Singapore

Amazon's auction trial: the case for approving agentic ad buying

The practical effect is a change in what you have to ask before you grant automation more control over budget and bids: who can approve a change before it ships, what record survives it, and what happens when the vendor's own numbers are the only ones you see. The FTC alleges Amazon kept a surcharge hidden from advertisers; Amazon disputes it. Whatever the court decides, buyers reviewing ai automation for business tools now have a concrete case to test vendors against.

Key takeaways

The FTC and 22 states filed an unredacted complaint alleging Amazon charged advertisers up to 50% above auction results on ordinary days and up to 250% on major shopping days.
The complaint says average cost per click rose 20% in the 2021 holiday season, and some advertisers saw increases of over 200%. Small budgets feel it too: at that rate a campaign buying hundreds of clicks a day pays for a full extra day of traffic every week.
A large food and beverage advertiser faced price increases of more than 100% during Prime Day in 2023, per the complaint.
Agentic buying moves the same risk from a human bidder to a model, so the control that matters is a recorded approval before spend, not a dashboard after it.

What the complaint actually alleges

The FTC filed the case with state attorneys general in the Western District of Washington, and a later filing made public the parts that had been redacted. The central allegation is that Amazon's search ad auction prices were raised above what competitive bidding would have produced, and that advertisers asking why received an explanation about a crowded marketplace.

The numbers in the unredacted complaint are the useful part for a buyer, because they describe how far a single auction design can move a line item. According to the FTC press release of 2026-11-13, the alleged surcharges reached 50% above auction results on ordinary days, 250% on Prime Day or Black Friday, a 20% average cost-per-click increase in the 2021 holiday season for some advertisers over 200%, a $87 million revenue lift from surcharges in the 2022 holiday season, more than 100% price increases for one advertiser on Prime Day in 2023, and an internal discussion of finding an additional $40 million in revenue to meet targets that year.

Amazon disputes the allegations. Treat every figure above as what the complaint claims, not as an established finding.

A media buyer studies a rising price chart beside an auction invoice while bidders raise paddles behind.

Why this is a trust problem for agentic buying

An agentic system is asked to do what a media buyer used to do manually: read auction signals, shift budget, adjust bids, decide where a campaign stops. When the auction itself is the thing under investigation, the question of who supervises the automation stops being abstract.

The complaint's mechanics matter here. If a surcharge can sit inside the auction result and still look like market movement, then an automated buyer reading that result has no way to tell the difference, because the system trusts the same number the advertiser is being asked to trust.

So the first thing to ask of the agent is not how well it optimises but what it will show you: the raw auction or cost data behind an action, not only its outcome. A change above an agreed threshold should then wait for a named approver, and the record of that approval ought to outlive the campaign so a later audit can rebuild the decision. Anyone comparing sem agency partners, or weighing in-house automation, can use that sequence as a simple test of whether the layer in front of the auction is accountable at all.

None of these prevent a bad actor on the platform side. They do make it impossible for an automation layer to hide behind the platform's own explanation. That is also where the cost of ai automation for business sits: the licence or retainer is the visible line, and the unreviewed bid changes are the invisible one.

What platform-native AI cannot see

The structural case for advertiser-side oversight rests on boundaries that no single platform algorithm crosses, regardless of how good its bidding is.

Start with cross-platform budget weight. Meta's automated budget sits at campaign level and depends on its own bidding framework; Google's Smart Bidding reallocates within its own traffic. Neither can decide how much of the month's budget belongs to the other platform, so that decision has to be made outside both systems, and whoever makes it needs a measurement standard both sides accept.

The attribution window narrows the picture further. The platform also takes CRM conversion data for training, but under a strict time limit, and conversions that land outside the official attribution window are dropped. Meta and Google each set their own window, both measured in days rather than months. A B2B deal that closes weeks after the click never enters the model that was supposed to learn from it, so the platform's optimisation is trained on a subset of reality. These limits are the framework's premise as stated in platform documentation, not a legal reference.

Then there is the spend that happens outside the ad account. Landing pages, form completion and sales follow-up sit in the funnel but outside the platform's reporting, and that is where part of the wasted budget accumulates without anyone seeing it. A sem agency working from the ad platform's dashboard sees the clicks and misses the drop between click and enquiry.

