Episode #8
#8: The non-negotiables in an AI-ready tech stack, with Brendan Powers
When an agent gets something wrong, the model is the first thing anyone blames. Bumping it up a tier is the cheapest available fix, which is why it happens so often and why it so rarely works. The fault is almost always upstream, in the prompting or in the data sitting underneath it. An agent that needs an account's industry can't do much with the half of your accounts where that field is empty. In this episode of Revenue Architects, John Lloyd is joined by Brendan Powers, Principal Go-to-Market Operations and Engineering Manager at ZoomInfo, on what a stack has to look like before an agent can run on it without a human checking the output. His answer starts underneath the agent: enrichment data, CRM data, and product usage in the warehouse, because the play for a customer who barely logs in isn't the play for one about to blow through their credits. Then the working rule. Make as much of the agent deterministic as you can, and hand it data instead of asking it to decide. When it breaks, you want to know which layer failed, and a heavy prompt doing five jobs will never tell you. Then Salesforce. Rep time in the CRM is trending down and to the right, and Brendan is building custom apps in his terminal without being an engineer, but it's still the source of truth and still the trigger for almost every agent they run. Klarna is the cautionary tale for anyone in a hurry. John and Brendan also get into why shadow AI is usually an enablement gap rather than a discipline problem, why the account notes rotting in someone's Evernote are suddenly worth something, and why the job is wide guardrails, not gatekeeping.
