How is the marketing leader's role changing in the AI Era? 12 lessons from Sunil Subhedar.

That was the premise behind a recent Futureproof Project session — Traction's community for marketing leaders navigating the shift to AI. More than 60 people showed up, including brand teams from Pepsi, GM, Talkspace, ADP, Atlassian, and the ANA, to hear a live conversation between Traction CEO Adam Kleinberg and Sunil Subhedar, an Entrepreneur in Residence (EIR) at Storm Ventures and an advisor in Traction’s ‘Liquid Workforce,’ who built the AI behind Canva's Magic Resize feature and most recently led growth marketing at Anthropic.

Here's what came out of it.

1. Lead with grace and patience.

Sunil's top-line advice wasn't a framework or a tool. It was grace and patience. Most marketing leaders he talks to have been handed a mandate — go and create AI adoption on your team — while still running the business and hitting the same numbers they hit last quarter. Nobody budgeted time for the part where you also have to learn how to design agentic systems. Understanding and acknowledging this reality will make life easier for everyone.

2. Three critical roles every marketing team needs.

Sunil sees three roles CMOs need onboard to build successful AI teams:

None of the three functions alone. According to Sunil, "If you have those three working in tandem, you start to see and observe what's happening at this larger level of where AI adoption is being successful, where the blockers are happening, and then where your role is to go and find out how to unblock them."

3. Why do you need a harness, not a wrapper?

Sunil pulled up a diagram midway through the conversation — data, reasoning and orchestration, execution, and observability and learning, all wrapped around the model itself. Most AI solutions being delivered by marketing teams are what Sunil described as “wrappers.” A wrapper is a set of skills bolted onto an existing workflow. It looks like progress until the underlying model changes, and then it breaks.

A “harness” is everything built around the model — a data layer, a reasoning and orchestration layer, an execution layer, and, critically, a learning loop that grades outcomes and feeds them back in.

"If you don't think about everything around the model — whether it's Claude or OpenAI or Gemini, “Sunil warned, “you are restricting the capabilities to where the model sits today."

4. Building evals to create a learning loop.

The layer almost nobody builds is the learning loop. Most teams collect feedback. Almost none of them grade it.

Most teams skip evaluation because it sounds technical, but Sunil's version starts simply with a scorecard that defines what "great" actually looks like for a specific piece of work. His team paired it with a learning file that logs the corrections you make so the agent stops repeating them. Start with pass or fail to eliminate subjectivity of grading, not a 1-to-10 score, and force you to name the non-negotiables instead of hiding behind subjectivity. Once that's solid, you can share it across your team as a plugin, and now your content marketer, your social lead, and your email writer are all pulling from the same brand voice instead of each reinventing it in their own thread.

One long-form content agent his team built needed more than 25 quality examples before the output actually held up. "The problem with the scoring is that it becomes more subjective," Sunil explained. "If you have a pass fail, it becomes very clear if it actually met the bar or not."

5. Ontology: from content to context to meaning.

Adam framed the arc of the last three years in three words: "2025 was the year of content. 2026 is the year of context. And 2027 is going to be the year of meaning." Traction is building an AI strategy and context engine built on marketing and industry-specific ontologies to give marketers data meaning.

Ontology is the study of concepts and relationships. For example, wine is the parent of red, white and pink. Dry is an attribute of any of those. Oaky is a characteristic that could describe Chardonnay or Cabernet, but not for Sauvignon Blanc or Beaujolais. In a vector database, rules like this are defined — that’s the difference between AI that guesses at the next word and AI that reasons because it understands your business.

Sunil agreed, and pushed it further — the more comprehensive the ontology underneath an agent, the more effective that agent is at actually doing the work, because it's no longer guessing at what a term or relationship means. It's working from a rules engine instead of pure probability.

Traction is building a context engine built on top of bespoke industry marketing ontologies. If you’d like to learn more about it, let’s talk.

6. A simple human-in-the-loop framework.

Sunil's rule for when to pull a human into the process: stay in the loop on any foundational layer until you're genuinely confident it holds up on its own, and weigh the risk. Customer-facing work earns more scrutiny than internal drafts. "A human has to be in the loop for every single foundation layer until you are very confident that the foundation layer meets its requirements," he said.

An example of this is how we’ve built the Strategy Engine in TractionOS, our own version of our AI operating system. It automates some steps like competitive analysis and message matrix development, but stops the bus when it comes to things like customer and stakeholder interviews that work far better when humans do them, but every step that has shown variance from the expected output has a gate to ensure a subject matter expert gets their eyes on it.

7. Protecting teams from “AI sprawl.”

Left ungoverned, every team member ends up building their own private experiment, and none of it compounds into anything larger. Sunil's fix at Anthropic was almost stubbornly simple: he made every person on his team block off an hour a week to build something with Claude Code, no exceptions. "It was just that time and energy," he said — plus office hours afterward to look at what people had built and figure out what was worth sharing.

Having the team surface and socialize their experiments is critical. AI sprawl is another term for the steaming hot mess of uncontrolled growth of AI tools, models, integrations and agents across an org without governance or visibility. Without intentional effort, disconnected AI across your team is inevitable.

8. Calculate the “click-based tax.”

Before his team touches AI, Sunil has them map every micro-task involved in getting something done — what he calls the "click-based tax."

