Agentic AI in Supply Chain: from insights to autonomous planning

AI is used by almost all supply chain experts. It makes forecasts, identifies any deviations and delivers a conclusion in understandable form. But here the matter stops because the decision-making is left for the planner who should interpret the notification and decide on further actions.

Here comes the agentive AI in the process of supply chain planning. Unlike regular AI which just provides predictions or recommendations, the AI agents can not only observe the situation but also make decisions within the framework specified by your organisation.

In this article, we will talk about what this means in practice, where AI-based agents can be used to improve your supply chain management, what makes them different from existing AI and automation tools and how to choose the right supplier.

What is agentic AI in supply chain?

The easiest way to understand how the process takes place is looking at it as a transition from predicting to recommending to acting:

• Predict: The models predict the demands or lead time in question through various approaches of statistical and machine learning nature.
• Recommend: The recommendation engines use the forecasts to provide recommendations regarding what action needs to be done, whether ordering something, moving some stocks, or changing some parameters.
• Act: The agents help in decision-making process such as placing orders, communicating with the vendors or doing whatever necessary if the situation exceeds their boundaries.

This is both an improved prediction and an artificial intelligence-powered interface on top of your ERP. Language skills and the ability to summarise are strengths of generative artificial intelligence. But agentic artificial intelligence uses all those skills as well as a loop, which includes observing, deciding, acting, and learning. Such a loop is necessary for supply chain planning since the value of the decision drops fast. A late response to the suppliers’ delay will be more expensive than the same response on day one.

It is crucial to be pragmatic. The agentic AI system cannot be a switch that will pass your whole supply chain into the hands of the system. The steps towards independence are achieved through a process where tasks are gradually delegated and are dependent on logic.

The pragmatism is key here. An independent Agentic AI system should not be considered as an ON/OFF switch, which can pass all your supply chain operations into the system’s hands. Independence is attained via the process of delegating tasks and depending on logic.

How does the AI agent impact the process of supply chain planning?

Planning has always been a process of sensing, analysing, decision making and implementation. The degree of automation and continuous execution of the process irrespective of the involvement of the planner is the important distinction. Four major abilities define the shift.

Interpret the changes in the supply chain environment

In the course of regular activities, planners examine exceptions, either at certain intervals, such as daily or weekly or upon any occurrence of problems. The agent can analyse the demand signals, inventories, open orders, confirmation by suppliers and deviations from the lead time on a regular basis and determine changes in the situation. The most important aspect of the process is interpretation. A 10% increase in sales for one SKU may be just an outlier, caused by some promotion, while a trend is beginning.

Evaluating options and trade-offs

Logistics decisions do not necessarily follow an ‘either-or’ dichotomy. Quick delivery provides service but increases costs. Transfer of stock from one site to another helps one customer and risks losing the other. Early procurement ensures supply but lowers liquidity. What is needed of agents is their capability to make choices from among available options in terms of pre-defined objectives, such as minimum level of service, budget availability and limitations of suppliers.

Action recommendation or execution

Not all actions have to be automated. In good system design, it enables the firm to decide on whether to have the agent recommend an action or automatically take the action depending on the decision. Small and insignificant decisions can be done through automation whereas decisions of greater value should remain a recommendation only.

Escalate decisions that need human judgement

Escalation is an ability, not a weakness. Every time an agent encounters an out-of-scope situation, whether it is a high-level client requiring protection, issues with the supplier, or a trade-off requiring business acumen, it should escalate with details: What happened? What was its decision? What were the options that it considered? And what does it recommend? This will ensure that the planners are constantly aware of important decisions.

The below-given scenario is fictional and does not present any customer case. Let us consider the situation where the main supplier of the company notifies about three-week delay in delivery of the certain quick-flowing range of products. The agent will:

1. Recognise the change in the delivery date and understand the orders and customers that will be affected by that.
2. Examine the inventory and the forecast of demand to recognise the period in which the organisation will face the risk of being out of stock due to low service level.
3. Find out the cost of different alternatives such as partial delivery, shipment from other stores, different supplier, or even air freight.
4. Take action if the cost is acceptable or escalate the issue if not.

The important thing is not to perform tasks that cannot be performed by humans, but to ensure that the right course of action is performed after recognising the signal without waiting for anyone to identify the problem or take any action.

Where can we use AI agents to help the supply chains?

The best use case scenarios exist wherever decision-making happens regularly, in a data-rich environment bonded by specific rules.

Demand & Supply Planning

Accuracy and bias in forecasts can be determined, changes in demand trends can be identified, and re-forecasting can be initiated. Agents can reconcile demand and supply needs and make recommendations prior to the planning meetings such that demand planning sessions would be backed up by analysis and not simply figures.

Optimisation inventory

Inventory is the first place where one could seek proof because of the ability to measure the goals. It’s possible to optimise service and investment in inventory at the product level. Safety stocks, reorder points and order quantities may be set depending on current lead times and demand volatility, rather than on parameters established a few months ago. Further reading about inventory optimisation is available here.

Supply and procurement

The agent will be able to manage the suppliers’ confirmations, seek the missing details and also synchronise the order timing and quantity with the supplier’s schedule and order volume needs.

Managing risks and disruptions in supply chains

Supplier reliability, inventory safety stock levels, and demand provide an approach that focuses on the risks that will impact their ability to serve customers and those risks that may be overlooked. This can be important when dealing with the flood of alerts.

