when colleagues have invested time or credibility in it. A broad request to “rethink the model” may sound like criticism of someone’s work. A focused question gives the group a clearer way to inspect the reasoning: “Would the ranking change if we included service costs over the expected product lifetime?”

Amy Edmondson’s 1999 study of 51 work teams in one manufacturing company found an association between psychological safety and team learning behaviour. She defined psychological safety as a shared belief that the team is safe for interpersonal risk-taking. In a decision meeting, that can mean giving people a clear invitation to identify assumptions that need checking, even when they support the option sustainability prefers.

Agree on what could change the decision

For the repairability comparison, the team could test how different repair uptake assumptions and service costs affect the ranking, drawing on validated calculations and suitable lifecycle evidence. A language model could help organise questions or summarise evidence supplied to it, while relevant specialists check the sources and methods.

If plausible assumptions produce different rankings, decision-makers should see that sensitivity before choosing. If key information is missing, the next step may be a technical test or targeted customer research. The team can agree on what evidence it needs, who is responsible for obtaining it and when the decision should be revisited.

Sustainability professionals can connect these questions across functions: where a cost lands, who carries a risk and which consequence disappears beyond the assessment’s boundaries. That scrutiny also needs to apply to the preferred sustainable option. An assessment should be open to evidence that challenges it.

Before approving an AI-supported recommendation, record the criteria, unresolved assumptions and person accountable for the decision. Once the product is in use, compare actual results with expectations, including reliability, service demand and environmental performance where relevant.

In the hypothetical review, the sealed design wins because manufacturing cost and reliability shape the assessment. Including repair costs and product lifetime could change the outcome. The team’s next decision is to establish whether those criteria represent what it wants the product to achieve—and what evidence it needs to choose with confidence.