Marketing's agentic shift, mid-2026: three layers, one AI operating model

What are the three layers of agentic marketing?

The word "agentic" is doing a lot of work in marketing right now, and most of it is imprecise. A Google AI Max toggle is called agentic. AdCP (Ad Context Protocol) is called agentic. A desktop workflow tool that books your social posts is called agentic. These things are not the same. They operate at different points in the value chain, have different governance requirements, and fail in different ways.

Three layers. Different decisions at each one.

The operational error is treating all three as interchangeable acceleration tools. They are not. Channel agents run on the platform's rails, toward the platform's goals. Protocol agents reshape the economics of the entire buy/sell relationship. Workflow agents automate your internal tasks. Each layer needs a different governance response.

The leapbuzz position on this: the agency-execution layer (assembling campaigns, trafficking media, building reports) is the part that AI eats first. When the platform ships the co-pilot and the protocol ships the negotiation agent, the billable hours for those mechanical tasks shrink toward zero. What becomes more valuable, not less: the judgment about what to automate, the measurement infrastructure to verify it, and the design of the escalation boundaries. That is the consultancy work. State it plainly, because it is the answer to the "AI kills agencies" framing that floods every LinkedIn thread.

What did Google and Meta ship, and whose objective do they optimise?

Two significant platform releases in the five weeks before this post. Both are real. Both are worth activating. Both optimise toward the platform's objective function first.

Google Marketing Live, May 20, 2026. The headline announcement was AI Max for Shopping: a one-click toggle in Merchant Center that converts standard product feeds into conversational ad creative, targeting long-tail high-intent queries Google's own systems identify. Also announced: Ask Advisor, a unified AI co-pilot across Google Ads, Analytics, and Merchant Center that responds to plain-language questions and surfaces recommendations; and a Multimodal Asset Studio that generates text, image, and video creative from a single prompt via Gemini (blog.google, May 2026). The "one-click" framing in Google's own language is accurate technically. The implication for advertisers is that one click sets the platform's objective as the primary optimization signal.

Meta Business Agent, June 3, 2026. Announced at the Conversations London conference, Meta's Business Agent deploys across WhatsApp, Instagram, and Messenger. The cited capabilities include customer question-and-answer, product recommendations, appointment booking, lead qualification, sales closing without human intervention, and multilingual responses (Meta newsroom, June 3, 2026). The rollout targets businesses of all sizes; initially free, with paid tiers announced for the coming months. Meta ran extended business-messaging pilots in markets including India, Mexico, and Brazil before this launch.

The Google and Meta channel agents are both worth trialling, and this is where an AI performance marketing practice earns its keep: harnessing the platform agents for execution while keeping the objective function under your control. The activation checklist before you flip either switch: (1) confirm your measurement infrastructure can attribute outcomes independently of the platform's reporting; (2) set a spend floor below which you will not read statistical significance into the result; (3) design the escalation rule for any agent action that touches the customer relationship directly.

The protocol layer: AdCP and the collapse of execution margin

The Ad Context Protocol (AdCP) is a different category of development from either of the platform tools above. It is not a feature inside a walled garden. It is an open standard for agent-to-agent advertising, built on Anthropic's Model Context Protocol (MCP), governed by AgenticAdvertising.org (a 501(c)(6) pending membership organisation with four equally-weighted voting classes: brands, agencies, publishers, and ad tech).

The first agent-to-agent media buy under AdCP executed on October 16, 2025: real money, real inventory from LG Ads, real AI agents handling creative, targeting, and approval with human oversight. Since then, 99 or more companies have joined, including Yahoo, PubMatic, Scope3, Samba TV, and LG Ads (AgenticAdvertising.org, verified June 2026). AdCP 3.1 is in active development as of mid-2026.

The protocol works like this: a buyer briefs an agent in plain language ("reach eco-conscious car buyers on connected TV in the US this week within a defined budget"). The buyer agent discovers available inventory by querying publisher agents, negotiates terms, and transacts, all without a human trafficking a single insertion order. The publisher agent on the other side is doing the same thing in reverse: discovering qualified buyers, evaluating their briefs, accepting or declining. The exchange is bilateral agent communication, not an auction.

This is what "the agency is dead with AI" actually means at an operational level. The labour that media agencies billed for, specifically the assembly of media plans, the trafficking of creative, the management of insertion orders, collapses into protocol execution. What does not collapse: the judgment about which brief to write, the measurement of whether the outcome matched the brief, and the escalation when the agent's decision diverges from the brand's intent. Those three remain human work, and they are the higher-value work.

The practical question for a marketing leader in mid-2026 is not whether to engage with AdCP. It is at what pace. Early movers who integrate their brand briefs and creative assets into AdCP-compliant format get earlier discovery by buyer agents and earlier learning cycles. The window is not closing as fast as the SEO window closed in 2004, but it is directional.

What should you never let an AI agent do unsupervised?

Meta's Business Agent is capable of closing a sale without human intervention. The product is designed for this. The marketing copy leads with it. For a significant class of companies, particularly those operating in regulated sectors, that capability is a liability disguised as a feature.

The more useful framing for regulated-sector clients comes from Kachilu, a macOS desktop agent for marketing and sales teams launched on June 16, 2026. According to its launch announcement, Kachilu plans campaign workflows and performs browser-based actions autonomously, but pauses when human judgment is required. That pause is the product feature. The escalation boundary is not an afterthought; it is the engineering deliverable.