Agentic AI vs. traditional AI and automation

The terms are frequently confused, making it difficult to converse with vendors about the issue. A clear distinction helps.

A useful test: what happens after the insight?

With traditional AI and copilots, a person carries the decision forward. With rule-based automation, the system acts, but only on scenarios someone anticipated. Agentic AI aims to handle situations where the right response depends on weighing several factors, while staying inside rules you set.

This also means agentic AI doesn’t replace the layers beneath it. Forecasting models, inventory optimisation logic and rules still do essential work. The agent orchestrates them and acts on their output.

What does an agentic supply chain need to work effectively?

It is a fact that agents are as good as what they stand on. There are 4 foundations that are important to consider:

Trusted planning logic

Those who forecast poorly or calculate their safety stocks poorly will simply make bad decisions more quickly. Firstly, proven demand and inventory planning techniques. This has been the base of Slimstock’s own customer results, understandably, so it’s worth saying out loud: The examples below are based on the planning engine in Slim4 and not on results from agents driven execution.

In a time of over-doubling the turnover, Schrauwen managed to decrease their stock by over a quarter and increase their level of service to over 98%. Quaker Chemical Brazil reduced their stock levels by 20% in a year without compromising on service. Sabco in Australia and New Zealand decreased stockholding by approximately 15% while maintaining the availability figure. Sappi Europe says it has at least 98% service levels for over 3,000 SKUs. These are what the customers as well as vendor report, and will differ depending on every company.

Consistent and clean Master Data

Each calculation for planning is based on lead time, minimum order quantities, and details for the supplier and product attributes. When they are in error, all decisions down the stream are impacted. Hence, Slimstock sees data quality as a precondition for autonomy and that’s why they use the Agentic Data Management capability first on the data layer: validate product & supplier information, hunt missing information and fix issues at source.

With this method of data input, Slimstock can help reduce manual work by up to 80% and improve data accuracy by as much as 98%. These figures represent the potential benefits, but the actual results will depend on your organisation’s data environment, processes and systems. That’s why recommend evaluating the impact within your own business to understand what improvements you could achieve.

Defined guardrails

Decision rights must be clear for the agent, what he can do without approval, what he needs to approve, and what he can never do. All value limits, supplier restrictions and service priorities are here.

Integrating with your systems

Agents need to read from and write to your ERP, warehouse and supplier channels. When they can’t act, they are advisers in the systems in which the work takes place.

How to evaluate Agentic AI vendors for supply chain planning

The marketplace is saturated and the word agentic is used to describe anything from chatbots to real decision automation. These 5 areas help to distinguish between substance and labelling.

The ability to plan and knowledge of the field of study

Discuss planning issues the seller has overcome over the past years, not months. Is the system able to cope with seasonal and intermittent demand, multi echelon networks, shelf-life and supplier constraints? In many cases, generic AI platforms are not as deep and established vendor planning platforms may be further up the agentic journey. Search for a clear explanation that agents do not supersede the foundation of proven planning logic.

The integration of data and systems

Discover how the solution integrates with your ERP and other systems, how long this process usually takes and what it does with the data when it isn’t perfect. Inquire about how the vendor handles data quality, as it may be the biggest obstacle to success.

Independence, fences and people

Request to observe the settings for autonomy. Are there different levels of autonomy per decision type, product class or value threshold? Do there exist an approval workflow and is it easy for planners to override? If the vendor only positions itself as an independent entity, not in a gradual process, it is a red flag.

Explainability, governance and security

Planners will neither trust nor accept decisions that can’t be explained, and auditors will not accept them. Ask if there is a reason behind each action, if there is a full audit trail and where your data is processed and stored. Discuss access controls and how the vendor’s agents are tested prior to production.

Measurable business impact

Request customer names for whom measurable results have been achieved and an explanation of the process that led to the results. Beware of general statements which are not specific to a customer, a time frame and a baseline. Better yet, define success metrics before you begin (service level, inventory value, planner time spent on routine tasks or decision lead time) and test a pilot against those metrics.

What does the future of Agentic AI in supply chain look like?

There are three developments that are likely, but will vary in time between organisations.

The first is that autonomy will start to grow over time. Organisations will begin with data quality and regular refresh decisions then expand responsibility for agents as results develop trust. The planners’ job changes from exception processing to establishing goals, evaluating performance, and dealing with the difficult cases.

Secondly, agents will be more likely to work in multi-functional roles. Finance, sales and procurement are impacted by planning decisions. Integrating supply chain planning with commercial and financial planning enables agents to make more informed decisions on trade-offs, such as margin versus availability. This is where Slimstock’s collaboration with Vena, the link between supply chain and financial planning, goes.

Third, governance will be a differentiator. As more is added to the agents’ job, the ones who can demonstrate what the agent did, why the agent did it, and how the agent did it in a way that was legal will be the ones who win.

Slimstock believes that agentic AI should be given a chance based on actual results, rather than on theoretical optimisation, and that these results should be measurable. For most organisations the best thing to do is not to purchase agents. It’s to test your planning logic and master data for them.

If you’re trying to figure out where agents might fit in your planning process, discuss your supply chain and the decisions that take up a lot of your team’s time with our experts.