Agentic AI in Supply Chain — What It Actually Does

Agentic AI in supply chain means software that takes actions on your behalf — raising a purchase order, chasing an unconfirmed delivery, reading a supplier’s reply, rescheduling a run when a material slips — rather than only producing a forecast or a dashboard for a person to act on. The distinction is not the model. It is whether the system is wired to do something and is judged on the outcome.

That distinction matters commercially, because most “AI in supply chain” for the last decade has meant better prediction, and prediction was rarely the bottleneck. The bottleneck is the hundred small follow-ups a week that nobody has time for.

The four levels

It helps to be precise about how much autonomy is actually on offer:

1. Analytics. The system shows you what happened. A dashboard of stock cover, service level, supplier performance. Every ERP has this.
2. Prediction. The system tells you what is likely — a demand forecast, a probability that a delivery is late. Better numbers, same human workload.
3. Recommendation. The system proposes a specific action: order 18 blocks of butter by Friday. The human still executes it, so the effort saved is the deciding, not the doing.
4. Agentic. The system executes the action within bounds you set — drafts and sends the PO, sends the chase email, reads the reply, updates the ERP — and escalates what falls outside those bounds.

Most tools marketed as agentic today sit at level 3 with a nicer interface. The honest test is simple: does the work get done while nobody is watching? If a person still has to open the tool for anything to happen, it is a recommendation engine.

What actually works today

In food and beverage planning, a handful of jobs are genuinely well suited to an agent right now, because they are high-frequency, low-ambiguity and easy to verify:

• Chasing acknowledgements. A PO without a confirmed date is not supply. Chasing it on day two rather than day ten is pure clerical work with real consequences, and an agent does it identically every time.
• Reading supplier replies. “We can do 15 of the 18, rest week after” is a plan change buried in prose. Turning that into a structured status — quantity confirmed, date changed — is exactly what language models are good at.
• Detecting the consequence of a slip. A delivery moving three days matters only if it breaks a run. Connecting the delivery date to the production schedule and the bill of materials is arithmetic the agent can do continuously.
• Raising the routine orders. The replenishment POs that follow directly from the material plan, within agreed suppliers, prices and quantity bounds.
• Keeping master data honest. Comparing quoted lead times against what actually happened, and flagging the ones that have drifted — the root cause behind a surprising share of shortages.

What does not work today, and you should be sceptical of anyone claiming otherwise: negotiating commercial terms, choosing between suppliers on anything beyond the criteria you gave it, deciding which customer gets short-shipped, or planning around information that lives only in someone’s head. Those are judgement calls with commercial and relationship consequences, and they belong to a person.

What to demand from a vendor

If a supplier of software says “agentic”, these are the questions worth asking, and the answers we would give for Leio:

What exactly can it do without asking? There should be a specific list, and you should be able to change it. “Sends chase emails and raises POs under £X with approved suppliers” is a usable answer. “It’s autonomous” is not.

What does it escalate? An agent that never escalates is not bounded; an agent that escalates everything is a notification system. The interesting design work is in the threshold.

Can I see what it did, and undo it? Every action needs an audit trail with the reason attached — what data it saw, what rule it applied. Anything it sends externally, a person should be able to see before or immediately after, and correct.

What data does it need? Realistically: the ERP for stock, bills of materials, open orders and suppliers, and the mailbox, because most of what actually changes a plan arrives as email. Ask what happens to that data and where it goes.

How does it behave when it is unsure? The correct behaviour is to stop and ask, with the specific question. The failure mode to fear is confident action on a bad assumption — which is why bounded scope matters more than model quality.

The honest limitations

Three worth stating plainly.

It will not fix your master data. If lead times are wrong and stock figures are fiction, an agent acts on fiction faster than a human does. A short, unglamorous data cleanup before deployment is worth more than any feature.

It will not remove the planner. It removes the clerical layer around the planner — the chasing, the re-keying, the reconciling — and gives back the time to do the part that needs judgement. Anyone promising an unattended supply chain is selling something.

It changes what you measure. Once an agent is chasing every order, “number of orders chased” stops being a meaningful metric. The measures that survive are outcome ones: on-time-in-full, expedites per month, service level, and how many exceptions were caught before they hit a run rather than after.

Why this is the right shape for food and beverage

Food and beverage manufacturing is a particularly good fit for agents, for reasons that have nothing to do with fashion. Shelf life makes the cost of both over- and under-ordering immediate. Supplier bases are fragmented, and much of the communication is email rather than EDI. Production schedules change weekly. Margins do not support a large planning team. The result is a planner with too many items, too many suppliers and too little time — the exact conditions where a bounded agent doing continuous clerical work is worth more than a better forecast.

That is the bet Leio is built on: keep the material plan live against the production schedule, raise and chase the purchase orders, read what comes back, and put the exceptions that threaten a run in front of the planner while there is still time to act. The planner decides. The agent does the following-up.

If you want to see that on your own orders rather than in the abstract, book a demo — or start with the supply chain planning process for the loop an agent has to fit into.