One company. 120,000 brands’ worth of marketing, running autonomously. Okara published that number on June 11, 2026, in a Vercel case study 1 Source 1 Vercel, “How Okara Runs CMO Agents for 120,000 Companies,” June 11, 2026. that landed in every CMO’s inbox before lunch.
The reaction in most marketing leadership teams was predictable: a mix of genuine excitement and the low hum of FOMO. Someone forwarded it to the CTO. The CTO asked what it would take.
Here’s the question nobody in that meeting knew how to answer: does our stack even qualify?
Okara built on Vercel’s infrastructure. Their orchestration layer talks to their clients’ stacks via API. That means 120,000 companies’ marketing environments had to pass a structural test before a single autonomous campaign could run. Most stacks, if you ran the same test today, would fail it.
Here is the infrastructure checklist CMO agents actually require before they can run autonomous marketing workflows. Five prerequisites. If your stack passes them, you’re ready to pilot. Fail even one, and the agent hits a wall — spending budget on theater instead of automation.
What CMO Agents Actually Do (And Why Stack Architecture Matters)
CMO agents are not a chat interface that writes subject lines. They are multi-step autonomous workflows: scheduling campaigns, segmenting audiences, generating content variants, monitoring performance, adjusting budget pacing — the whole chain, running without a human in the loop for each decision.
They don’t operate through a UI. They operate through APIs.
The agent calls your CRM’s contacts endpoint to pull the audience. It reads from your analytics layer to check campaign performance. It writes to your CMS to publish the content variant. It triggers your email platform to send. Each step is a structured API call — not a mouse click.
Which means any system in the chain that lacks an accessible, documented API breaks the workflow. A data silo the agent can’t read is a dead end. Most mid-market marketing stacks were built for humans clicking dashboards, not autonomous processes calling endpoints.
For context on how multi-agent coordination works technically, see Multi-Agent Systems Grew 327% in Four Months — the architecture explains why closed SaaS platforms block agent coordination at the network layer, not just the UI layer.
The Five Infrastructure Tests Your Stack Must Pass
1. Open API Coverage Across Every System Agents Will Touch
Every touchpoint in the workflow — CRM, CMS, analytics, email platform, ad platforms — needs a real API. Not a webhook workaround. Not a Zapier bridge. A documented external API that an autonomous process can call without a user session.
Can an external process read contacts, write content, read campaign performance data, and trigger sends? Walk the workflow step by step. For each step, identify the API call required, then verify that call is available to an external caller outside the vendor’s ecosystem.
This is where lock-in shows up. Some platforms expose APIs to their own automation tools but not to external orchestrators. One closed leg, and the whole chain fails.
Practical check: map every workflow you want to automate, identify the API call each step requires, verify the call is available to an external caller.
2. A Unified, Queryable Data Layer (Not Four Separate Dashboards)
CMO agents make decisions based on data. If performance data lives in one system, audience data in another, campaign data in a third, and content data in a fourth — without a common schema — the agent can’t correlate signals across the full picture.
The requirement: a data warehouse or event layer that ingests from every source and is queryable via SQL or a structured API. One place where all the data lands. One query the agent uses to ask cross-system questions.
This is the single highest-failure-rate prerequisite, based on our own infrastructure audits with mid-market companies leaving HubSpot. Most stacks have four or five data silos with no single source of truth. The CMO who can’t answer “which blog post drove the most pipeline last quarter?” in one query doesn’t have an agent-ready data layer — they have a reporting problem the agent will inherit and amplify.
For how this data layer fits into a modern AI-native web stack, see AI Agents for Websites.
Practical check: can you answer that pipeline question from a single query in a single system? If not, your data layer isn’t ready.
3. Identity That Resolves Across Systems (The Contact Graph Problem)
A CMO agent personalizing outreach needs to know that one email address in your CRM is the same person as a visitor ID in your analytics and a subscriber ID in your email platform. Without that connection, the agent operates on fragmented data.
Cross-system identity resolution is an architectural property, not a lookup table. Either your stack has a contact identifier that resolves across every system the agent touches, or it doesn’t. If identity resolution is currently a manual CRM operations task, like someone deduplicating records and linking IDs, agents will compound that problem at scale.
Practical check: does a single contact ID resolve the same person across your CRM, analytics, and email platform without manual intervention? If the answer requires a spreadsheet, you have identity chaos.
Most teams discover their stack isn’t agent-ready from a vendor in a pitch meeting. Running the checklist yourself, before that meeting, gives you a different kind of leverage.
