How AI Is Changing Online Marketplace Economics

In the last quarter of 2024, Etsy kept $22.80 of every $100 that buyers spent across its marketplaces. AI has barely touched that number. What it is doing instead is eating into the costs that sit underneath it, and it is doing so faster than it is changing any marketplace’s take rate. That gap, not the chatbot on the homepage, is what will decide who wins the next five years of platform commerce.

I have spent most of the past decade on marketplace builds and advising the people who run them, from holiday rental platforms to B2B spare parts exchanges. The board conversation about AI almost always starts with revenue. A smarter search bar. A premium tier. A shopping assistant with a friendly name. The conversation it should start with is the income statement, because that is where the money is actually moving.

The take rate barely moved. The costs underneath did.

A take rate is a price, and prices on two-sided platforms are sticky. Sellers compare fees across platforms, competition authorities watch them closely, and every percentage point a marketplace adds has to be defended on a results call. Raising it is slow, public and politically expensive.

Costs are a different matter. Nobody holds a press conference when a support queue shrinks or a fraud review team stops growing. That asymmetry is what I call the take-rate squeeze: the headline rate holds steady while AI presses cost out of the lines beneath it, and the real contest becomes who pockets the pressed-out margin. The platform. Its sellers. Its buyers. Or, increasingly, a new intermediary standing between all three.

Here’s the problem with most AI business cases I see. They are written as product documents, with screenshots and adoption targets, and the finance team sees the spreadsheet last. The AI work worth funding usually lives in two or three cost lines. Rarely all six.

Walk the income statement and the savings show up line by line

Picture a mid-sized marketplace doing $1 billion in gross merchandise value at a 15% take rate. That gives it $150 million of revenue to cover marketing, payments, support, risk and the people who recruit sellers. The figures in that example are mine and illustrative. The data points under each line below are not.

Customer acquisition

Marketing is the largest controllable cost for most marketplaces, and Etsy shows how large. In its full-year 2024 results, published on 19 February 2025, the company reported marketing expense of $856.6 million against revenue of $2.81 billion. By my arithmetic, that is about 30.5% of revenue spent on getting buyers through the door.

Before AI, the cost drivers were the auction price of a click on Google or Meta, the agency hours behind creative, and the analysts tuning bids. Generative tools have collapsed the cost of producing ad variants, and automated bidding now does much of the analyst’s work. But here’s the catch. Every competitor has the same tools and bids in the same auctions, so much of the saving on buying attention gets competed back to the ad platforms. The production saving is real. The media saving mostly isn’t yours to keep.

That is why acquisition sits low on my list when an operator asks whether to hire a marketplace development company or build AI capability in-house. I ask for the cost lines before the feature list, and the work worth funding tends to sit further down the statement.

Matching and conversion

Matching is where AI behaves least like a cost cut and most like a revenue multiplier, which lowers acquisition cost per order without touching the marketing budget. Amazon offers the clearest public number. On its third-quarter 2025 call on 30 October, the company said 250 million shoppers had used its Rufus assistant that year, that customers who engage with it are 60% more likely to complete a purchase, and that it expects Rufus to generate over $10 billion in annual incremental sales.

A finance reader should notice two things in that disclosure. The $10 billion is Amazon’s own estimate, built on an internal downstream-impact metric with a seven-day attribution window. And the gain shows up as conversion, not as a lower line on the cost side. There is also a risk buried in it. Sponsored listings are a large slice of marketplace revenue, and an assistant that answers the question directly may show the shopper fewer paid placements. If conversational search cannibalises ad inventory, part of the take rate erodes from the inside. That last part is my inference, not Amazon’s.

Customer support

Support cost used to be simple to model: contacts per order multiplied by cost per contact, with a seasonal spike that forced you to overhire or outsource. AI changes the first variable. On its second-quarter 2025 call, Airbnb said its AI customer service agent, rolled out across the US, had cut the share of guests who need to contact a human agent by 15%. Etsy’s full-year 2024 release noted it had expanded conversational AI in its support channels too.

On one rental marketplace I advised, the refund policy generated more tickets than any feature on the site. Those are exactly the contacts an AI agent resolves well. The trap is counting deflection as resolution. A bot that makes a guest give up does not save money. It moves the cost into refunds, bad reviews and churn, where it is harder to see and more expensive to fix.

Trust and safety and fraud

Most marketplaces understate fraud, because much of it sits inside chargebacks, refunds and declined orders rather than in a line labelled fraud. A Mastercard study with Datos Insights, reported in April 2025, found that 45% of chargebacks in 2024 were either first-party or third-party fraud.

The pre-AI defence was a rules engine plus a manual review queue, and its hidden cost was false declines, good buyers turned away by blunt rules. AI scoring on behaviour, device and listing images cuts both the review headcount and the false declines at the same time. That is unusual. Most risk controls trade one for the other. The counterweight is that fraudsters use the same models to write convincing fake listings and fake support messages, so this line will never reach zero. It gets cheaper per order, not cheap.

