There is much excitement about the disruption that AI is likely to bring to traditional business workflows. This applies especially to supply chains, the interconnected systems that provide us with everything from daily staples to luxury goods. The move to AI is likely to present opportunities for improvement in areas from inventory planning and management to manufacturing, production and delivery.
So far, AI agents’ contribution to supply chain operations has been limited, and for companies not already on a path to digitalisation it can offer no shortcuts. While digitalisation adds value through algorithms, machine learning and data management, such tools have been around for years. In many cases, these are instruments that have been rebadged to tap into the current frenzy.
Many of the services marketed as agents are “base-level programs” that have existed for a while, says ManMohan Sodhi, a professor of supply chain management at Bayes Business School, London. “The heavy lifting is being done by the mathematical models already in use.”
The trade disruptions so far in the 21st century — the Covid-19 pandemic, protectionism and wars — have shown that resilience and adaptability are essential. New barriers are bound to appear and logistics networks will have to be ready. Meanwhile regulation over sustainability and transparency has increased, requiring companies to be more aware of provenance.
In this context, new forms of AI, including autonomous agents, are likely to add value to operations if deployed cohesively. While generative AI can boost productivity at the individual level and augment support functions, more sophisticated agents will help those companies that are prepared to restructure their supply chains.
Maria Jesús Saénz, the director of the supply chain transformation lab at the Massachusetts Institute of Technology, says AI means that supply chains can be automated and autonomous. “This is a vision for most companies [but] the reality for only a few of them – the most successful ones,” she says.
This Tech for Growth Forum report examines how and where new-generation AI can help companies manage supply chains against what is a demanding backdrop, and how businesses must consider a digital transformation if they are to benefit.
Old tech
Despite the hype, AI has been applied to supply chain operations for decades, helping with route planning, supply procurement, inventory management and exceptions identification (when events do not go to plan).
Goods routing is an example: avoiding traffic jams and delivering items to schedule have been handled by existing systems for years. Likewise, demand forecasting relies on older statistical models, while so-called agents, which book freight slots or manage disruptions, are based on rules engines and workflow automation. These are traditional tech models, even if they are advertised as being autonomous and able to reason.
In many cases it is only the marketing that has changed. Vendors might now refer to “AI demand sensing” and “autonomous replenishment” but often that technology has been in use for a long time.
New capability
While this suggests that the revolution is limited, evolution has taken place. Better data-handling and faster compute now allow enterprises to run continuous model updates. With real-time data and always-on optimisation, supply chains can respond to circumstances much more rapidly.
Danijel Lolic, the chief operations officer of Formic, an Illinois company that deploys fully managed robotics systems, says faster data processing can be transformative. For instance, to fulfil an order for a specific type of apple of a certain colour, models can run statistical analysis on normal distribution, meaning that sorting 1,000 or 10,000 apples can be based on the first 100.
That data might have been available previously but dealing with any change during the sorting process, for instance if a customer altered their demands, was time-consuming. It could take hours to stop lines and recalibrate. “It was manually intensive. It was a lot of almost gut feel or instinct,” Lolic says.
Now technology sorts in real time, he says. “There is this power that’s been unlocked in demand planning and scheduling and processing inputs . . . [We now have the] ability to make better decisions in the moment across a wider range of inputs and parameters. This enables someone who doesn’t have 30 years’ experience to make an as-good-if-not-better decision than someone . . . who’s been doing this for 30 years and who knows what to do because they worked with this retailer for x number of years.”
Large language models can add value as an overlay. A paper in the International Journal of Production Economics found that in a theoretical model, AI LLMs helped to improve supply chain agility and responsiveness thanks to automated data interpretation and real-time decision support.
Put another way, LLMs’ new interface has changed how operational data can be handled. Generative LLM-based AI can rapidly condense large volumes of cross-functional data into natural language summaries, enabling workers to act on more information than they could handle manually. Summaries of real-time data and events improve workflows and speed up decision-making, creating measurable operational benefits. Complex scenario analyses are now possible without extensive manual labour, which will help companies plan for and handle disruptions.
Next generation thinking
Companies at the leading edge of the transformation can find even more value in bespoke agents. Saénz of MIT says: “When we talk about deep transformation, this is execution of the supply chain in a completely different way. It is more automated, more autonomous, more self-learning, self-correcting, more resilient.”
