MIT survey maps warehouse AI adoption and control gaps

MIT survey maps warehouse AI adoption and control gaps

Warehouse AI adoption is accelerating across surveyed international logistics operations. MIT CTL and Mecalux found that reported improvements frequently outnumber quantified gains, while respondents expect more autonomous decisions by 2028.


IN Brief:

  • A survey of 1,967 supply chain and warehousing professionals across 21 countries examined generative and agentic AI.
  • Of use cases described in detail, 87% of generative AI and 81% of agentic AI applications reported at least moderate improvement, but fewer recorded quantified significant gains.
  • Some 57% of respondents expect AI to make or execute more than one-quarter of warehouse operational decisions autonomously by 2028.

MIT’s Center for Transportation & Logistics (CTL) and warehouse systems supplier Mecalux surveyed 1,967 supply chain and warehousing professionals in 21 countries about the deployment and performance of generative and agentic artificial intelligence. Among respondents, 57% expect AI to make or execute more than a quarter of warehouse operational decisions autonomously by 2028. That is a forecast from the surveyed group; the study also examines how far current implementations have progressed and whether their benefits have been measured.

The study, titled The state of generative and agentic AI in warehouse operations, collected online responses between 14 August and 1 October 2026. It separates generative AI, which can produce instructions, explanations, and other content, from agentic systems designed to take actions in pursuit of defined tasks. Those categories overlap in practice, but the difference becomes material when a system moves beyond advising an operator and is permitted to change a warehouse workflow or issue an instruction to another application.

Among the professionals surveyed, 61% reported that generative AI was in large-scale use or fully integrated at their organisations, compared with 56% reporting the same maturity for agentic AI. Those are self-reported adoption levels within the survey sample, not a measurement of every warehouse worldwide. Implementation conditions also differ: inventory accuracy, shift coverage, software integration and customer service requirements influence how consistently an AI application can operate after a pilot.

Reported operational improvements were more common than quantified significant gains. Among use cases respondents described in detail, 87% of generative AI applications and 81% of agentic AI applications produced at least moderate improvement against their target metrics. The respective proportions associated with significant gains that had been quantified and documented were 39% and 36%. A perceived reduction in daily administrative work is a different measure from recorded changes in order cycle time, picking accuracy or stock availability.

Warehouse decision systems rely on a chain of information that can include order priority, stock availability, item locations, labour availability, equipment status, and outbound transport cut-offs. A recommendation engine can rank the next tasks or suggest the location from which to replenish a picking face. For it to execute that decision, the warehouse management system must accept the instruction, update the transaction record, and maintain an auditable account of what changed. The survey does not establish that every deployed AI product can complete that sequence.

Integration remains an operational constraint. Forty per cent of respondents identified connecting AI with legacy systems as the principal obstacle to scaling generative AI, while 54% said improved integration with warehouse management and enterprise resource planning systems would make the largest difference to adoption. An AI layer working with incomplete stock records or out-of-date order information can produce persuasive recommendations that are not executable. The dependency on accurate transactions and permissions is therefore as important as the sophistication of the model.

Expectations of autonomy were higher among respondents whose organisations reported full AI integration. In that group, 70% expected agentic systems to make or execute more than a quarter of warehouse decisions by 2028, against 57% across the complete sample. Integration may give respondents greater familiarity with the software’s capabilities, but the survey does not establish how many of those forecast decisions will eventually be approved for automatic execution.

Ninety per cent of surveyed organisations expected to increase AI spending over the next two to three years, with 29% forecasting an increase of more than 25%. The comparable figure expecting any spending growth in the previous year’s survey was 87%. Expansion budgets are therefore being considered while many applications still need reliable baselines, defined measurement periods and controls that distinguish an AI-assisted task from the work previously performed by people or established software.

In the workforce responses, 56% said AI had primarily shifted employees towards higher-value tasks, while 11% reported lower headcount and 9% higher headcount. Seventy-three per cent reported upskilling warehouse-floor staff, and 61% were creating new technical roles. These figures describe the organisations that answered the survey. They cannot isolate AI from other influences on staffing, including changes in throughput, mechanisation, operating hours or outsourcing arrangements.

As software gains permission to act on warehouse instructions, responsibility shifts from reviewing recommendations to controlling executable decisions. An incorrect inventory adjustment or order release can affect picking, dispatch and the accuracy of customer commitments. The system therefore needs clear approval limits, a transaction record and a procedure for reversing errors. Such boundaries are already incorporated into warehouse AI systems that separate analysis from authorised execution.

The researchers recommend establishing governance before extending deployments, treating data quality and system integration as essential groundwork, and observing results over 12 to 24 months before deciding whether to expand or withdraw an application. That gives operators time to compare the AI-assisted process with a defined baseline across normal and peak conditions. A well-performing pilot with one product range cannot, by itself, validate a warehouse-wide productivity claim.

The study’s 2028 forecast sets a measurable test for AI suppliers and warehouse operators: how many operational decisions are executed autonomously under defined controls, and what effect can be documented on stock accuracy, safety and service. The researchers’ proposed 12-to-24-month observation period would allow individual systems to be assessed across routine operations and trading peaks before their deployment is widened.


Stories for you


  • GXO plans wider US warehouse automation rollout

    GXO plans wider US warehouse automation rollout

    GXO plans warehouse automation deployments across selected United States facilities. The 2027 programme combines pallet robots, automated unloading, vision systems, and sortation, with initial results still to be measured.


  • SpecialtyRx adds robotic sorting to pharmacy fulfilment

    SpecialtyRx adds robotic sorting to pharmacy fulfilment

    SpecialtyRx is automating prescription sorting for American long-term care facilities. The Tompkins Robotics installation raises reported peak station capacity from about 1,000 to 1,400 prescriptions an hour.