Payments and disputes

Disputes are a volume problem, and the volume is heading the wrong way. The same Mastercard work projects global chargebacks of 324 million transactions in 2028, up from a 2025 forecast of 261 million. Each one used to mean a person pulling tracking data, message history and delivery photos into an evidence pack, with a fee charged whether the merchant won or lost.

AI now assembles that evidence automatically and, more usefully for a CFO, triages which disputes are worth fighting at all. Who benefits depends on a contract term most people skim: the merchant of record. On many marketplaces the seller carries the dispute loss, so the seller captures most of this saving. Where the platform is merchant of record, it lands on the platform’s margin instead.

Supply onboarding

Every marketplace lives or dies on supply, and recruiting sellers has always been labour-intensive. Account managers, catalogue clean-up teams and listing quality reviews all sit somewhere in the cost base, and a new seller who takes weeks to post a decent listing often never makes a first sale. Amazon’s annual small business report, published on 30 April 2026, said independent sellers created more than 12 million listings using generative AI in 2025, and that its Seller Assistant had more than 230,000 monthly users.

Here the saving lands first with sellers, as time. The platform’s gain comes later, through better catalogue quality and faster supply growth, which is why it is so often left out of AI business cases. It shouldn’t be.

Calling AI a feature misreads the income statement

There is a line you hear at almost every investor day and product offsite.

AI is a feature, not a margin story.

It is half right, and the wrong half is expensive. A feature is something a customer sees and might pay for. Most of what moved in the walk above is invisible to customers. A ticket that never got opened. An evidence pack nobody typed. A listing written in minutes rather than an afternoon. None of it appears on a pricing page, and all of it appears in operating costs.

The deeper point is that support, risk and disputes are variable costs. They grow with every order. When AI takes a slice out of the cost per contact or the cost per dispute, it changes the slope of the cost curve, not just its level, so the saving compounds as GMV grows. That isn’t a feature. It’s a different cost structure.

Where the sceptics are right is on durability. If every marketplace in a category buys the same tools from the same vendors, the saving gets competed away into lower fees or more marketing. Which is precisely why who captures the saving matters more than its size.

Who keeps the saving decides who wins

The table below is how I map an AI roadmap against a marketplace P&L. The final column is the one boards argue about.

In my experience, operators do one of three things with the money. Some bank it as margin, which pleases investors for a few quarters. Some pass it to sellers as lower effective fees, which wins supply from rivals. Some recycle it into buyer acquisition. What goes wrong is failing to choose, so the saving leaks to whoever happens to hold it.

The marketplaces I expect to do well are the ones whose own data makes their models better than a vendor’s default: years of dispute outcomes, fraud labels and support transcripts. The ones I worry about built their edge on a manual service layer, a concierge team or a hand-curated catalogue. Their moat was labour, and labour is exactly the line AI is repricing.

The take rate is next, and the platform may not set it

So far AI has worked on the costs. The take rate itself is where the next fight sits, and it is arriving from outside the platform. In late September 2025, OpenAI launched Instant Checkout in ChatGPT, starting with US Etsy sellers, with merchants paying a fee on completed purchases. A new intermediary between the buyer and the marketplace is good for reach and awkward for economics, because any fee that intermediary charges has to come out of someone’s slice.

Rufus is the defensive version of the same idea: keep the assistant inside the store, so the shopper never leaves. Not every marketplace has Amazon’s budget to do that. My read is that the cost savings from the squeeze are the war chest for this next phase. A platform that has already taken cost out of support, fraud and disputes can afford to share revenue with an AI shopping agent. A platform that hasn’t will have to raise fees to do it, at the worst possible moment.

Questions a CFO should ask before funding the AI roadmap

1. Which P&L line does this project move, and by how much at today’s volume? If the answer is a feature description, send it back. A fundable AI project names a line, a current unit cost and a target unit cost. Anything else is still research, and should be budgeted as research.
2. What is the attribution window behind the revenue claim? Amazon was candid that its Rufus estimate rests on a seven-day window and an internal impact metric. Your vendor or product team should be equally specific. A conversion uplift measured over a generous window, on a self-selected group of users, will flatter almost any tool.
3. Does the saving scale with orders, or is it a one-off? Savings on variable lines like support and disputes compound with growth. Savings on a one-time catalogue clean-up do not. Both are worth having, but they deserve different discount rates and very different board slides.
4. Who will capture the saving, and is that a decision or an accident? Use the table above. If the merchant of record is the seller, the dispute saving belongs to sellers unless you change the contract. If the saving is meant to fund lower fees, say so in the plan, so nobody counts it twice.
5. What does it cost to run at peak, and who owns the data it learns from? Model usage is billed per call, so a seasonal spike that once meant temporary staff now means a larger compute bill. Ask for the peak figure, not the average, and check that the support transcripts and fraud labels feeding the model remain yours if you switch vendors.
6. What happens to the line when the system gets it wrong? A support agent that deflects without resolving pushes cost into refunds. A fraud model that tightens too far raises false declines and quietly loses good buyers. Every AI line item needs a counter-metric, owned by someone in finance, that would show the damage before the quarterly results do.

By the time the take rate moves, the winners will already have been decided on the cost lines nobody was watching.