How Agentic AI Is Automating Campaigns on Google Salesforce and Meta

What agentic marketing actually means for workflows

Google’s AI Max, Salesforce Agentforce, and HubSpot’s Prospecting Agent all launched or expanded in 2026 with the same basic promise: hand over more of the campaign execution to software, and get better results with less human input. That promise is real in some places and overstated in others, and the gap between the two matters a lot if you are deciding where to invest your team’s time and budget.

Agentic marketing, as a category, means AI systems that do both recommend actions and take them. A traditional optimization tool surfaces a bid suggestion; an agent adjusts the bid, rewrites the ad copy, and expands the keyword set without waiting for a human to approve each step. That shift changes what marketers actually do all day. It does not eliminate the job, but it does move the work upstream, toward goal-setting, data governance, and quality control, and away from tactical execution.

The honest version of this story is that these three platforms are at very different points on the autonomy spectrum. Google AI Max operates closest to full automation inside paid search. Salesforce Agentforce is a configurable platform that can be highly autonomous or barely autonomous depending on how you build it. HubSpot’s Prospecting Agent, despite the marketing, is mostly a drafting assistant that still requires a human to select a contact before it does anything. Understanding those differences is the starting point for any rational adoption decision.

How Google’s AI Max replaces Dynamic Search Ads

Google announced AI Max for Search at Google Marketing Live 2026, framing it as an optimization layer that sits on top of existing search campaigns rather than a replacement campaign type. [12] In practice, it does much of what Dynamic Search Ads used to do, including crawling your landing pages to generate headlines and expanding your keyword targeting to queries you never explicitly bid on, but it goes further by rewriting final URLs and selecting ad formats dynamically. [1]

Google’s April 2026 blog post confirmed that AI Max is also expanding into Shopping campaigns, adding text customization, Final URL Expansion, and automatic format selection for retail advertisers. [3] The AI Brief feature, also announced in that update, gives advertisers a plain-language summary of what the system changed and why, which is Google’s attempt to address the “black box” complaint that has followed Performance Max since launch. [1]

The cost question is where AI Max gets complicated. Digiday reported that one year into AI Max’s rollout, media buyers are seeing real CPC increases, with some accounts absorbing higher spend without proportional ROAS gains. [4] Google’s position is that broader match and URL expansion reach higher-intent queries that manual campaigns miss, so the higher CPCs are justified by better conversion quality. That argument is plausible in theory, but it is very hard to verify in practice when the system controls both the targeting and the reporting. [5]

From what I have seen in accounts running AI Max alongside legacy exact-match campaigns, the system consistently pulls budget toward broader queries. That is not always wrong, but it does require tighter negative keyword management than most teams are used to running. If you opt in without auditing your exclusion lists first, you will spend the next quarter cleaning up irrelevant traffic that the AI decided was “close enough.”

Marketing Brew noted that Google is pushing AI Max as the successor to manual search optimization, not just a supplement to it. [2] That framing is worth taking seriously. Google has a long history of deprecating manual controls once an AI alternative reaches sufficient adoption, and AI Max looks like the next step in that pattern.

Salesforce and HubSpot launch their own AI agents

Salesforce’s Agentforce platform is not a single product. It is a framework for building autonomous agents that run inside the Salesforce ecosystem, pulling from CRM data, workflow rules, and external signals to take action without waiting for a human trigger. [11] Salesforce has released domain-specific versions including Agentforce Operations, Agentforce Sales, and industry editions for public sector and life sciences. [6] There is no formally branded “Agentforce Marketing” SKU in current Salesforce documentation, which matters if you are evaluating it for a marketing use case specifically.

The Agentforce Prospecting Agent, described in Salesforce’s Sales release notes, does the most relevant work for revenue teams. It researches target accounts using signals from Salesforce data and the web, then surfaces that research to sales reps rather than sending outreach autonomously. [9] Salesforce’s Summer 2026 release expanded these capabilities further, with tighter integration between agent actions and Salesforce Flow. [18] Moderna selected Agentforce for AI-powered customer engagement, which gives some signal about enterprise readiness, though the specifics of that deployment are not fully public. [16]

HubSpot’s Prospecting Agent sits inside the Breeze AI suite and works differently than most of the marketing around it implies. The HIC Global implementation guide is direct about this: the agent “is not a tool that independently finds leads or sends emails automatically” by default. [7] A user selects a company and contact in the CRM, and the agent then researches that contact and drafts an outreach message. That is useful, but it is a long way from autonomous prospecting. HubSpot offers 28 days of free access to the agent for evaluation, which suggests the company is still building adoption rather than charging at scale. [14] Pricing beyond the trial sits inside the Breeze AI credit model, which Resolve247 analyzed as adding meaningful cost for teams running the agent at volume. [15]

The Hublead guide describes a more autonomous configuration where the agent can research and send messages with less human review, but that mode requires deliberate setup and carries obvious risks around message quality and compliance. [10] Most teams will want to run the agent in assisted mode for at least a few months before letting it send anything without approval.

