Infinity Campaigns, Zero Leaks

On 15 June 2026, Hightouch published its manifesto for the agentic CDP: background agents that generate hypotheses, investigation agents that test the evidence, verification agents that challenge every finding, all connected directly to the warehouse without copying data. On 16 June, Databricks launched CustomerLake, an agentic CDP whose Profile Agents and Campaign Agents build the Customer 360 in the lakehouse and run what it calls infinity campaigns, continuous, agent-driven engagement loops. Databricks' CEO put it plainly: marketing stops being a series of campaigns and becomes a continuous loop, agents that constantly analyse, decide, and act on every customer in real time.

This is not a positioning war between two vendors; it is a category settling. Hightouch's AI Decisioning agents have made over ten billion marketing decisions since 2024, with an average published lift around twenty-five per cent and named results at WHOOP and PetSmart, and the company raised $150M at a $2.75B valuation on the back of it. Braze answered by acquiring OfferFit for $325M. Adobe shipped its Agent Orchestrator; Salesforce rebuilt Marketing Cloud as a full-funnel agentic product. And the gap the agents close is real: Salesforce's own State of Marketing finds three quarters of marketers have adopted AI while 84 per cent admit their campaigns still feel generic. Pre-planned, segment-blasted marketing is exactly what the agentic CDP exists to replace.

In October 2025 Google open-sourced an official MCP server for the Google Ads API. In February 2026 Amazon Ads opened its own MCP server to beta, campaign creation, reporting, and account operations as tools an agent can call. In April, Meta launched Ads AI Connectors, an official MCP server with write access from day one: agents can create and edit ads, ad sets, and full campaigns in natural language, and by July Meta had opened it to any developer with an app. In May, TikTok announced its Ads MCP Server and an Agentic Hub, a marketplace of AI skills, so that marketers can, in TikTok's words, connect their own AI agents directly to the ads platform to plan, launch, and optimise campaigns.

The platforms' own automation is racing the same direction: TikTok's Smart+ and Symphony Agent, Meta's Advantage+, Google's AI Max and its new Ask Advisor agent spanning Ads, Analytics, and Merchant Center. Press reporting has Meta working toward fully automated ad creation and targeting. Put the two layers together and the arc from strategy to planning to creative to activation is, for the first time, a software pipeline: your agents reason over your first-party data and your brand, then operate the channels through governed sockets. The campaign brief becomes a prompt; the media plan becomes a loop.

What this does to the operating model is the point. The work moves up a level: from executing campaigns to setting goals, guardrails, and taste, and from channel hands to signal engineering. The figures in our chart are modelled, but the direction is unambiguous, and it is why we tell clients that the marketing organisation they are hiring for in 2027 looks nothing like the one in their plan.

Agentic buying is only as good as the conversion signal it optimises against, so every channel now runs a server-side conversion API: Meta's Conversions API, TikTok's Events API, Snap, Pinterest, and LinkedIn CAPIs, Google's Enhanced Conversions. In March 2026 even Netflix announced its own Conversion API for its ads business, and Amazon pairs its Conversions API with the Amazon Marketing Cloud clean room, which only answers queries aggregated over at least a hundred users. With Google retiring most of Privacy Sandbox in late 2025 while keeping third-party cookies on life support, the industry has effectively standardised on this pattern: first-party, server-side, hashed, minimised.

The trap is on the device. Under Apple's App Tracking Transparency, embedding a third-party ad SDK that combines your users' data with other apps' data is tracking, even if you never use it for that, and Apple holds the developer responsible for every SDK in the app. Drop the TikTok or Meta SDK naively into your iOS app and you have bought yourself the ATT prompt, and the opt-out rates that come with it. The privacy-centric route is Apple's own: SKAdNetwork and its successor AdAttributionKit attribute installs and re-engagement through signed, aggregated, crowd-anonymised postbacks, and, in Apple's words, you don't need to use the AppTrackingTransparency prompt to use them.

The same discipline applies server-side. Hashing an email before sending it to a CAPI is data minimisation, not anonymisation, and the EDPB's 2024 guidelines put pixels, URL tracking, and unique identifiers squarely inside the ePrivacy consent rule: moving tracking to the server does not move it out of the law. Consent Mode v2 is now the price of Google measurement in Europe. So the winning architecture is precise per channel: consented, hashed, minimal conversion events out; aggregated attribution back; nothing that lets a platform reconstruct who your customer is. That is signal engineering, and after the Reset it is a core marketing skill.

Follow the data. The agentic CDP unifies your customer file, identity, behaviour, propensities, and hands autonomous agents the run of it. Hightouch is a US corporation; it offers EU workspace regions, and its own documentation is honest that it cannot verify where your destinations store your data. CustomerLake is in preview with no published EU-residency terms. None of this is a criticism of two excellent products; it is a jurisdiction fact. A US-headquartered provider can be compelled under the CLOUD Act regardless of where the servers sit, a point Microsoft's own France leadership conceded under oath in 2025, and the EU-US Data Privacy Framework that papers over the gap is once again before the CJEU.

European regulators have already shown what happens when martech meets that gap: the CNIL ruled Google Analytics transfers unlawful in 2022, Meta took a €1.2B fine over EU-US transfers in 2023, and TikTok was fined €530M in 2025 over transfers to China. An agentic CDP that decisions your customers from a jurisdiction you cannot audit is the same exposure, with autonomy added: you will have automated the leak.

The sovereign adoption pattern keeps the upside and closes the hole. Use Databricks or Hightouch for what they are best at, the composable, no-copy context layer over a warehouse you govern, EU-resident. Run the decisioning itself, the models that read the customer file and choose the next action, on sovereign inference inside your perimeter. Let signals leave only as the consent-clean, hashed, minimal events of chapter three. The agents get everything they need to run infinity campaigns; no one outside your control ever holds the memory of who your customers are.

Audit your signal paths first. List every ad SDK, pixel, and tag across app and web, flag what forces the ATT prompt or fires without consent, and replace the client-side sprawl with first-party, server-side collection behind a consent mode. This is the fastest risk you can retire, and it usually improves the data.

Re-plumb conversion signals channel by channel: SKAdNetwork and AdAttributionKit on iOS, hashed and consent-gated CAPI events to Meta, TikTok, Netflix, Amazon and the rest, clean-room queries where a platform offers them. Design each feed to the minimum the algorithm needs, and document the legal basis as you go; the machine optimises just as well without the identity.

Then wire the agentic loop with guardrails: platform MCP sockets behind a gateway you log, budgets and brand rules as hard constraints, a human approving strategy while agents run the middle. Stand up the decisioning on sovereign inference over your own warehouse for one lifecycle programme, retention is the classic proof, and measure the lift. From there, expansion is a routing decision, and your competitors are still planning campaigns by hand.