An analyst faces separate platform screens showing partial funnels, with a gap between click and enquiry.

Best marketing automation software for a small business: the checks that matter

Small teams are the ones most exposed to the scenario above, and the question they actually ask is simpler than a feature comparison: what should I check before I let this software move my money.

Judge the product on controls, not on the model behind it. Four things are worth verifying in a trial:

Control	What to check in the trial	Why it matters
Approval before spend	Whether a budget or bid change can execute without a person approving it, and who that person is	A hidden platform-level change cannot be caught, but an unrecorded software-level one can
Run records	Whether every action links to the task, the data and the approval that produced it	An audit months later needs the reasoning, not just the change
Shadow mode	Whether the tool can analyse without touching the account, and for how long	Comparing its recommendations against your own costs nothing while you test
Rollback	Whether a change can be reversed, and whether timeouts and failures are recorded	Agents fail silently more often than they fail loudly

A tool that answers all four can still produce mediocre results. One that cannot answer them gives you nothing to audit when results disappoint, which is the same position the advertisers in the complaint describe. Running the sem agency comparison this way also costs nothing beyond the trial period.

How Topkee structures oversight around the spend

Topkee's position is that the trust problem is a governance problem, not a model problem, and its product lines are built to keep approval inside the execution path rather than in a review afterwards.

The TTO Funnel Tuning Agent runs one agent instance per campaign, joins ad, site and CRM data into a full funnel, and outputs a P0/P1/P2 action list. Every action starts as pending confirmation; until the client grants auto-execution rights, the system audits, prioritises and records reasons without changing the account, and reports zero unauthorised account changes at delivery, with four hand-off modes covering agent executes, route to Topkee expert, await approval, and archive. On the platform side, the Topkee AI Agent platform gates pending steps through Approvals and links the task list to its evidence in run records. The AdOps Vanguard paid-media template starts from the KPI plan, the channel data and the budget, and holds any task involving spend or publishing until approval. Its related paid-ads capability has a shadow mode that analyses only. Diagnosis stays traceable: each recommendation carries its target level, object ID and the data behind it, and each report follows the metric, judgement, weighting and action chain with the evidence kept as a document.

The limitation is worth stating plainly. Approval and run records are controls, not proof that a platform's pricing is honest; they make your own layer auditable, and the platform's auction remains the platform's to explain. For a reader deciding between a platform-run agent and an agency-run one, that asymmetry is the point, and it is the reason the sem agency contract should name the approver rather than the tool.

A hand hovers above a confirmation switch on a control panel, a clipboard of approvals beside it.

Recommended next step

Run a free account audit with read-only access before you grant any agent write permission. The diagnosis shows the funnel stage that leaks budget, the evidence behind each finding, and the actions ranked P0/P1/P2, and you can withdraw access at any time.

Wider ai automation for business projects tend to start here, with one paid channel that can be audited end to end. If you want the context on how ad delivery changed before you decide, read what Google's Q3 results mean for your SEM budget, and for the technical side, server-side ad calls and what they change for SEM.

FAQ

Did the court find Amazon guilty?

No. The filing made public allegations that had been redacted. Amazon disputes them, and nothing in the complaint has been established as a finding.

Does an approval step slow down campaign optimisation?

It adds latency only where the change crosses a threshold you set. Routine bid and budget moves below that threshold can execute automatically once you grant the rights; the approval gates the large or irreversible actions.

Can automation detect a hidden platform surcharge?

Not reliably from inside the platform's own reporting. Comparing cost per click and conversion quality across channels with one measurement standard, and reconciling against CRM outcomes, is what surfaces a gap between what the platform reports and what the business earns.

What does shadow mode prove?

That the tool's recommendations match or beat your current decisions without touching the account. It does not predict spend and it does not validate a budget plan; it is a comparison run over real data.

Is this relevant to a small business on one platform?

Yes, because the exposure is per click, not per channel. A 20% cost-per-click increase on a small daily budget is small in absolute terms and still changes the maths on every sale the campaign produces. The control that scales down is the same one: an approver, a log, and a way back.