It is the difference is between visibility and action. And it is at this precise moment in time that AI in logistics operations has ceased to be an experiment and has become a standard aspect of risk, quality and cost management instead. According to a study carried out by BCG and Alpegi in January this year among 180 experts from logistics service providers (LSPs) and shippers, more than 40% of shippers are already taking their logistics service provider’s AI capabilities into account when choosing a partner. AI may not, as yet, be a requirement, but it is no longer simply a novelty factor.

In logistics, traditional OHS systems operate reactively. They analyse something that’s already happened; a collision, damaged goods, complaints, a delayed time slot... With this method, an organisation only finds out about a risk after incurring the loss. This creates what’s known as an information gap; in other words, the period between the emergence of a hazard and its disclosure in the report. Vision AI closes that gap.

Computer vision models use existing Internet Protocol (IP)/infrared (IR) camera networks, analysing images in real time and applying an interpretation layer; who is where, how fast a forklift truck is moving, the distance between a pedestrian and a machine, whether the operator is wearing full PPE... The algorithms are capable of simultaneously classifying hundreds of objects, monitoring their motion vectors and calculating the precise distance between forklift trucks and members of staff. System decisions are taken in less that 300 milliseconds; in other words, faster than humans can react.

The scale of the problem addressed by Vision AI is measurable. According to the USA Department of Labor’s Occupational Safety and Health Administration (OSHA), forklifts are responsible for between 35 000 and 62 000 injuries across the country annually, with 36% of deadly accidents involving a forklift occurring with pedestrians. These are not incidents that can be prevented by training and signage alone. They represent a systemic area of risk that cannot be shut down by classic tools.

A common simplification crops up in discussions about AI in warehouses, claiming that it is one system, one camera and one alert. In practice, mature architecture consists of three layers that operate in parallel and provide mutual support. The BCG and Alpegi study has confirmed that picture at the priority market level. Shippers and LSPs agree that the three most vital areas of AI today are planning transport, forecasting and end-to-end visibility. This encompasses three fundamentally different ways in which AI is beginning to operate in the functional zones of a warehouse, at the gate and in the back office.

This is the most common starting point. Operating in real time, Vision AI assesses the movement trajectory of a forklift and a pedestrian, identifies the hazardous zone, verifies the completeness of the PPE and takes action before contact is made. The signal makes its way directly to the operator’s human-machine interface (HMI) and to a local warning system or programmable logic controller (PLC) that can halt a production line or gate. In the industry literature, this architecture is described as ‘collaborative intelligence’, a synergy between human and machine, where the system doesn’t replace the member of staff but provides them with information that simply wouldn’t be available at all without the AI layer.

Crucially, every event of this kind is recorded as a near miss and it is data that will be missing from a classic model, since no one reports a near miss, either because of pressure of time or on account of fear of the consequences. On the basis of these data, dynamic heat maps of traffic are created, showing the points in functional zones of the warehouse where violations most often occur. Adjusting signage or transport routes at those spots permanently reduces the risks, the operating costs and, in some instances, insurance premiums.

The second area is the gate and control tunnel. Traditional sea containers, or twenty-foot equivalent units (TEUs), trailers and tankers take eight to twelve minutes per unit to process more or less entirely on the basis of human perception. When traffic is heavy, this generates a bottleneck. At the same time, the evaluations themselves are subjective and susceptible to mistakes arising from haste or tough weather conditions. The absence of objective, hard documentation of technical status at the moment when responsibility is transferred is one of the most frequent causes of lengthy complaints disputes.

Convoluted neural networks such as You Only Look Once (YOLO), single-shot multibox detections (SSD) and YOLO-Neural Architecture Search (YOLO-NAS) are adapted from the automotive industry and analyse images at the gate in a split second. Optical character recognition (OCR) recognises the container and vehicle numbers. Vision AI segments any defects; corrosion, dents, scratches, leaks... The results are sent directly to the terminal, yard or warehouse operating system (TOS/YMS/WMS). When no damage is found, the gate opens automatically; in the event that damage or violations are detected, the vehicle is redirected for reworking. Irrefutable photographic documentation linked to the waybill is also created, as is a timestamp, all of which streamlines the claims process. With a management-by-exception model, manual processing is reserved solely for disputed cases, The remaining 90% of the traffic is processed automatically.

