Inventory replenishment with an AI agent: how to avoid stockouts and rush purchases
What checking stock by hand costs, what the rush purchase that follows costs, and what an agent does at each step of the replenishment.
Iván Itzcovich
Co-founder
This post is about inventory replenishment: how an AI agent avoids stockouts by checking every item every day, and how it gets implemented without touching the stock system.
It starts from an uncomfortable fact. A buyer has two thousand active items and one morning a week to look at them.
What they should do is, for each one, compare today's stock against the reorder point, and that reorder point against that particular supplier's lead time. Two thousand times. Then build the list of what has to be ordered, confirm lead times, issue orders and chase them until they arrive.
Nobody does that. What gets done is looking at the fifty you carry in your head, and reacting when somebody at the store says there is none left.
The cost isn't checking, it's the rush purchase
This is what makes this process different from the other eight: here the manual work barely happens at all. It is not that it takes many hours, it is that it does not fit.
So the cost is not in the hours. It is in what happens when the check arrives late.
When a shortage is caught past the point where it could still have been ordered in time, there are two ways out and both cost. One is the rush purchase: list price, special freight, sometimes another supplier. The other is not replenishing and losing the sale.
Neither shows up in an inventory report. The first is spread across purchase invoices and the second is nowhere at all, because a sale that did not happen leaves no trace. That is why this is the process where the real cost is the largest and the least visible of the nine.
The same replenishment, today and with an agent
One replenishment cycle, end to end
- the agent does it
- a person does it
- always a person
the steps of the replenishment
- Check every item against its reorder pointStock system
Looks at all two thousand every day, not the fifty somebody carries in their head, and flags the ones that crossed the point.
- Cross it with each supplier's lead timeSuppliers
Uses that supplier's real lead time rather than an average, which is the difference between ordering in time and ordering late.
- Build today's order listPurchasing
Delivers the day's list with the reason for each line, instead of a spreadsheet somebody has to interpret.
- Confirm the lead time with the supplierEmail and chat
Writes or calls to confirm the lead time still holds before issuing, which is when lead times stretch without anybody saying so.
- Assemble the purchase orderPurchasing
Leaves the order assembled with quantities, current prices and the date it is needed, ready to authorize.
- Chase the order until it arrivesSuppliers
Chases the supplier when the date passes, which is the follow up abandoned the moment the order is issued.
- Flag the stockout that can no longer be avoidedEmail and chat
When an item is caught past its window, tells sales before a customer finds out.
- Authorize the purchasePurchasing
A purchase order is money committed to a third party. A person authorizes it, on a list already assembled and justified.
Today that is 70 hours a month of checking and 12 items ending in a rush purchase or a stockout. Between the time and the premium, USD 1,320 a month.
With the agent holding 7 of the 7 steps, 4 hours and USD 66 are left.
The difference is USD 1,254 a month, USD 15,048 a year.
And 36% of that difference is not time: it is the premium on rush purchases that stop being necessary.
Of what a person still does, the heaviest step is: Authorize the purchase.
You set all five numbers. How the minutes split across steps is ours, taken from how it gets done by hand. The modelling assumption: the stockout born of checking late falls in the same proportion as that check stops depending on somebody having a free morning.
Leave only the first step on, checking every item every day, and switch off the rest. It is the minimum implementation and the biggest mover, because the problem in this process is not the execution but the frequency of the look.
This object also has a page of its own, in case you want to send just that.
What the agent does that a person cannot
It looks at all two thousand, every day. It is not that it does it better: a person cannot do it, which is why the real process is looking at fifty and reacting to the rest.
It uses each supplier's lead time, not an average. The average lead time is what makes an item from a thirty day supplier get ordered as though it were ten.
It chases the order after issuing it. Follow up gets abandoned the moment the order goes out, and that is exactly where lead times stretch with nobody saying anything.
What stays with a person
Authorizing the purchase. An order is money committed to a third party, and it deserves somebody saying yes, at least for the first few months.
And the policy decision: which products carry safety stock, which get bought to order, which to drop. That is commercial judgment and the agent does not have it.
Where this approach doesn't work
Worth saying so as not to overpromise.
It does not work on very irregular demand, where one month's consumption says nothing about the next. There a fixed reorder point produces false alarms.
It does not work for short seasons, when the lead time is longer than the whole season: the decision was made at the initial buy.
And it does not work when system stock does not match warehouse stock. There you fix counting first, because automating on top of a wrong figure only makes the error arrive faster.
How it gets implemented
Read access is all you need to start: stock, reorder point and lead time per item, compared every day, with a daily notice of what has to be ordered.
The most common friction point is not technical: it is that each supplier's lead time is recorded in no field. That is the preliminary task, and it is less work than it looks because most suppliers fall into three or four typical lead times.
Writing, which means issuing the order, comes later and with authorization.
What to measure
The first effect is counterintuitive: orders go up. Not because more is being bought, but because replenishments stop being postponed that used to be discovered late.
The number that matters takes a quarter: how many rush purchases were made. That saving is spread across invoices and is usually larger than the lost sales everybody was trying to measure.
And the third, the easiest to forget: how many issued orders arrived after the committed date. That one tells you whether the problem is still the checking or has moved to the supplier.
Frequently asked questions
What does an AI agent do with inventory replenishment?
It checks every item every day against its own reorder point and its own lead time, builds the list of what has to be ordered today, asks the supplier to confirm the lead time before issuing, leaves the purchase order assembled, and then chases that order until the goods arrive. It does not forecast demand: it compares stock against lead time.
Do you need demand forecasting to avoid stockouts?
In most operations, no. Before forecasting there is a simpler problem: the stock figure exists and nobody reads it often enough. Warning early is not forecasting, and it is where most of the value sits.
What is the real cost of not checking in time?
The rush purchase. When the shortage is caught late, replenishment gets solved at list price with special freight, or worse, it does not get solved and the sale is lost. That cost is spread across purchase invoices and appears in no inventory report, which makes it larger than anybody believes.
Iván Itzcovich · Co-founder, StudioChat
Want agents like these working for your team?
Talk to us now