AI inventory management
Most AI inventory management projects break before the model ever gets a chance to. Your marketplace, your warehouse and your production system each hold a different number for the same item, nobody has joined them, and a faster reader of bad data only makes a confident mistake sooner. That gets fixed first. Then an agent on top is worth having.
Your systems hold three different numbers for the same item, and a faster reader of bad data only reorders the wrong one sooner. Your shelves, your rules — not mine.
One honest number per item, then an agent that reorders inside your limits
Tell me where your stock numbers live
Two lines is plenty. Where the numbers live, and which two of them disagree. I read it myself and answer within one working day.
Sergey Avakyan. You are writing to me, not to an inbox — I am the whole company, and I am the one who builds it. More about me →
Routine reorders inside a threshold you set run unattended; a fall that might be demand and might be a data error stops and routes to you with the reason attached.
Projects go in the order they arrive — the earlier we start, the earlier yours is built.
- One honest number per item, joined from the marketplace, warehouse and production systems you already keep.
- A reorder agent with a threshold you set — it acts inside the limit and stops, with the reason, outside it.
- A guardrail that watches its own decisions and resets itself, plus a monitor that notices the moment it stops running.
- A written list of what it must never reorder alone — a stock fall that could be demand or a bad feed, anything above the limit — enforced in the code.
The shelf robot underneath this is not new work: it tags every product and rebuilds the shelves itself, 968 products tagged, running as one of the thirteen systems of mine live every day — and its guardrail has already fired for real, out of sixty-five built.
A quarter of whatever we agree is due at the start, and that quarter is the only money at risk under the two-week promise — nothing else is due until it is running. What puts you at the bottom of that band: stock that lives in one or two places, a reorder rule you can already state, and items that behave the same way. What puts you at the top: a marketplace, a warehouse and a production system that each hold a different number and were never joined, seasonal or promotional demand a dumb rule has to catch, and thresholds nobody has written down. It sits in the middle band because the model is the easy half — the work is agreeing one honest number per item first, and then what the agent may reorder without you.
- Two weeks, or the first stage is free. The first stage is one named piece, written down before we start. If it isn't running two weeks after the start, you pay nothing for it. The clock pauses on any day I'm waiting on you for an access or an answer.
- You write the pass mark before I write the code. Before anything is built we agree one sentence that says what working means — for example, "it takes the call, puts the booking in the calendar and never promises a price, on nine of every ten of fifty real calls". If it misses that, you choose: I finish it at my own cost, or you take that stage's money back and the work stays with me.
- Three months of fixes included. For three months after handover I fix whatever breaks in what I built, already paid for. Something you ask for later, or a service someone else changes underneath it, is new work — and I'll tell you which it is before I touch it.
“I” here is the whole company: AVAROPA SLU, one founder and one employee, both of them me. The guarantee is the company's, and I am the person who honours it — you deal with the same one throughout.
Every stage you have paid for is yours the moment it is paid — the code and the accounts, handed over as we go. Nothing here depends on you staying.
Prices exclude VAT — 21% in Spain, reverse charge for an EU business with a valid VAT number, and normally none outside the EU — confirmed on your invoice. The band is a starting point; the exact figure is agreed and fixed before any work begins, and it does not move after that. The two-week first piece is the part I guarantee; the finish windows are what a job this size usually takes.
How it works, what it refuses, and when you don't need it
Where it actually goes wrong
- The agent sees half your stock. Goods in production sit in one system, warehouse stock in another, and what the marketplace thinks you have is a third number. An agent that cannot see inbound goods will reorder things that are already on a truck. That is not a model problem — it is reading exactly what it was given.
- A fall in stock is not always demand. Stock dropping looks like selling well. It can just as easily be an incident at someone else's warehouse. Without a baseline to compare against, an agent buys inventory to solve a logistics problem.
- Nobody checks the forecast against a dumb rule. A moving average is free and genuinely hard to beat. Before anything gets built, the forecast is run against that baseline on your own sales history and you see both error rates side by side. If it does not win, it does not ship.
- The tool your platform sells you only sees your platform. It cannot see your production, your second sales channel or your own store, so its picture of stock is incomplete by design — however good the interface is.
- It fails quietly. A reorder job that stopped running looks exactly like a calm week. You find out from the revenue, a month later, and by then you cannot tell which decisions were wrong.