The third related area consists of vision tunnels on packing lines in distribution centres. Every shipment is compared with a digital template for label accuracy (OCR), packaging integrity and conformity of size and colour. If there are any discrepancies, the PLC isolates that shipment before it leaves the dock and the WMS generates a task for the operator. This has a direct impact on the overages, shortages and damages (OS&D) which, according to industry data, can consume 5-10% of logistics budgets. Each shipment held back in the dock means no return transport route, no added CO₂ emissions and no refund on a complaint that has dented the end client’s trust.

The third layer involves data that have never been combined to produce a single picture before. Vision AI transforms warehouse and yard images into a live map of occupancy, queues at loading docks and vehicle dwell time. A YMS grounded in vision data prioritises vehicles on its own, allocates docks on its own and sends communications to drivers on its own. In operational terms, this facilitates a move from static planning to the dynamic orchestration of time slots. The cameras installed on the gates and mobile vision systems installed on VNA trucks and/or drones can also carry out inventory tasks, updating statuses in the EMS with an accuracy exceeding 98%, without the need to interrupt warehouse operations.

For procurement/purchasing departments, large language models (LLMs) are integrated with contract life cycle management (CLM) and enterprise resource planning (ERP) systems. They analyse contracts from global carriers and create comparisons of responses to RFIs/RFPs in terms total cost of ownership (TCO) and bidders’ financial stability, while the autoredline function amends draft contracts in line with an organisation’s internal playbook. This oversight doesn’t end once a contract is signed. The system then analyses deviations from KPIs and from regulatory and market reports, flagging ESG risks at subcontractors and/or financial problems at key partners. Combining an LLM with an organisation’s knowledge graph reduces the risk of hallucinations, which is identified in the literature (Richey et al., Journal of Business Logistics, 2023) as one of the main operational risks of LLMs within supply chains. It also provides full explainability for every recommendation.

Benchmark data from deployments at leaders in the sector paint a fairly consistent picture of what happens in the first months after model calibration. However, they are best viewed as a reference point and not a promise, since concrete results will always depend on the specific characteristics of a facility, its operational culture and the quality of the deployment process.,

•   The number of near miss events falls by some 30-50% in the first three to six months after model calibration.

•   Compliance with PPE requirements rises to above 95% after eight to twelve weeks of continuous monitoring and team training.

•   Response time is reduced, being faster by 15% to 25% from the moment of threat detection to a fully secured incident site.

•   Effective detection of people in critical zones reaches a recall level of above 95%.

•   Automation of up to 90% of terminal traffic at gates fitted with Vision AI. Manual processing is only required in disputed cases.

The economic dimension of those changes is not only fewer accidents. It also encompasses lower costs for post-accident downtime, a stronger negotiating position with insurers and a tangible reduction in costly returns logistics. The BCG and Alpegi study shows that almost 80% of shippers and LSPs identify operational effectiveness and cost reduction as their primary reasons for turning to AI, well ahead of strategic motives and competitive pressure.

In 2024, Gartner forecast that, by 2027, half the companies running warehouse operations will be using Vision AI systems to replace traditional cycle count processes based on manual scanning, In December 2023, 506 supply-chain professionals were surveyed; 20% of them had already deployed a system of that kind. Gartner’s latest forecasts, dating from April 2026, go further. By 2030, half of the new warehouse facilities in developed countries will be designed as human-optional. In other words, they will be designed to operate with minimal human presence.

A recent, more detailed picture is provided by the BCG and Alpegi study, where around 40% of the LSPs surveyed declared that they have AI deployed beyond the pilot stage, but only one in ten had integrated AI into their core operations across their entire organisation. Just 13% of them reported a measurable business effect. The figures are even lower for shippers. Almost 70% are still at the exploratory or pilot stage, with a mere 7% able to show concrete improvements in their logistics operations.