Meta opens its platform to third-party agents

Meta’s position in the agentic marketing story is structurally different from Google’s or Salesforce’s. Rather than building a single AI agent product, Meta has been opening its ad platform APIs to third-party agent frameworks, which lets tools like Agentforce or custom-built agents read campaign data and make changes through the Marketing API. This is less a product launch than a platform posture, and it has significant implications for how enterprise teams manage Meta campaigns at scale.

Google Marketing Live 2026 made clear that Google sees agentic campaign management as a first-party capability it wants to own. [12] Meta’s approach is closer to infrastructure: provide the API surface, let the ecosystem build on top of it, and capture the ad spend regardless of which agent placed the buy. For marketers, that means Meta campaigns are increasingly manageable through the same agent frameworks they use for CRM and email, which reduces the number of separate tools a team needs to operate.

The tradeoff is accountability. When a third-party agent makes a targeting or budget decision inside Meta’s system, the audit trail runs through two vendors instead of one. If something goes wrong, whether a brand safety issue or a budget spike, figuring out which system made the call takes longer than it would in a native tool. That is not a reason to avoid the integration, but it is a reason to build explicit logging into any agent workflow that touches paid media.

How marketing team roles and skills must evolve

The skills that made a strong paid search manager in 2022 are not the same skills that make a strong one in 2026. Manual bid management, keyword research at the individual match-type level, and ad copy A/B testing are all being absorbed by AI systems faster than most job descriptions have caught up. What remains, and what actually gets harder, is the work that sits above the campaign layer.

Prompt engineering for ad systems is one emerging skill, though “engineering” overstates it. What it really means is knowing how to write asset descriptions, audience signals, and campaign briefs that give AI systems enough context to make good decisions. Google’s AI Brief feature is a step toward making that feedback loop visible, [1] but the input side, what you tell the system before it starts optimizing, still requires human judgment about brand voice, competitive positioning, and audience intent.

Data governance is the other skill gap that is widening fast. Agentforce agents run on Salesforce data, which means the quality of your CRM records directly determines the quality of the agent’s output. [11] HubSpot’s Prospecting Agent has the same dependency: if your contact records are incomplete or stale, the agent drafts messages based on bad information. [7] Teams that have been tolerating mediocre data hygiene will find that agentic tools make that problem much more visible, and much more expensive.

The role that is genuinely at risk is the mid-level specialist who executes tactics but does not set strategy. Campaign coordinators who spend most of their time on manual optimizations, reporting pulls, and copy variations are doing exactly the work these systems are designed to automate. The teams that will adapt well are the ones that move those people toward analysis and experimentation rather than letting the headcount reduction happen passively.

A framework for adopting new agentic AI tools

Adoption sequencing matters more than most vendors will tell you. The instinct is to turn on every AI feature at once and see what improves, but that approach makes it nearly impossible to attribute results or diagnose problems. A more defensible path is to run one agentic tool against a controlled baseline before expanding.

For Google AI Max specifically, the practical starting point is a campaign audit before enabling Final URL Expansion or broad match automation. [5] Negative keyword lists need to be current and comprehensive, because the system will test queries that manual campaigns never touched. Running AI Max on a subset of campaigns while keeping legacy exact-match campaigns live gives you a real performance comparison rather than a before/after that conflates seasonal changes with AI impact. [17]

For Salesforce Agentforce, the configuration question is where to set the human-in-the-loop threshold. Salesforce’s platform allows agents to take action autonomously or to surface recommendations for human approval before executing. [13] Starting with approval-required workflows and loosening that constraint as you build confidence in the agent’s output is a lower-risk path than going fully autonomous on day one, especially for outbound communications where a bad message reaches a real person.

HubSpot’s 28-day free trial for the Prospecting Agent is genuinely useful for evaluation, but the metric to watch during that period is not volume of messages sent. [14] It is reply rate relative to your current baseline, because that is the only signal that tells you whether the AI-drafted outreach is actually better than what your reps write manually. If reply rates do not improve, the agent is saving time but not improving outcomes, which changes the ROI calculation significantly.

Across all three platforms, the governance question that most teams skip is logging. Every action an agent takes should be recorded with enough detail to reconstruct why it happened: what data it used, what rule or model triggered the action, and what the outcome was. That logging is both for compliance. It is how you catch the cases where the agent is technically following its instructions and producing results that do not match your actual goals, which is the failure mode that is hardest to see until it has already cost you real money.