The news
A new piece in Martech.org from Allen Martinez argues that data sovereignty and decision authority are two separate infrastructure problems — and most martech stacks have only solved the first one. The core stat: 90.3% of companies report using AI agents, but only 6.3% have them fully integrated into their marketing stack. That's not a tooling gap. That's a governance gap.
Our take
Here's the failure mode: a team stands up an AI agent, connects it to their CDP or CRM, gives it read access to customer data, and calls it a day. The agent is "in production." Then it offers a discount tier that hasn't been approved, or commits to a service level that requires legal sign-off, or fires a sequence that should have had a human in the loop. Nobody catches it until the customer is already holding the email.
The instinct is to patch at the tool level — add a guardrail in HubSpot, add a review step in Salesforce, tell the chat agent to escalate anything that sounds like pricing. That works once, in one system, for one scenario. Three months later a different agent in a different part of the stack does something different and unauthorized. You're playing whack-a-mole with your own automation.
What Martinez is pointing at — and the real complexity in agentic GTM work — is that decision authority has to be designed as its own layer. Not bolted onto your CDP, not configured per-tool, but defined upstream: what can this agent do, in what context, with what constraints, and when does it hand off to a human. That's not a prompt engineering problem. That's an orchestration architecture problem.
The 84-point gap between "using AI agents" and "AI agents in production" exists almost entirely here. Teams can spin up agents. They can't yet govern what those agents are allowed to decide.
So now what?
This is the right week to pressure-test your agentic setup with three questions:
- What can this agent commit to on behalf of your company? If you don't have an explicit answer, the agent has an implicit one — and it's probably wrong.
- Where does data access end and decision authority begin? Map it. If those boundaries live only inside individual tools, you have patchwork, not governance.
- What's your escalation path? Every agent that touches a customer-facing workflow needs a defined handoff condition before it goes anywhere near production.
Giving an AI agent access to your data without defining its decision rights isn't automation — it's delegation without a job description.