Codex isn't a copywriter anymore. It's an operator.
In a demo, Codex completed full competitive research in nine minutes, from one prompt. What agentic AI already means for your Bol.com and Amazon operation, where the human stays in the loop, and why every agent needs a burden of proof.
Last week I read yet another article full of AI tips for e-commerce. A chat here, a prompt there. Meanwhile I was cleaning up my own Google Ads account after taking it back from an agency: draining and rebuilding negative keyword lists, pausing campaigns, shifting budget. By hand. Exactly the kind of repetitive, rule-based work where an AI with browser control makes a real difference. Not as a copywriter. As an operator that pushes the buttons.
Why "Codex tips" is the wrong question
Most decision-makers missed the moment this changed. In the video this piece responds to, Codex completed full competitive research on cameras, including the price delta against an existing model, ratings and decision data, neatly delivered in a spreadsheet. From one prompt. In roughly nine minutes.
The demo that makes the point
Demo conditions; expect slower in practice, but the same category of task.
Nine minutes is not the news. Look at what happened: browser opened, sites visited, data collected, structured, delivered. That is a different category from a chatbot tossing you a paragraph of product copy.
Why this matters right now to anyone selling on marketplaces: the channels themselves are moving the same way. My own webshop is live on ChatGPT Commerce, visible in the Dutch automotive aftermarket next to players like Winparts and Autodoc. Wait until it's proven and you start from behind.
From chatbot to operator: what agentic execution really means
If you still use ChatGPT as a copywriter, you're using a fraction of what it can already do. The real jump is agentic execution: browser use and computer control let the AI finish entire tasks instead of producing text.
Concretely, this touches marketplace work directly. The AI navigates your browser on its own and runs competitor research. In the demos it also handles refunds and negotiates with customer service; I have deliberately not put that part into production myself. And it cleans up your computer, from sorting files to archiving your mailbox. At my first marketplace launch I still did all of this work myself, before I dared to hand it off.
What most people overlook: an agent that does work also reports work. And those reports are not free truth. When I audited my own AI stack, roughly half of the success messages could not be traced to a real action. The operator promise is real, but it ships with a burden of proof nobody mentions in the demos.
Where this touches your marketplace today
Take the camera research and replace "camera" with "your bestseller on Bol.com". Automated price and competitor monitoring on Amazon NL and Bol.com, including delta analysis against your own listings. Same functionality, different data.
On these marketplaces that compounds fast. The Buy Box dynamics and pricing on Amazon versus Bol.com make real-time monitoring show up directly in your ROAS.
I've seen that dynamic up close. At a large Dutch consumer-electronics retailer I built the Bol.com operation from zero as E-Commerce Manager, to more than two-thirds market share in its category. We tamed Buy Box volatility through B2B price alignment with retail partners, without a price war. Price stability and stock reliability beat the lowest bid structurally. A targeted social campaign added around 20% weekly sales growth during the campaign period, and out-of-stock stayed under 2%.
Orchestration over prompts
The next shift is subtler but bigger. ChatGPT introduced native voice mode that lets you steer multiple agent threads at once and run them in parallel. You talk, an agent starts a new thread, you move on.
"Can the AI do it" has become the wrong question. The bottleneck is you: is your process built to orchestrate multiple agents?
And it goes further. ChatGPT sites lets you publish output as a spreadsheet, a PowerPoint or a site. You then hand that URL to another agent as input. Agent one's output becomes agent two's input, and that is the seed of an automated e-commerce pipeline. Quality-of-life features like pinning threads sound like gimmicks, until you're managing dozens of running AI processes. More on that in my full guide to these tools.
I learned what that looks like in practice on my own eBay DE operation. The dashboard reported zero order errors across 64 orders; reality had two. Two paid orders sat unshipped for days, with a ship SLA running out. Since then an agent pulls an order snapshot every 30 minutes and the not-yet-shipped list simply appears in my morning brief.
The NL/EU reality check
Here I have to draw a line. An agent that independently changes prices, emails customer service or edits listings introduces compliance, brand and error risks. In the regulated EU context those are not trivial. Think VAT rules and consumer law, plus the limits marketplaces themselves put on automated pricing behavior.
A price-changing agent in the Netherlands is fiscally and legally more sensitive than the American story suggests. The question is not whether you deploy agentic AI. The question is where you place the human in the loop.
And don't assume your own numbers are telling you the truth. When performance kept disappointing after a campaign overhaul, my first assumption was that my own work was the problem. It turned out Consent Mode v2 sat at 0% granted: over two-thirds of conversions were invisible to Smart Bidding, and I had been optimizing against incomplete data. GA4 showed steady growth all along. A pre-existing measurement bug, not a result of my changes, but my responsibility to diagnose and hand over. An agent on broken measurement data only automates your mistakes.
What this means for your team
Some roles shift hard. Research, monitoring, first-line customer contact, listing checks. That is exactly the work an agent can do end-to-end. If your listings don't convert anyway, this is the moment to rebuild your process.
My own shortlist is already running. A daily radar compares 204 feed rules in Channable line by line against yesterday, so configuration drift reports itself. Order snapshots every 30 minutes, unshipped orders on top. Mail triage that never archives money-related mail and learns from my corrections. And weekly competitor and SERP scans that arrive as a report, so I don't open dashboards anymore.
My delegation list in numbers
Every agent on this list ships proof with its work; a report without an artifact doesn't count.
The gap becomes your competitive edge
The gap between "AI as a toy" and "AI as an operation" is exactly where the advantage lives.
What I do next is settled. Every repetitive task in my operation gets an agent, and every agent gets a burden of proof. The machine may write; the truth stays handmade.
Book an AI e-commerce audit. I'll walk through your marketplace operation and show where an agent already saves hours. You get the same burden of proof I put on my own agents: every finding comes with an artifact, or it doesn't count.
Sources and further watching
Want to spar about your marketplace strategy?
No hype. A sober look at where your growth is and where margin leaks away.
Get in touch