At three in the morning, a good AI agent on a marketplace account does not sell more and does not make bold decisions. It watches, it compares, it carries out the few small and reversible actions you have authorised in writing, and it prepares everything else for the person who opens the laptop at eight. Anyone who chooses an agent for its autonomy is buying the wrong thing. What matters is the boundary: the precise list of what it may do on its own and what it must only report.
I am saying this now because in mid-October the subject became concrete even for people who do not work in software. On 13 October Google brought Gemini 4 Argon to its app with a tasks mode that, in the company's own words, works for you even when you are not at your computer, and it connects to more than forty applications, Shopify included. At the end of September Amazon had already upgraded its Seller Assistant with workflows that follow prices, stock and product performance continuously. Agents that work at night are no longer a conference promise. They are a feature you switch on with one click, and that is exactly why they need to be thought through before they are switched on.
Why the night matters for an AI agent on a marketplace
A marketplace never closes. While Italy sleeps, a German competitor drops a price, a customer in Spain writes to ask where their parcel has gone, and an item that was selling well runs down to zero units. Anyone selling across several European countries lives on hours that differ from those of their own office, and the hours between midnight and seven are the ones in which small problems become big ones, simply because nobody sees them.
For years there were two kinds of answer. Either you accepted the gap and discovered the damage in the morning, or you handed the night to rigid rules: the repricer that follows the lowest competitor, the automatic alert when stock falls below a threshold. Rules work as long as the world behaves as expected. The trouble is that the unexpected is precisely what happens at night, and a blind rule chasing a wrong price can burn margin for hours.
An AI agent differs from a rule because it reads context. It can notice that the competitor who halved a price most likely made a typing mistake, that a drop in sales comes from a listing that is no longer visible rather than from demand, that a message from an angry customer deserves a person and not a template. This ability to judge is its strength and also its risk, because a system that can interpret can also misinterpret.
What an AI agent does on your account while you sleep
Let us follow one through a night, on an account that sells on several European marketplaces and runs its own warehouse.
Around one o'clock it matches the evening's orders against real stock levels. If a product is running out on one channel but still available on another, it moves quantities according to the rules it has been given, because that is an action you can undo in a minute. If the goods are actually missing, it does not invent solutions: it flags the product, estimates how many days of sales are left and prepares a draft reorder for whoever handles purchasing.
Around two it looks at prices. A competitor has cut the price of an item you have been competing on for months by five per cent: if the pricing policy allows it and the margin stays above the set minimum, the agent follows within the limit it has been given. Another competitor has halved the price of an identical product, and here a good agent stops. A change that large is more often a mistake or a clearance than a new normal, and chasing it at night means selling at a loss until morning.
At three the messages arrive. Someone asking where their parcel is gets a precise answer with the shipping status, because that information is in the data and requires no judgement. Someone writing about a faulty product, threatening a negative review or asking for a refund outside the rules finds in the morning a reply that is already drafted, but a person reads it and sends it.
Before dawn the agent does the most useful and least spectacular thing of all: it writes the handover. What happened, what it did and why, what it left pending and what needs a decision. Whoever arrives at eight does not start from scratch and does not have to rebuild the night by looking at ten different screens. They start from one page.
What an AI agent must not do alone at night
The list of forbidden actions matters more than the list of permitted ones. An agent must not make bulk changes to the catalogue, because an error across a thousand listings shows only once the damage is done. It must not change prices beyond a written threshold, nor raise advertising budgets because it saw a spike in traffic. It must not reply to marketplace communications about the state of the account, which always stay with whoever is responsible for it. And it must not promise a customer refunds or discounts that nobody has approved.
The criterion for deciding is easy to state and hard to apply. You have to ask whether the action can be undone at no cost by the morning, and how much it costs if it is wrong. Moving availability between channels, answering a shipping question, pausing a listing that is selling below cost are reversible, low-risk actions. Rewriting titles and descriptions, changing the pricing strategy, touching account settings are not. The agent carries out the first kind and prepares the second.
The boundaries: how to decide what an AI agent can do
The boundary is not a feeling, it is a document. It states in numbers how far a price can move, above which minimum margin, on which products and under which conditions. It states which messages can go out without review and which cannot. It states what must happen when the agent does not know what to do, and the right answer is almost always to stop and report.
The second piece is the log. Every night-time action must leave a trace a person can read: what the agent saw, what it decided, which rule it applied. Without a log nothing can be corrected, and you cannot even trust the good results, because you do not know whether they came from merit or from luck. It is no coincidence that the marketplaces themselves are moving in this direction: in its September update Amazon lets sellers choose whether the agent only suggests or also acts, records every action and allows limits and approvals to be set.
The third piece is time. A boundary is widened step by step, by reading the logs. In the first weeks the agent only suggests, and you count how often it would have been right. Then it is trusted with the safest actions. Only later, if the numbers justify it, are the thresholds raised. Whoever switches everything on the first day usually switches everything off in the second month.
Software, people and AI agents: why you need all three
This is the model on which we built Selleee, the platform that combines account-management software, human operators and AI agents. The software keeps the data in order and applies the rules. The agent works in the hours when nobody is watching, reads the signals and prepares the work. The person decides, handles the difficult cases and widens or narrows the boundary according to what they find in the logs.
None of the three is enough on its own. Software without an agent reacts only to what someone foresaw. An agent without people ends up making decisions that are not its to make, with a confidence it has not earned. People without software and without an agent spend the morning rebuilding the night instead of growing the account. The value lies in the handover from one to the other, and that is exactly where most companies lose time today.
AI agent for marketplaces: where I stand
In the coming months every platform will offer you its own always-on agent, often included in the price and often behind a single button. The question is not whether to switch it on, because you will, and you will be right to. The question is who writes the boundary and who reads the log. If the answer is nobody, you do not have an agent working for you at night. You have something deciding on your behalf without anyone knowing.
I would ask any vendor a single question: what does your agent do at three in the morning when it does not know what to do? If the answer is that it finds a solution anyway, change vendor. If the answer is that it stops, writes it down and leaves the decision to a person, you have found someone to work with.
CEO & Co-Founder of BusinessLabs. I have attended major international trade shows since 2015 and share here experiences and insights for online sellers.
LinkedIn