"What are all the clicks that you have to do to get something done? If you put that down on a paper, it becomes really clear — it's like, oh my god, these are so many things that I'm doing unnecessarily."

One team's version of this: reparenting 400 campaigns in Salesforce went from a three-week ticket process to a 15-minute task. The same instinct applies to scale — designing for the audience size and market footprint you'll have in two years, not the one you have today.

9. Manage adoption plateaus.

One CMO of a well-known job community site raised something a lot of the room recognized — his creative team jumped on AI fast, but his director-level cohort has had the hardest time adjusting, which surprised him.

Sunil's take is that plateaus aren't a sign of failure; they're teams processing. When a team gets stuck, the answer usually isn't grinding harder on the same tool. "The real question is, how much juice is worth the squeeze," he said. "Is this incremental improvement on this AI model really worth the time, or should we go do something and explore something else?"

The biggest blocker to adoption usually isn't the technology — it's fear.

One attendee, who leads a young in-house creative team, put it plainly: "I have a very young in-house creative team... they're afraid of it, and they were never taught it." Sunil advised building a culture where wins get shared and celebrated openly, rather than one where being more efficient just means someone quietly gets handed more work: "It's how you create the culture in your team to feel like it's a collaborative and all-lifts-all-boats [environment], versus someone feeling like they're going to be left behind or could be taken advantage of."

10. Measure real outcomes, not adoption metrics.

"AI adoption rate" doesn't tell you anything useful, according to Sunil. In fact, many tech companies that tried to force AI adoption by measuring token usage reversed course when employees started gaming the systems and spending loads of money on token fees.

What holds up in front of leadership is a before-and-after: hours saved, a KPI that actually moved, a task that used to take three weeks now taking fifteen minutes. "That's something that's very tangible that you can show," he said. "But it's not going to happen overnight. This is a marathon, not a sprint."

11. Get off local machines.

An agent that only runs on one person's laptop isn't a system — it's a demo.

Sunil walked through how Anthropic moved skills off individual machines and into shared, governed infrastructure through GitHub and cloud deployment, with a central team providing the access and permissions that non-technical teams can't get to on their own. "A culture which recognizes we need to figure out how to give as much of the tooling and capabilities so that everyone can benefit from it is what Anthropic prides itself on," he said.

Adam shared that he’s had to learn his way around Github in the past year and gave a power-user tip for those trying to build and manage AI systems: tell Claude to explain what you need to know like you’re an 11-year old.

12. Autonomous Agents vs Process Automation

Fully autonomous, customer-facing agents are still too risky for most marketing use cases, in Sunil's view — but internal process automation is already viable today.

What's coming next is agents that behave more like co-creators, surfacing information before anyone asks for it, and eventually, agents coordinating with other agents. "The joke is that you're now managing agents, and your agents are managing agents," Sunil said.

Key Mindset Shifts for Marketing Leaders

What's next.

Traction's Marketing System Diagnostic applies this same click-based, systems-thinking approach to a live assessment of how AI is actually being used across an organization's people, platforms, and processes — surfacing the next AI opportunity grounded in reality, not guesswork. Traction has also built an AI marketing operating system that puts context, judgment, and a real learning loop into a working system marketing teams can use.

Join the community to get an invite to be part of the conversation.

FAQ.

What roles make up a successful AI marketing team?

There are three essential roles in an AI marketing team: you need a Builder to develop the agents and solutions, a Solutions Architect to be the plumber who handles the engineering and workflow integration, and a Technical Program Manager to observe usage, remove blockers, and scale what's working.

What is Marketing System Design?

Traction has a consulting practice that helps brands align the people, platforms, partners and platforms to succeed in an AI world. This practice includes a Marketing System Diagnostic and development of a custom AI operating system for brands. Unlike Deloitte or McKinsey, Traction leverages practitioner consultants who are actual subject matter experts who understand how work gets executed. In this case study, you can see how one $20B retailer will save over $15,000,000 in the next three years.

What’s the difference between a harness and a wrapper?

A wrapper is a superficial interface or set of Skills that just links back and forth with an LLM. A “harness” is everything built around the model — a memory layer, a reasoning and orchestration layer, an execution layer, a learning loop that grades outcomes and feeds them back in.

What is TractionOS?

Traction’s AI Operating System is our own custom implementation of Traction’s AI strategy and context engine that we use to operate our own business. The underlying platform includes a dynamic context engine with industry-specific marketing ontologies and a portal for AI agents that handle internal operations, marketing execution, and a Strategy Engine. The Strategy Engine is a workflow automation of AI agents that manage everything from discovery through brand narrative through customer journeys, media architecture and customer experience design recommendations.

What is AI sprawl?

AI sprawl is the steaming hot mess of uncontrolled growth of AI tools, LLMs, integrations and agents across an org without oversight or visibility. Organizations experience AI sprawl when AI adoption takes place without a coordinated strategy to unify context and application.

What is a click-based tax?

The measure of the cost of all the clicks that a team has to do to complete a task. It helps identify where workflow automation or AI agents can help reduce effort and make the greatest impact in a business.

What is an ontology?

Ontology is the study of concepts and relationships. In a vector database, rules are defined that help AI understand the meaning of data. It’s the difference between AI that guesses at the next word probabilistically and AI that reasons because it understands your business. Benefits are reduced hallucinations, higher reasoning capability and more consistent responses from AI.