This requires a different approach to AI. “Does that transformation need LLMs and GenAIs? Yes, of course. But it’s a completely different order of magnitude than [the AI] we use for our daily job. It’s like two different dimensions.”
Such change also demands a different mindset and a willingness to undertake a “long and tedious journey” during which companies will follow a “J curve”. Saénz says: “It’s difficult to make gains quickly. Why? Because you need to transform your processes and your organisation. You need to invest in technology… You need to monitor how quickly you are transforming and how well. You need also to bring your suppliers into the transformation.”
No technology can help if the premise is wrong. The first question a company should ask is “how do I envision my supply chain?”, Saénz says. Whether the goal is greater resilience or responsiveness or less friction, the vision must come first and the technology should be designed to achieve that. Since each company is different, they should train their own agents to achieve these goals using their discrete performance and learning indicators, as well as a well-structured and defined decision-making process. “No one system fits everyone.”
The most successful companies, Saénz says, began the transformation five years ago “and they continue transforming because there is no end”.
Compliance complications
AI can add value with functions that are ancillary to supply chain operations. Regulations are proliferating and adding to the complexity of doing business. Rules around sourcing, sustainability, carbon content and labour disclosures — along with AI-governance requirements and carbon border adjustments — make compliance increasingly complicated.
Companies are struggling to deal with this burden. In 2025, according to Industrial Arbitrage, the supply chain publication, businesses paid a total of $4.3bn in customs penalties and regulatory fines worldwide.
Things are about to become harder. Requirements around labour practices, sustainability and transparency are set to grow in 2026. From this month, vans in the EU will need to have tachographs fitted to record data which will include driver hours to ensure that minimum wage rules are met. In the US details that have been collected on lorry drivers for years must now include GPS tracking to avoid workaround violations. Meanwhile, the Uyghur Forced Labor Prevention Act has expanded its scope, increasing compliance requirements. Between its implementation in 2022 and April 2026, more than 42,000 shipments worth nearly $4bn have been detained. Further legislation will require supply-chain due diligence on ethical labour practices in geographies including the UK, France, California, Germany and Australia.
On top of these, the EU Corporate Sustainability Due Diligence Directive will come into force between 2027 and 2029. This will apply to EU companies and non-EU entities that have sales in Europe of more than €1.5bn and require them to identify, prevent and mitigate human rights and environmental harms in their entire supply chain.
The EU carbon border adjustment mechanism, which became operational in January, mandates that importers must disclose and certify the embedded carbon emissions of imports, while the EU Deforestation Regulation, which comes into effect in December, prohibits the sale or export of products linked to deforestation or forest degradation. The digital product passport, which attaches proof of origin and regulatory compliance to specific products, will become mandatory next year.
The cumulative effect of these new directives is that compliance has become a data-management problem that manual processes cannot realistically handle. Each regulation demands evidence, formats and timings; data sources may include supplier attestations, emissions factors, geolocation data and audit trails. Many of these extend beyond a company’s direct suppliers into networks where visibility is limited and data quality uneven.
As obligations increase, the burden shifts from interpreting rules to continuously gathering, validating and reporting information across thousands of transactions and multiple parties. AI can help to reduce this operational load. It can identify relevant regulations, extract required data from documents, map obligations to specific shipments or suppliers, flag discrepancies and assess risk exposure. AI-enabled software can also automate the creation of compliance documentation such as due-diligence reports, supplier questionnaires or emissions disclosures, based on data already held in enterprise systems.
There are numerous third-party LLMs — to name but three: Assent, Amber Road and Descartes Global Compliance — that are designed to take on these tasks. These free up people to focus on exceptions and interpretation rather than manual data processing.
Coping with shock
Supply chains must not only comply with regulations, they also have to be able to adapt to shocks. Resilience is a challenge not only for companies but for countries and communities that rely on imported goods.
Some countries, such as the US, are encouraging greater domestic procurement. The Biden administration required the federal government to prioritise domestic suppliers with the longer-term goal of beefing up its manufacturing base. While the government under President Donald Trump has rolled back sustainability targets, policies to encourage sourcing in America are high on the agenda.
Wars, trade barriers and shocks such as Covid-19 cannot generally be predicted but they can be prepared for.