Three decisions that should never run without a human in the loop, regardless of which agent is asking:

1. Regulated product sales and lead qualification. Any agent action that creates a binding commercial relationship or advances a regulated sales process needs a human sign-off in the audit trail. This is not optional for banking, insurance, or any other MAS-regulated activity in Singapore.
2. Brand-sensitive response generation. An agent answering customer questions on WhatsApp is generating statements attributable to your brand. Off-brand or legally imprecise statements are a reputational and compliance surface. The agent needs a review cadence and a human escalation path for any query that touches pricing, policy, or complaints.
3. Budget reallocation above a defined threshold. Channel agents (AI Max, Smart Bidding, Performance Max) can move significant spend across placements within an approved campaign. Set a daily delta threshold above which no automated action takes effect without human review. The threshold number depends on your budget and risk tolerance; the threshold existing at all is non-negotiable.

The counterintuitive finding from regulated-sector engagements: clients who design the escalation boundary first, before activating any agent, get more from the agent. Because the agent's scope is defined, the measurement is cleaner, the audit trail exists, and the trust builds faster. The ones who activate first and govern later spend the first 90 days unwinding problems the agent created in the first 30.

Why independent measurement matters more now, not less

A specific failure mode is emerging with the rollout of channel agents: the platform grades its own homework. Google AI Max reports its own conversions in Google Ads. Meta Business Agent reports its own lead quality in Meta Business Suite. Both platforms have a structural incentive to show you good numbers, because good numbers keep the budget flowing.

This is not a new problem. Smart Bidding had the same issue in 2020. Performance Max had it in 2022. The difference in 2026 is that the agent is making more decisions at a higher speed, across more surfaces simultaneously, which means the divergence between the platform's reported performance and your actual business outcome compounds faster.

Two measurement tools that matter more in an agentic environment than they did in a manual one:

• Incrementality testing. A controlled experiment (holdout group versus exposed group) that measures the lift attributable to the agent's activity versus what would have happened without it. This cannot be run inside the platform's own measurement tools without significant methodological caveats. Run it externally, using your own first-party data as the ground truth.
• Marketing Mix Modelling (MMM). Statistical modelling that attributes business outcomes to marketing channels using historical data, without user-level tracking. In an environment where agents are optimising across multiple channels simultaneously, MMM gives you a channel-agnostic view of what is actually moving the needle. Our earlier post on cookieless MMM and Bayesian incrementality in APAC covers the methodology in detail.

The analytics and insights function becomes a strategic asset, not a reporting backoffice, in an agentic environment. The organisations that build independent measurement infrastructure before they scale agent activation get accurate data. The ones that accept platform reporting as truth at scale find out how wrong it was during the next budget cycle, when the business outcome does not match the dashboard.

For clients who want to understand what this looks like with real numbers: the anonymised cases at leapbuzz Results show the before-and-after measurement redesign in a banking context (6x conversion rate, 60% cost reduction over seven quarters) and a fintech programmatic context (74% improvement in efficiency, 40% reduction in cost per acquisition). Both involved independent measurement as the first infrastructure layer, before any agent was activated.

Do you need an agent operations function?

The conventional response to a new technology surface is to hire toward it. There was a programmatic trading desk phase (2012-2016), a data science hiring phase (2017-2020), a martech stack expansion phase (2019-2023). Each generated new roles, new cost centres, and eventually consolidation when the technology commoditised.

The agentic phase calls for a different structure. The work is not adding more people to run more tools. It is building a small function that governs which tools run, where they run, and under what constraints. Call it agent operations. The analogy is data governance: you would not deploy a data stack without someone owning data quality, access policy, and schema management. Deploying channel agents, protocol agents, and workflow agents without governance produces the same class of downstream problem.

What agent operations owns, specifically:

• The agent registry. Which agents are active, for which campaigns or workflows, with what approved scope. The registry is the audit trail.
• Escalation boundary design. For each agent, the documented set of conditions under which the agent must pause and surface a decision to a human, rather than acting.
• Measurement standards. The independent measurement methodology (incrementality, MMM, or both) applied to each agent's output. No agent operates in production without a measurement standard attached to it.
• Agent audit cadence. A regular review (weekly for high-stakes agents, monthly for workflow tools) of agent decisions against intended outcomes. The audit catches objective drift before it becomes a budget or compliance problem.

The team size for most mid-market organisations is two to three people with clear mandates across these four areas. This is not a department. It is a governance layer with teeth. The AI strategy engagement is the natural entry point for designing this function, the media integration workstream is where the channel agent governance typically gets built, and the day-to-day spend lives inside the performance marketing function the governance layer sits above.

One more implication worth stating plainly: the org chart argument against building agent ops (usually "we already have a media team for this") misses the structural change. The media team runs campaigns. Agent ops governs the systems that run campaigns. When the same systems handle the campaign and the reporting, someone outside the system needs to be asking whether the campaign is actually working. Agent ops is that function. Build it before the agents are running at scale, not after.

Website development is the other infrastructure layer that often gets overlooked in this conversation. An agent that drives a consumer to your site but lands them on a slow, poorly-instrumented page loses the conversion the agent earned. If your site architecture is not ready to receive agentic traffic and attribute it correctly, talk to us at leapbuzz about website development as part of the engagement.