The geographic spread of deployments is also telling. The Asia-Pacific (APAC) region is in the lead as far as AI maturity is concerned, with 31% of LSPs reporting successful deployment in their core operations. The figure for North America is 14%, with Europe at just 6%. For LPSs, the consequences of this are straightforward. The time for asking ‘Is it worth implementing AI?’ has passed. The burning question now is this; what order of deployment makes the most business sense? And if someone has European horizons in mind, then the pace among competitors in Asia and the USA should be the argument that takes precedent in their decision.

The most surprising finding from the BCG & Alpega study relates not to the pace of deployment but to what’s holding it back. Approximately 40% of the respondents from LSPs and shippers alike identified ambiguous returns on investment and gaps in in-house skill sets as the main barriers. Cost and technical complexity came further down the list, particularly with major players. Three years ago, the conversation around AI was concerned with whether an organisation’s technology was ready or if it could afford deployment. Now, though, when the technology is available and the costs have fallen, the key question is different. Does the organisation know how to deploy it?

The regional picture is even more diverse. In the APAC region, 54% of the respondents identified a lack of skills as their primary obstacle. At 38%, trust and explainability are the main focus in North America as a result of expectations in terms of regulatory and governance oversight. A quarter of the respondents in Europe indicated organisational resistance to change as the chief barrier. These are three utterly different first steps in deployment. They also represent three different skill sets that an organisation’s technological partner has to provide.

Model quality is rarely a bottleneck nowadays. Instead, the bottleneck is often the foundation that the model is endeavouring to operate on.

•   Enterprise-class architecture: safety systems have to operate 24/7/365. If the infrastructure cannot guarantee this, critical decisions ranging from collision detection to signalling the OLC become meaningless.

•   Edge AI and privacy: image analysis carried out locally on an edge device means that raw footage doesn’t leave the facility. Only anonymised metadata reaches the cloud. This approach, known as ‘privacy by design’, really does facilitate compliance with the GDPR and the majority of European clients have now made it a condition for acceptance.

•   Integration with the operational level: Vision AI only generates value when its output feeds into systems where someone makes decisions; business intelligence (BI), WMS, YMS, TOS, ERP... Detection without integration is just an expensive recorder of events. In the BCG & Alpega study, around 60% of LSPs identified AI integration as an investment priority, placing it ahead of technological partnerships and finding talented staff.

•   Machine learning operations (MLOps) and continuous learning: lighting, rack layout, new kinds of damage, seasonal inventory rotations... a warehouse is an environment that changes every day. Without a retraining pipeline, detection efficacy drops rapidly.

•   The AI safety layer: a filter between the system’s logic and the staff in the functional zones of the warehouse, this layer analyses scenarios and only generates instructions after that. This reduces the risk of an incorrect message reaching the relevant zone and triggering decision-making chaos.

Sensible Vision AI deployments in logistics are a three-step process that moves from pilot to scale and not the other way round. This is also consistent with a BCG recommendation whereby AI success today is viewed more as a challenge in terms of the operating model than the choice of tools.

•   Discovery: involving the precise mapping of processes, identification of critical risk points and auditing of the available camera and network infrastructure, this is the stage that most often gives rise to the first savings. Numerous organisations already have more resources than they think.

•   PoC and pilot: deployment in one area that plays a major business role, such as the entry gate with the highest volume of traffic, the corridor with the highest number of reported near misses, the packing tunnel with the highest turnover... Model validation takes place under real conditions rather than in a test environment.

•   Rollout and optimisation: scaling up to the entire facility and full integration with the IT ecosystem. Implementing predictive analytics within the EHS system, continuous model training and traffic heat maps as a tool for reorganising the layout of the storage zones. At this stage, the technology ceases to be a design and becomes part of the process.