Dynamic scenario planning and alternative sourcing strategies are all within the capabilities of models that can react more quickly to real-time situations. Modern AI can read and interpret news, announcements and even social media chatter to give insights into developing events. Faster access to more data can help supply chain managers move more quickly when exceptions arise. Information on port or warehouse congestion, for instance, can be valuable to companies that ship perishable goods, helping them to find options that can save cargoes and costs, if not rescue revenues.
Lolic, of Formic, recalls the flood of food orders that overwhelmed the distribution centres when Covid hit. While these days a system would still not predict Covid, it can now inform a shipper “that the truck you are about to ship out from your facility is going to a distribution centre that has no pallet storage space available . . . Because there are 87 trucks inbound to that distribution centre . . . carrying 26 pallets on the 53ft trailer each . . . They have this much capacity to store products and their churn is this . . .”. While data analysis would not have been swift enough to prevent a company from such disruption-related losses in 2020, today the system has that capability.
Formic’s insights indicate the power of data for identifying trends or events. With an overview of patterns and pressures affecting the 200 or more manufacturing facilities it serves, the company and those like it have access that can help customers respond to change.
Productivity gains
Regular LLMs can also add value at a more individual level. AI language engines provide employees with far more information about marketwide trends and can speed up tasks such as specific data mining and analysis, creating presentations to order.
They can also support administrative tasks such as reading and creating emails, distilling contracts and extracting clauses. One such program is Microsoft’s 365 Copilot, which can reduce time spent on rote tasks while speeding up execution and outcomes.
So-called exceptions, for instance when goods do not arrive as expected, can often lead to a chain of emails between people who are trying to come up with a plan to resolve the issue. “With AI you can train it on those emails and then agents can do those things faster and more consistently,” says Sodhi of Bayes Business School. Enterprises can buy in agents or train their own on communications or operational data which trigger predefined actions.
With so much individual excitement around the use of generative AI models, Sodhi says, it is tempting to believe they can be as useful for the entire supply chain. “We are extrapolating from our current personal experience all the way to the supply chain.”
The utility here is still nascent but one overlooked area for AI application, he says, is “the team aspect”. If AI is really that individually productive, the next step should be to make teams more collectively productive, an area in which Sodhi says he has so far seen only limited development.
Poor foundations
Despite the potential benefits of AI overlay, its deployment is still piecemeal, especially in operations. A Gartner survey in November of 140 senior supply-chain leaders found that only 17 per cent of respondents were implementing a transformational redesign of their processes and workflows, while the rest were either applying AI incrementally for specific uses or gradually scaling it into their processes. Gaps in data readiness, employee capabilities and fragmented vendor landscapes hampered progress on tech adoption and deployment near term.
AI solutions cannot be deployed unless the enterprise is in position for digital operation. At the basic level, cohesive, clean data is essential. Stocking and inventory systems cannot find supplies or products that are inconsistently labelled, for instance.
To then digitise processes, Saénz says that AI agents need clear performance indicators and decision pathways. “The process should be structured in a way that AI can identify the points of my decisions – the data that represents those decisions – so that AI can start learning. If you don’t have a clear, standardised process with good anchor points for your expected performance, it will be more difficult for AI to learn how to improve.”
A definitive figure is elusive but analysts suggest that only a third to two-thirds of supply chains are digitalising their processes, depending who is doing the measuring and what they are measuring. When it comes to automation, Lolic estimates that among SMEs the figure is as low as 10 per cent.
The initial work of digitalising a system presents opportunities that are often overlooked. Gartner says more than half of chief supply chain officers have no plan to get to digitalisation. Of those that do have an idea of the way ahead, a third have no aligned strategy under a single governance plan.
Having a joined-up approach is an advantage. Blue Yonder, the specialist in intelligent supply chains, says in its Supply Chain Compass 2026 report that the respondents to a survey, whom it dubs “the optimists”, have a less fragmented approach to tech adoption, giving them better visibility across the chain than those it terms “the less optimistic”, who have taken a piecemeal, siloed approach to deployment.
The report says that while 95 per cent of the 678 supply-chain leaders to whom researchers spoke were using machine learning and predictive AI, only a quarter had taken to generative AI, and only 8 per cent were using agentic AI.
Limitations on agentic application
It is important to understand the limitations of the application of AI, particularly LLMs, in supply chains. These include a lack of understanding of AI models, over-reliance on the tools, insufficient skills to benefit from them and inherent shortcomings in the models themselves, especially for complex problems. A paper from the International Journal of Production Economics says: “LLMs are generative, rely on unstructured data and introduce interpretability risks, creating distinct opportunities and challenges for supply-chain management”.
