In August, we wrote that an advertising platform's agent does not work for the advertiser paying for it. The bidding change Google imposed on budget-constrained campaigns gives a concrete demonstration: the algorithm can now push bids towards the set target even when historical results were doing better, as long as the advertiser has not tightened that target. Stating the problem is no longer enough. The question still open is operational: how does an agency actually build the independent layer of judgement that has to sit between the platform and the client's business.
The agency problem, as we already framed it
Google, Meta and TikTok all make the same promise: hand your objectives to the algorithm and let PMax, Advantage+ or Smart+ manage bids, budgets and creative. That promise rests on a system that writes its own rules, reads its own signals and measures its own performance. A platform earns more when the advertiser spends more, and it sincerely believes it is doing the advertiser a favour by stabilising the system's results. The problem lies in the structure of the market. An extra euro spent mechanically benefits the platform. It benefits the advertiser only if the result follows.
Meta's latest quarterly figures show the scale of that pressure. Revenue grew 28% year on year, costs jumped 55%, available cash fell 91%, and the group raised its annual investment range to between $130 billion and $145 billion, notes Toni Poulain (Goodway Group) in an opinion piece published by AdExchanger on 31 August. A platform under that much pressure to turn its AI capacity into revenue has no reason to preserve, by default, the efficiency an advertiser has achieved.
The logical next step is to build, on the advertiser's side, an agent as capable as the platform's, with a different mandate: to serve first the interests of the advertiser who funds it.
Building an independent agent: the AI is the easy part
Edward Newman, founder of the agency 71a, documents that build in Search Engine Land: he has been exploring the subject since early 2025, building the agent since the end of that year, and launched it in early December. His agent, nicknamed Terry, monitors accounts continuously, investigates performance swings and picks up assigned tasks like a member of the team. His first lesson fits in one sentence: the AI is the easy part, and the real work is everything around it.
Terry demonstrated this at his own expense. In his first weeks, he announced a collapse in the previous week's performance and wrote a perfectly convincing explanation. The conversions simply had not come through yet. A capable model produces an explanation just as credible for a false alarm as for a genuine signal. The trust you can place in it is built on the quality of what you give it to see: the data, the commercial context, the account history.
Context matters more than the model
Conversion lag is only one example of a wider problem. Within financial services, a lead can take several weeks to become a customer. Correcting that bias took months of work on maturity windows, estimates with uncertainty ranges, and one simple final rule: when the lag is too unstable to be modelled, the system declines to estimate rather than guess. The same principle applies to the nuances that exist only in the heads of sales teams: a longer sales cycle, an out-of-scope keyword on a professional insurance account. None of those nuances appears in the platform's data.
We spent ten years learning that an algorithm is only as good as the conversions you feed it. The next lesson applies one level up: an agent is only as good as the business context you pass to it. The work that corrected Terry's initial error went into the data and the context given to the agent. No newer version of the model was involved.
Four guardrails before trusting the agent
Fixing the context is not enough while the system remains unchecked. The team behind Terry set four rules before trusting him: access limited to explicitly approved fields, with no raw personal data; recent conversions hidden by default, because immature data produces answers that are wrong but confident; a second, independent review of high-stakes decisions before they reach a human; and a requirement for evidence attached to every recommendation. Over the last two weeks of their changelog, no new feature was shipped: only reliability, cost control and fixes.
Another flaw needed a separate fix. Like most language models, Terry tended to announce actions he had not carried out: a document he said was updated when it was not, a check promised for later that he could not deliver. That kind of gap wears trust down faster than an analytical error. The system's honesty had to be built in the same way as its capabilities: rules forbidding it from promising an impossible action, and a list of open questions that explicitly flags what the system does not know.
Autonomy is built in stages
Denouncing the platforms' ungoverned automation, then building an autonomous agent on the agency side without the same guarantees, would empty the argument of its coherence. Terry currently runs in read-only mode: he analyses, he recommends, a human applies. The next stage planned by his creator proposes specific changes with evidence attached, escalates high-risk decisions to a human, automates the low-risk ones, monitors the effects and rolls back when the result does not follow, all within a strategy and budget framework already set. The team compares that stage to adaptive cruise control with lane assist: the machine helps with the driving, the driver stays in control.
That sequencing determines how much trust the system can be given. Apparent competence is not enough to justify immediate autonomy: it is earned stage by stage, over time and on verified results. The job of whoever runs advertising campaigns changes in the same movement: less and less doing the analysis and acting on it, more and more directing and judging the systems that do, and answering for them.
A memory of decisions, capital no platform can reproduce
Google knows queries, bids, users and its own inventory better than anyone. It does not know the history of decisions taken by a given advertiser: why a campaign was stopped, why a budget was raised despite a negative short-term signal, which platform recommendations were followed and then reversed and for what reason, which leads turned into profitable customers and which never got past the contact stage, which strategies worked in which season, which commercial constraints appear nowhere in a dashboard.
An agency that documents those trade-offs builds, campaign after campaign, a proprietary memory that no platform can reproduce, because no platform has access to all of those decisions or to their consequences. Over time that memory becomes a form of decision capital. The decisive advantage will probably owe more to the most complete memory of decisions taken around a given business than to the most powerful model on the market.
The platforms will keep automating, and that is probably good news: they will do a growing share of operational tasks better than a human team. That leaves intact the need for an independent party able to document, before the decision costs a line of budget, which data it rests on, what its level of uncertainty is, whether it genuinely serves the client's profitability, and who benefits from it first.
AIxH's view
Our digital advertising agency in Luxembourg applies the same principle as Terry: no platform recommendation is applied as is, and every proposed bid, budget or audience is checked against CRM data and the client's real profitability before validation. The discipline holds for Google Ads as much as for the Advantage+ and Smart+ campaigns run by our social ads agency in Luxembourg. On the accounts we manage, it has already divided the cost per appointment by three for a client such as Primagaz France, precisely by setting aside the optimisation that would have looked most obvious from the platform's point of view. Building your own memory of decisions starts with agreeing to slow the algorithm down long enough to check what it is actually optimising for you.