What gets built
- One honest number per item. Every system feeds one clean store: the raw data kept, the same product matched across systems, ready numbers on top. Until this exists, everything above it is decoration.
- A reorder agent with a threshold. Routine top-ups inside the threshold run on their own. Anything past it stops, names the reason, and waits for a person. Spend caps, a margin floor and service tiers are where the threshold comes from.
- A guardrail that resets itself. After a real change it measures against the new normal. A guardrail that fires every day teaches everyone to click through it, which is worse than not having one.
- Something watching the watcher. A monitor whose only job is to notice when the reorder job stops running, and say so the same day.
What a refusal looks like
On 20 July the guardrail on my own shelf robot fired for the first time since launch. Over one week the pool of healthy stock — items sitting at 50 units or more — fell from 488 to 339, about 30%. Two things had happened at once: a surge of orders, and an incident at the marketplace's own warehouse.
The robot did not top up into that fall. It held shelf 082 and said why, in words a person could act on: healthy pool down 30% against last week.
Then a person did three things. Confirmed the drop was real and not a data glitch. Pushed one bestselling shelf through by hand, past the guardrail. Re-read the live storefront to check the result — 20 items expected, 20 items showing.
No code was changed. The guardrail reset itself and now measures against the new baseline of 339.
That refusal is the feature. The smarter-sounding version would have seen stock falling, concluded demand was up, and refilled hard: confident, fast, and wrong.
Too early for you if
- You sell on one channel with a few dozen items — the platform's own analytics is enough, and a project here would cost more than it returns.
- Your cost price is not tracked anywhere. Start with the accounting side: a forecast sitting on top of unknown cost tells you nothing about margin.
- You want a one-off report rather than something that runs every day and can act.
The line I do not move
The agent drafts, a person signs. Routine work inside the limits runs unattended; anything unusual routes to a human with the reason attached. Every system on this site is built that way, and this one is where it matters most — because here the agent spends money.
Already running on real stock
- Recommendation shelves — the shelf robot from the story above. It tags every product and rebuilds the shelves itself; 968 products tagged.
- Sales funnel report — the same idea applied to the funnel. It reads roughly 750 products every day and flags only the ones with a real problem.
- Systems watchdog — the monitor that notices when an automation stops running, so nothing dies quietly in the background.
- Promotion guard — the neighbouring problem: stock you priced correctly, swept into a discount nobody agreed to.
Honest answers
How much does AI inventory management cost?
€4,800 to €8,300 as a one-off build, higher when a marketplace, a warehouse and a production system each hold a different number and were never joined. A quarter is due at the start and is the only money at risk under the two-week promise; nothing else is due until it runs. The first hour is free.
Is AI inventory management worth it?
It pays when stock lives in more than one place and the numbers disagree, and a wrong reorder costs real money. If you sell on one channel with a few dozen items, it is not worth it yet — your platform's own analytics is enough and a build would cost more than it returns.
What does AI inventory management actually do?
Three things, at its best: it keeps one honest picture of stock across your systems, it forecasts what is about to run out, and it places routine reorders inside limits you set. What it should not do is decide anything unusual on its own.
How long before it is useful?
The forecast needs a couple of months of real sell-through per product line to be worth anything, and brand-new lines stay poor until then. The data work — joining your systems into one number per item — is usually the longer half, and it pays for itself before any model runs.
Does it work with marketplaces?
Yes, and marketplaces are their own animal: several warehouses, different fulfilment models, and demand bent by your own advertising. The platform data is clean and can be pulled; the value comes from joining it with your production and warehouse numbers.
Can the agent order stock without me?
Inside a threshold you set, yes — the routine top-ups. Outside it, no. Spend caps, a margin floor and service tiers are the limits, and anything past them routes to a person with the reason attached.
What if my stock data is a mess?
Then that is the project. Take twenty of your best sellers and count them in three places on the same day: the platform, the warehouse, production. Write down every disagreement and how old each number is. That one afternoon tells you more than any vendor demo.
Send me your stock picture
Write a couple of lines about where your stock numbers live and which two of them disagree. I will tell you plainly what can be automated now, what has to be fixed first, and what it would take. The first conversation is an hour and it is free.
Book that hour →