•   Treating Vision AI as a substitute for OHS: Vision AI supplements and reinforces existing procedures. It doesn’t replace them. A safety culture doesn’t spring from a camera.

•   Conducting a pilot in a zone where nothing much ever happens: without real traffic and incidents, estimating business value is impossible and the pilot itself is interpreted as ‘unclear ROI’.

•   No integration with the WMS, YMS, TOS or ERP system: an event that is detected but doesn’t reach the decision-making system doesn’t change the process. It simply generates a report that nobody reads.

•   Leaving the GDPR out at the design stage: correcting the architecture after deployment is significantly more expensive than building privacy into the design from the outset.

•   An LLM in procurement without a knowledge graph: the flexibility of a language model alone, without hard links to ERP/CLM data, increases the risks of hallucinations in high-stakes decisions. The problem is described in deployment practice and the academic literature.

•   Deployment without a reskilling plan: the BCG & Alpega study demonstrates that around 50% of LSPs anticipate crucial retraining needs for their teams. If this aspect is omitted, then AI ends up alongside the process rather than within it.

At MakoLab, we don’t see AI in logistics as a standalone product that can be plugged in to a warehouse. We treat it as a layer of collaboration between people and system; in other words, collaborative intelligence, where technology takes on the parts that people cannot do faster or more accurately. Our work structure consists of three mutually complementary pillars.

1. Deep expertise: the proprietary damage detection standards we transfer from the automotive industry to road and intermodal transport provide levels of precision unknown to logistics until recently. Our MakoLab Automotive Damage Detection solution is grounded in YOLO, SSD and YOLO-NAS convolutional neural networks adapted from the stringently regulated automotive sector.

2. Full-stack integration: we connect Vision AI with the BI data layer, procurement-support models and ERP, WMS, YMS and TOS operational systems. Detection is merely a signal. The value springs from the decision that follows.

3.Enterprise-class security: we leverage on-premises edge AI, anonymised metadata, full GDPR compliance and, for procurement, an isolated cloud environment or on-premises installation, providing full privacy for your commercial data and 100% source traceability for every decision.

We are not involved in selling cameras or models. Our role is to help organisations make the transition from vision to action and ensure that what awaits them at the end of their journey is not a display board with a report pinned to it, but a more secure, more effective and more predictable process.

Sources

1. BCG & Alpega (January 2026): AI Is Already Moving the Logistics Industry Forward; https://www.bcg.com/publications/2026/ai-is-already-moving-the-logistics-industry-forward.

2. Gartner (2024): Half of Companies With Warehouse Operations Will Leverage AI-Enabled Vision Systems by 2027; https://www.gartner.com/en/newsroom/press-releases/2024-06-12-gartner-predicts-half-of-companies-with-warehouse-operations-will-leverage-ai-enabled-vision-systems-by-2027.

3. Gartner (April 2026): Half of New Warehouses Built in Developed Markets Will Be Human-Optional Facilities by 2030; https://www.gartner.com/en/newsroom/2026-04-13-gartner-predicts-half-of-new-warehouses-built-in-developed-markets-will-be-human-optional-facilities-by-2030.

4. U.S. Department of Labor Occupational Safety and Health Administration (OSHA): Powered Industrial Trucks – Pedestrian Traffic; https://www.osha.gov/etools/powered-industrial-trucks/workplace/pedestrian-traffic.

5. OSHA.com: 5 Common Forklift Accidents and How to Prevent Them; https://www.osha.com/blog/5-most-common-forklift-accidents-and-how-to-prevent-them.

6. Richey, R. G. et al. (2023): Artificial intelligence in logistics and supply chain management: A primer and roadmap for research, in Journal of Business Logistics, 44(4), 532–549; https://onlinelibrary.wiley.com/doi/10.1111/jbl.12364.

7. SupplyChainBrain: How AI-Enabled Vision Systems Will Transform Yard and Warehouse Management; https://www.supplychainbrain.com/blogs/1-think-tank/post/43151-how-ai-enabled-vision-systems-will-transform-yard-and-warehouse-management.