4. Workflow APIs That Accept Writes (not just reads)
Reading data is table stakes. CMO agents need to write: update contact properties, create CMS content, trigger sequences, pause campaigns, adjust bid strategies. This is where agents move from advisory to operational, and where most stacks expose a gap that wasn’t visible before.
Many platforms expose read APIs but gate writes behind their proprietary automation tools or workflow builders. It looks like a technical limitation. It isn’t. It’s a product decision. HubSpot’s Breeze agents run inside HubSpot’s own API surface; an external CMO agent from a third-party orchestrator can’t access the same data through those same APIs. Salesforce Marketing Cloud gates campaign writes behind Journey Builder. The platform keeps orchestration inside its walls so you keep paying for the automation layer. An external agent hits a read-only wall exactly when it could be most useful.
The audit question: for every system in scope, can an external API caller create, update, and trigger — not just read? If the answer for any system is “only through the vendor’s own automation layer,” that’s lock-in by design.
Stacks built on open-source CRM, headless CMS, and open analytics expose full read/write API access to external callers by default. There’s no proprietary automation layer to protect.
5. Governance Infrastructure — Approval Gates, Audit Logs, Rollback
CMO agents acting autonomously on live campaigns carry real business risk: wrong audience, wrong message, wrong budget pacing. The guardrails are not optional. The CTO who signs off on autonomous marketing agents without a governance layer is the CTO who gets called at 2am because an agent sent 40,000 emails to the wrong segment.
Before agents go autonomous on live campaigns, three guardrails must be in place. Human approval gates for irreversible actions: sends, publishes, budget changes. A full audit log of every agent action with timestamps and parameters. Rollback capability within five minutes.
Deloitte’s 2026 Tech Trends report projects that more than 40% of agentic AI projects will fail by 2027. 2 Source 2 Deloitte, “2026 Tech Trends.” The primary failure causes are legacy architecture and missing governance infrastructure. The failure mode isn’t the AI performing poorly. It’s the AI performing at speed, without guardrails, on a stack that wasn’t built to supervise it.
The Readiness Checklist
For a 30-minute architecture review, give your answers to these five prerequisites:
| Prerequisite | Passing State | Your Stack |
|---|---|---|
| Open API coverage | Every system has a documented external API | ☐ |
| Unified data layer | One queryable source of truth across CRM, CMS, analytics | ☐ |
| Cross-system identity | Canonical contact ID resolves across all systems | ☐ |
| Write-capable workflow APIs | External callers can create, update, and trigger — not just read | ☐ |
| Governance infrastructure | Approval gates, audit log, and rollback in place | ☐ |
Score: 5/5 means you’re ready to run a pilot. 3–4/5 means you have addressable gaps — sequence the infrastructure work, then move. 0–2/5 means replatform first. Agents on a broken stack don’t fail quietly; they automate at the speed of the problem.
If your evaluation of prerequisites 1 or 4 surfaced a vendor lock-in pattern, then 97% Deploy AI Agents. 29% See ROI. explains exactly why bolt-on vendor AI fails this test, and the five questions to ask before your next renewal.
Frequently Asked Questions
What is the difference between marketing automation and CMO agents?
Marketing automation (HubSpot workflows, Marketo programs) is rule-based: if X, then Y. CMO agents are goal-based: given an objective, the agent decides what steps to take, calls multiple systems via API, and adapts based on real-time results. The infrastructure requirements are fundamentally different because agents operate across system boundaries, not inside one platform’s workflow engine.
How long does it take to make a stack agent-ready?
It depends on which prerequisites you’re failing. Open API coverage and identity resolution are often a 4-to-8-week engagement — mapping gaps and selecting replacements for closed systems. A unified data layer is a larger infrastructure project, typically 2 to 4 months. Governance infrastructure can be layered on at the end. Companies that try to skip the sequencing end up building governance on top of a broken data layer, which doesn’t hold.
Do we need to rebuild our entire stack to run CMO agents?
Not necessarily. The checklist is diagnostic: some companies find they pass 4 out of 5 tests and need targeted work on one system. Others fail all five because they’ve centralized everything in a single closed platform. The replatform question only arises when the closed platform is both the data silo and the write-gated system, which is common in HubSpot-centric stacks.
Okara’s 120,000-customer deployment proves CMO agents aren’t speculative. They’re operating at scale today on production stacks. The question is no longer “will this technology work?” It’s “does our stack qualify, and do we know?”
In our experience migrating marketing stacks off proprietary platforms, the gap between “we run marketing automation” and “our stack is agent-ready” is almost always three missing things: a queryable data layer, write-capable APIs, and governance infrastructure. Most teams discover this from a vendor in a pitch meeting. Running the checklist yourself, before that meeting, gives you a different kind of leverage.