While the models can generalise across huge and diverse data sets, they differ from forecasting models or optimisation engines in that they are not task-specific. They might be able to reason across multimodal data and interpret contracts into natural language, but the risks are real: hallucinations and interpretability can create unreliable results. When the input is unstructured data there is little way of knowing whether an output is reliable.
LLMs are good for research, automation and decision-making support based on understanding patterns at scale but they cannot be relied upon for inventing strategies in unstructured environments. Their persuasive appeal lies in the easy accessibility of their outputs, in natural language.
People problems
The barriers to adoption that the researchers found most critical were related to people. Problems with interpretability and human-AI interaction were more significant even than data-related concerns. This is borne out by the real-world improvements shown in UPS’s Orion system. The redesign of its core technology was transformational but restructuring work processes, staff incentives and providing easy-to-understand explanations for its decisions were essential for the program’s success.
When deploying LLMs, it is critical that an enterprise has the right people to understand outputs and that those outputs are easy to understand. It is important to be aware that workers can be resistant to change, especially when a redesigned process may make their roles obsolete.
Saénz says: “If you are having humans change their process for automation in order to substitute themselves, this is a very perverse thing… They want to keep their salary and might boycott the AI. [They may] be algorithm averse.”
One way to overcome this might be to create an agent to monitor when someone is not facilitating the automation and advise management who can then clarify to the worker how the transformation benefits their career development. “This is why [transformation is] so complicated. There are multiple dimensions of change management and learning.”
Uncertainty and a lack of trust and clarity are the biggest hurdles to AI adoption, so even if AI monitoring seems Orwellian, ensuring that workers have clarity around the destination for the company and their relevance in new processes is crucial. Viewing AI as a capability-building tool rather than a cost-cutting tool is also important in nurturing attitudes.
Great expectations
The holy grail is that fully autonomous agents should manage the entire chain, carrying out decisions in the place of humans. Forecasting models that previously were run and acted upon weekly might then run continuously.
For most companies this is not yet viable, Sodhi says Always-on models could cause too much responsiveness as they constantly generate new decisions. Overactivity could actually worsen performance.
There is also limited confidence that AI can make good decisions as well as a limited understanding of how those decisions are made. “Errors can build on top of errors… the agent might react to errors and get outcomes that are further [from desirable].” Given the speed of operations, this makes it hard to intervene. Plus with agents operating with other agents in an opaque system whose logic is internal, “if things go wrong, no one would know what to do”.
In such a scenario, human oversight would remain essential, presenting another hurdle to agents operating 24/7. “Humans don’t work 24 hours a day. They are not going to stay on the keyboard and keep hitting “yes” to OK the agent’s decision to upgrade the forecast,” Sodhi says.
While this is a challenge for smaller companies, Saénz says it might be manageable by larger businesses with worldwide handovers and agents that are designed to manage certain occurrences, flagging only more significant problems.
Control issues
While companies can benefit from building their own agents, Sodhi cautions that there are risks with tools that rely on models based in the cloud, including a lack of control over pricing and opacity around how the models work. “You don’t have ownership, you don’t have firm pricing and you don’t know what it is doing to create the outputs, because none of it is your program. It’s not a simple program where you can say, ‘Please add these 10 numbers’. and it adds the 10 numbers and it stops. Even the creators of these LLMs don’t know what their software is doing… they don’t have these things in their control.”
Saénz adds that building an agent from scratch to suit specific needs and with clearly defined parameters can get around this. Should any technology prove to be expensive or, for instance, not offer significant benefit, it can be designed to find something more basic that fulfils the task.
Unrealistic goals
Do not expect AI to fix all your problems, especially if you do not know what they are, or you have inadequate systems. AI cannot run operations in systems that are not already digitalised, and it cannot (yet) make decisions beyond the specific, constrained tasks each model is designed for. AI also cannot solve problems that an enterprise has still to identify: deploying a technology in search of a problem seldom brings a benefit.
There are immense gains to be had from new technology but companies must first identify their goals and realise that agentic solutions are not a cure-all — and also be aware that many are simply polished-up old tech under a new name.
As with any new technology, before signing up, an enterprise must understand what lies beneath the tool and what it is trying to achieve with a deployment.
