Why Marketers Are Adopting AI Instead of Operationalizing It

Introduction

Ninety-seven percent of executives deployed agents in the past year, but only 29% are seeing ROI. Isaac Ferreira, VP of Growth Systems and AI at Shift Paradigm (and a MarTech Podcast sponsor), argues the gap isn't model capability — it's that most organizations treat AI like another application instead of an operating system. With 25+ years building systems across defense, healthcare, and enterprise technology, Ferreira breaks down how marketing teams move from task-level automation to an agentic foundation that compounds.

Tasks Are the Wrong Unit of Improvement

Most teams are optimizing at the task level: clean up my inbox, draft this email, generate this piece of content. Ferreira's point is arithmetic — small scope means a small ceiling on returns. Processes are where efficiency, quality, margin, and customer experience actually change. He frames adoption as four levels: AI as assistant, AI inside workflows, AI as an employee, and finally AI as an operating system ru ing across all workflows. Most companies are stuck at the first two.

What an Agentic Foundation Actually Contains

At enterprise scale, the foundation is a contextual base plus governance — role-based access control, ISO certifications, SOC 2. The context layer answers: what am I trying to do, what are the limits on what I can do, what operations am I allowed to perform, what context surrounded past decisions, what was the performance, and what are my improvement goals. Get that in place and you can layer any agent on top: pull data into the warehouse, train on context, deploy. The alternative — buying a SaaS tool or coding an app per task — produces 300 underused tools and no governance.

The Recursive Learning Loop

What makes AI an operating system rather than automation is that it improves itself. Ferreira's loop is goal, context, measure, performance, levers — Lean Six Sigma for AI. Without context, "AI otherwise is just guessing," and the outputs are guesses too. The orchestration is the easy part. As he put it, "The hard part's getting the data there." Human approval of learnings should be the default early on, and permanently for high-risk work: a supply chain agent might report 3% throughput improvement that a human recognizes came from suppressing sales. Low-risk processes can iterate freely; customer-facing ones need human continuity in the loop.

Marketing Shifts From Journeys to Signals

Ferreira reframes personalization as a marketing operations problem, not a MarTech one. As AI intermediates what buyers see based on contextual relevance, predefined journeys lose their grip and marketers start operating on signals — and creating conditions for signals to appear. That inverts the segmentation process: start broader within your ICP, push several message variants, then build sub-segments from who responds and what attributes they share. Practically, that looks like intent tooling such as 6sense, or four email versions carrying four value propositions, then following each responder down their own value chain rather than a standardized campaign. Adobe is furthest along at the platform level, layering agentic AI across its creative, real-time CDP, and activation stack.

Fund It by Use Case, Not by Data Project

The budget conversation is where most of this dies. Ferreira's advice is to refuse the fight entirely: "I do not now, and I would never go to a company and say, centralize your data." Instead, pick one use case — operationalizing campaign development inside your project management system — and unify only the data that use case requires. Build the base layer, add the agents, show value, then snowball into the next use case, like the content supply chain, on top of what you already built. On ROI, he pushes back on dollars-per-FTE as the metric. Real returns show up as revenue, throughput, speed to market, and margin, with client engagements producing 10 to 15 point margin improvements without cutting headcount.

Key Takeaways

Stop scoping AI to tasks and start scoping it to processes. Build the contextual foundation — data with meaning, permitted functions, rules, and performance history — before layering agents on top. Instrument a goal-context-measure-performance-levers loop so workflows improve rather than just repeat, and keep humans approving learnings wherever risk is real. Fund the whole thing one use case at a time, and measure it in margin and throughput, not headcount reduction. As Ferreira puts it, "AI should be a tool for growth, not a tool where we fire everyone and stop growing."