When Publishers Get AI Agents: Andrew Mole on Agentic Trading and the Future of Media

Full analysis

Andrew Mole, CEO and co-founder of PubX, went on Signal & Noise to argue that AI agents negotiating directly for buyers and sellers make the whole programmatic stack (DSPs, SSPs, agency trading desks) structurally redundant. PubX is live, moving a few thousand dollars a day across 3,000+ publisher sites, and Mole expects real market uptake in 2027.

Here's the frame. The news isn't "PubX transacts a few thousand dollars a day." That's a rounding error. The news is that the plumbing to bypass the stack is entirely open source: agents call a stripped-down bidder over OpenRTB, reconcile inside Prebid (the header-bidding wrapper most publishers already run), and use price priority in the ad server. No DSP fee, no SSP fee, same mechanics. This is a Type 2 decision for most operators (cheap to experiment, easy to walk back) and a Type 1 for anyone whose P&L depends on being a pass-through toll booth. Forcing function: Mole says holdco engagement accelerated over the last six months, and CMOs who use Claude at home are starting to ask why their agencies don't.

The Market Analyst. Mole is talking his book, loudly. He builds the thing that kills DSPs, then predicts DSPs die. Weight accordingly. But strip the self-interest and one claim survives: he's saying The Trade Desk gets "resized" to Viant scale because its value is pure utility, and utility is exactly what agents replicate for free. The plain-English version: if a machine can do the buying for a tenth of the cost, you don't pay a premium for the buying software. The counter he ignores is that The Trade Desk saw this coming and built OpenPath, CTV supply deals, and planning tools that are harder to commoditize than open-auction bidding. SSPs with real publisher data relationships survive. Pure resellers do not. That part isn't controversial.

The Skeptic. For this to work, publisher first-party data has to be worth transacting on at machine speed, and agents have to price inventory off honest ad-server forecasts instead of the inflated bid stream. Mole says 80% of bids are duplicate, which means the "open web audience" is a fiction inflated by the same impression getting counted five times. Fine, grant it directionally. But the reason the stack has DSPs and SSPs isn't only fee extraction. It's trust arbitration: who guarantees the impression was real, brand-safe, viewable, and paid for. Mole's answer to brand safety is "same as today," plus some conversations with IAS and DoubleVerify about encoding signals via MCP. He also notes Experian shipped MCP servers with zero usage. That's the whole problem in one sentence. The rails exist and nobody's driving on them.

The Operator. Tuesday morning, what breaks? The demo runs on Prebid and GAM, which every premium publisher already operates, so the integration story is genuinely light. That's real. What's not modeled: reconciliation and disputes. When an agent negotiates a price and the campaign underdelivers, who eats it, and which log is the source of truth? Scope3 is handling the payment float, which tells you money movement is the unsolved part, not bidding. And the publisher salesperson story is optimistic. Mole says they move to "high-value creative partnerships." Some will. Most sit inside a comp plan tied to transactional volume the agent just automated away. That fight happens at 90 days, not in the deck.

The CFO. Run the arithmetic Mole hands you. Publishers get 41 cents of the advertiser dollar today. PubX models a 10 to 15% take rate. Even granting his numbers are illustrative, a publisher moving open-market spend onto agentic rails plausibly keeps meaningfully more per dollar. That's the pitch, and it's a good one. But the real CFO risk is switching cost, not take rate. You're routing revenue through a startup transacting single-digit thousands per day, with payment float run by a second startup. The opportunity cost of a pilot is near zero. The opportunity cost of moving real budget before the money rails are proven is a bad quarter you can't reverse. That's exactly the innovator's dilemma Mole names for holdcos, and it applies to publishers too.

The tensions. Three real disagreements. First, is the DSP's value utility or trust? If it's utility, Mole is right and agents strip it out. If a meaningful chunk is verification, fraud defense, and someone to sue when a campaign goes wrong, the stack thins but doesn't collapse. Second, does open-source plumbing plus first-party data actually replace the demand aggregation a DSP provides, or does demand just re-aggregate somewhere new, with a new toll? Third, the timeline. Mole says 2027 for real uptake. The Experian "MCP servers, no usage" detail argues the gap between plumbing shipped and money flowing is longer than the believers think.

Synthesis. This hinges on one belief: whether the programmatic stack is mostly fee extraction on top of commodity mechanics, or whether the intermediaries are doing trust work that agents can't cheaply replicate yet. Mole is right that the mechanics are commoditizable, and he's proven it runs. He's overselling how fast trust, reconciliation, and money movement follow. The council leans toward "the direction is correct, the timeline is soft, and the first casualties are pure resellers, not The Trade Desk." What to verify before moving budget: can an agentic transaction clear payment and resolve a delivery dispute without a human, and will a verification provider actually stand behind a signal delivered over MCP. Until both are yes, this is a pilot line item, not a channel.

Prediction: No independent agentic ad-transaction platform (PubX, Scope3, or a peer) will be clearing more than a low-single-digit-percent share of any major premium publisher's open-market programmatic revenue by the 2027 upfront season (May–June 2027). The mechanics will be proven well before then; the money movement will still be the bottleneck.

Confidence: Medium. Plumbing is easy; payment and dispute rails are not.

Why: Mole has already shown the hard technical part works. Agents call OpenRTB, reconcile in Prebid, use price priority in the ad server, all open source. Yet the volume is still a few thousand dollars a day across 3,000 sites, and the payment float is being handled by a separate company (Scope3) precisely because money movement isn't solved. The clearest signal in his own telling is that Experian shipped MCP servers for agent integration and reports zero usage, which is what adoption looks like when the rails exist but nobody trusts them with real budget yet. Verification liability and delivery-dispute reconciliation are unsolved, and no publisher CFO routes material open-market revenue through a startup's payment float before those are settled. Getting agentic trading to real share by mid-2027 would require verification providers, payment rails, and holdco comp structures all to move in eighteen months, and holdcos face the exact innovator's dilemma (short-term revenue miss) that Mole himself flags.

Revisit by 2027-06-30: We're right if agentic direct-negotiation platforms remain sub-5% of open-market programmatic revenue at the top premium publishers heading into the 2027 upfronts. We're wrong if any major premium publisher reports agentic workflows clearing north of 5% of its open-market spend by then.

One more thing worth saying plainly. The pure-reseller SSP call is the safest thing in this episode and still worth stating: an intermediary that only passes bids and takes margin, with no direct publisher data relationship, has no answer to a machine that replicates the pass-through for free. That business was already under pressure. Agents just name the date.