"The action is where the value is." Thomas Brooks of BHS said it at the 2026 MRF Operations Forum in Chicago, and it summed up what operators wanted from AI and automation: tools that change what happens on the next shift, not dashboards that describe the last one.
Material recovery facility operators at the MRF Operations Forum spent much of the week on the same few problems. Data is scattered across systems, maintenance is mostly reactive, sorting results vary widely, and labor is hard to find and keep. AI came up constantly, but the questions had become more pointed. Operators wanted to know what a technology would deliver in practice, in an industry where costly capital projects have not always met their promises.
Data that leads to a decision
MRFs already record a great deal: equipment status, downtime events, throughput, material composition, recovery rates, staffing and maintenance logs. Much of it sits in separate systems. A supervisor often has to spend time assembling it to work out what happened during a shift.
Attendees described a risk of "data fog," where adding more data makes the picture harder to read. The task they described is to pull the information together, show it to the right person at the right time, and help that person improve the next shift.
The practical advice was narrow. Pick one metric that regularly causes trouble, such as unplanned downtime, throughput variation or material leakage. Then ask whether it can be measured reliably, whether its cause is known, and what action will follow when it changes. If the answer to the last question is unclear, another dashboard is unlikely to help.
Supervisors first, with people still in charge
Breakout sessions pointed to supervisors as the most immediate beneficiaries of AI. A large share of a supervisor's day goes to reconstructing events on the floor: checking several systems, walking the facility and comparing notes with the team. Software can take over parts of that work, leaving more time for strategic decisions.
Attendees did not present this as a replacement for experienced operators. One put it this way: "You need people to keep a good pulse on the plant." The suggested starting point is the manual reporting that supervisors repeat every shift, especially tasks that draw on multiple sources, such as downtime reports and shift summaries. A person stays responsible for interpreting the results.
Maintenance comes before prediction
Jams were named as a major cause of MRF downtime, and a stopped machine affects the whole facility. Manufacturing plants have built preventative and predictive maintenance practices over decades. Recycling maintenance tends to be more reactive, shaped by harsh operating conditions and staffing limits.
The two approaches differ. Preventative maintenance follows a schedule, such as inspecting a component weekly or replacing a part after set operating hours. Predictive maintenance uses equipment and operational data to flag developing problems before a failure. In a MRF that could mean spotting recurring jams on one conveyor or identifying which equipment faults repeatedly cut throughput.
Detection is only part of the job. Operators also need people, parts and expertise to act on it. The advice was to track downtime by equipment, cause, duration and shift, look for patterns, and use them to set maintenance priorities before buying more advanced predictive systems.
Sorting as a statistical process
Bastian Küppers of STADLER Anlagenbau GmbH described sorting as "a statistics process." Material composition, contamination, volume and equipment performance all shift. Another attendee said, "The only thing consistent in a MRF is inconsistency."
For that reason a single snapshot says little. Operators want trends across shifts, days and material streams: whether recovery is improving, whether a commodity keeps escaping to the residue line, and whether a process change helped. Extended producer responsibility programs add pressure, because they bring new expectations for recovery, measurement and reporting. That raises the value of item-level visibility into what is being recycled. Computer vision and AI could supply it, although measurement accuracy is still being worked on.
The guidance was to compare performance over time, account for changing composition, and set a baseline before adjusting a process so the effect can be judged.
Trust, and what remains unproven
Trust was the dominant theme. Experienced employees are aging out, difficult roles are hard to keep filled, and operators want to hold or raise output with fewer people. Robotics and automation fit that need, but operators remember projects that promised high performance and delivered less, and they now ask for evidence first.
Glacier, a recycling robotics company that took part in the forum, said all of its robots in the field remain operational. That is a claim about availability, not results. The company said it also cares about whether customers recover more material, improve quality and reduce downtime, but the figures offered do not show those outcomes. No recovery rates, downtime reductions or customer numbers were given.
The week's discussion points to a gap between what the industry is asking for and what has been shown publicly. Operators want outcome data from ordinary production days, and the vendor evidence described does not yet supply it. An installed-and-running count is a weak substitute, and without a stated denominator or time period it cannot be compared across suppliers.
The maintenance discussion carries the same caution. Manufacturing has long experience with predictive maintenance, but nothing presented showed that the approach transfers cleanly to MRFs, where material is variable and staffing is thin. Operators were told to get downtime tracking right first, which implies that many sites lack the basic records predictive tools need. If so, the near-term gain is likelier to come from consistent data capture and automated shift reporting than from sophisticated models.
The EPR point is the least settled. Item-level measurement could become a compliance requirement, yet the industry itself acknowledged that accuracy is still being worked on. Baselines built on imperfect measurement could mislead as easily as they inform. The test for vendors will be whether they can show, over time and against a baseline, that a change in a MRF produced a measurable improvement.
