Agent Hub, Agentforce, and the Data Question: What Scales Agentic Marketing Automation

Within just a few weeks, the two largest CRM platforms have taken the use of AI agents in marketing to a new level. On July 23, HubSpot launched Agent Hub and Agent Builder into public beta as a central control layer for AI agents: a console where teams can see, manage, and build new agents in natural language across marketing, sales, and service. Shortly before, Salesforce had made its Agentforce commerce agents generally available — including connections to ChatGPT and Google Search. The direction is clear: AI agents are shifting from individual tools to infrastructure. All the more clearly, it shows what determines their productive use in practice — not the agents themselves, but the data foundation and orchestration underneath.

What the platforms actually built

What is interesting about HubSpot's approach is less the individual agent and more the promise of control. Agent Hub bundles pre-built agents, custom agents, and agentic workflows in one place; all access the same customer context in the CRM — deal history, contact data, conversation notes, purchase signals. Agent Builder lowers the entry barrier: business users describe in natural language what an agent should do, instead of configuring workflows technically. And: Custom agents consume credits, meaning they are priced and measured based on usage.

Salesforce pursues the same logic in a commerce context with its commerce agents: Shopper, Buyer, and Merchant agents take on defined roles in the buying process and are connected to external interfaces like ChatGPT and Google Search — meaning the point of customer interaction partially shifts away from owned channels.

Both announcements thus address an operational problem that is already visible in many organizations: agents working in isolation. The often-cited negative example from industry discussions: A sales agent contacts an existing customer in the same week that a service agent is handling their open complaint — and neither system knows about the other. This is exactly the coordination problem that the new control layers address.

The uncomfortable prerequisite: connected customer data

As different as the products are, they share a silent assumption: that the organization's customer data is complete, up-to-date, linked, and access-controlled. Industry coverage of the July releases rightly noted that many companies are deploying agentic AI on a data infrastructure that was never designed for it.

For banks and financial service providers, this weighs twice as heavily. Customer data there is typically scattered across core banking systems, CRM, campaign tools, and service channels; regulatory requirements for data protection and traceability add to this. An agent that serves product recommendations while the customer simultaneously has an active complaint is annoying in retail — in financial business, it damages trust and can become regulatory relevant.

The consequence: The data question comes before the agent question. Three building blocks have proven to be sustainable:

• A connected customer profile. Transactions, interactions, consents, and ongoing processes must be consistently available to all systems — otherwise, every agent operates on a different slice of the truth.

• Event-based timeliness. Agents react in minutes, not in monthly cycles. A data foundation that is only updated nightly or weekly makes fast agents slow — or wrong.

• Rules before autonomy. Contact pressure control, channel preferences, exclusion reasons, and escalation rules belong in a central decision logic, not in each individual agent. Orchestration means: The framework decides which agent is allowed to act and when.

Orchestration in practice: decision logic instead of channel thinking

What such a control layer looks like in practice can be described by the principle of the next best action: Instead of every channel and every agent pursuing its own triggers, a central logic evaluates per customer and point in time which action — offer, service note, advice impulse, or deliberate inaction — yields the greatest value. The agents thus become executing units of an overarching decision, not competing senders.

This also changes how teams work together: Marketing, sales, and service jointly define priority rules and exclusion criteria instead of coordinating retrospectively over contradictory customer approaches. And it creates the prerequisite for learning loops — if all actions and reactions converge in one place, it is possible to systematically evaluate which approach works for which customer segment. It is precisely this feedback loop that turns individual automations into a learning system.

Measurability becomes a disciplining factor

An underestimated side effect of the new platform generation: Agents are measured and priced. When custom agents consume credits, every automation gets a visible price tag — and needs a robust business case. That is healthy. It forces teams to quantify the benefit per use case: What conversion impact does the agent have? What manual efforts does it replace? What error costs does it avoid? Marketing organizations that can run these calculations will defend budgets — the others will bury pilot projects.

Added to this is the regulatory dimension: From August 2, 2026, the transparency obligations of the AI Act require that the use of AI in customer contact be disclosed and AI-generated content be labeled. Anyone using agents in customer communication needs processes that systematically ensure this — another reason to anchor control centrally instead of in isolated solutions.

What decision-makers should check now

1. Data audit before agent rollout. What customer data is currently available for automated decisions, how up-to-date is it, and where are the gaps and contradictions?

2. Plan orchestration as its own layer. Who coordinates agents across marketing, sales, and service? What central logic governs priority, contact pressure, and exclusions?

3. Start with a few, measurable use cases. One agent with a clear value proposition and clean success measurement beats five experiments launched in parallel.

4. Build in labeling and disclosure from the start. Integrate transparency obligations into processes and templates from day one instead of retrofitting them.

Conclusion

In July, HubSpot and Salesforce demonstrated where marketing automation is heading: from individual AI functions to orchestrated agent landscapes with a shared customer context. The platforms deliver the control layer — they do not deliver the data foundation. Whether agentic automation creates value or chaos is decided by connected, up-to-date customer data and a central decision logic. Organizations that create these foundations now can use the new generation of tools productively. Everyone else is just automating their inconsistencies — only faster.