Agentic AI is cheap and fast to deploy, and it breaks where nobody checks: what it knows, how far it can go, and whether anyone can see it. Five experts explain where marketers should look first.
Nearly every marketing team now has an agent roadmap. Fewer have asked what the agent will actually work with once it goes live. Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027, and MIT’s Project NANDA found that 95% of enterprise generative AI pilots showed no measurable return. Neither figure, on its face, points to weak reasoning.
Read side by side, the contributions from seven MartechView experts converge on one question: What separates an agent that works from one that merely demos? They come from adtech, loyalty, go-to-market and enterprise automation. They differ on tactics but agree on where the risk sits.
Is the architecture built for objectives or for scripts?
Anurag Gurtu, CEO of Airrived, starts with the money. UiPath, once valued at more than $35 billion in private markets, now trades about 70% below its peak, according to him. Torq, meanwhile, raised $140 million at a $1.2 billion valuation in 2026. Gurtu argues that older workflow platforms share a flaw: They start from “rules rather than objectives,” so they fail when conditions change. Investors now back systems that interpret intent, plan across tools and adapt to uncertainty, he says. He calls this an architectural rewrite, not an upgrade.
Lisa Sharapata, VP revenue growth, Blackboard, makes the same argument for B2B revenue teams. Account-based marketing relies on fixed target lists, she says, and those lists freeze a view of a market that changes almost weekly. Buyers also now ask private questions inside language models, where traditional intent signals can’t follow. In her model, teams define the conditions that make an account worth attention. Agents then scan the whole market for matches, pull back where interest cools and open up the long tail. People shift from deciding every move to setting goals and constraints.
What does the agent actually know?
Kenneth Rona, chief AI officer at JWX, says the first question in agentic advertising is not how good the agent is. It is “What does the agent know?” Foundation models are converging in capability and price, he argues, so the edge moves to the quality of the inputs. A buyer agent fed only domain, URL and category taxonomy makes decisions no better than the rules it replaced.
He lists seven dimensions that inventory data needs: provenance, viewability, ad environment, attention and outcome, context, brand safety, and salience. Each must be linked and queryable before a dollar is committed, not audited afterward. His example is the difference between a page tagged “automotive” and a scene tagged “car chase.”
A piece on customer data platform (CDP) memory layers makes the same point about customer data. An agent working from nightly batch updates may pitch an upsell to a VIP who filed a delivery complaint 12 minutes earlier. The authors describe such an agent as having “severe short-term amnesia.” Their proposed fix is a real-time CDP that returns a unified profile in under 100 milliseconds, enforces consent at the moment of decision and sets discount limits from live margin data. The piece also cites Moffatt v. Air Canada, in which a tribunal held the airline liable for what its chatbot told a customer.
How far should it be allowed to go?
Sharapata’s follow-up piece turns the autonomy question into policy. She lists five principles: clear decision boundaries, specialist agents that share context, real-time decisions instead of static segments, measurement designed before launch and humans in the leadership seat. Her rule for the first is that “autonomy is granted, not assumed.” Bid changes, creative rotation and send times sit with the agent. Pausing a campaign or changing brand messaging goes to a person. Boundaries widen as trust builds, and she frames them as brand policy, not technical configuration.
The CDP piece reaches the same place from the data side. It argues that guardrails should be deterministic, with opt-outs and commercial limits enforced in the data layer so they hold whatever the prompt says.
Where does the learning happen?
Greg Collison, head of product & design at Adobe Advertising, looks at what cheaper execution makes possible. Most sophisticated buyers stop at two or three demand-side platforms (DSPs) because each new one costs time, not because they lack ambition. He expects DSP-level agents for setup, optimization, troubleshooting and insights to lower that cost. Agencies could then build their own agentic operating systems that coordinate across platforms.
The larger prize, he says, is learning: Buyers could “taste and see” which platforms suit each objective instead of defaulting to familiar ones. He notes the industry is still early and expects these tools to mature over the next six to 12 months.
Can the agent see you at all?
Denise Holt, SVP, strategy, growth, marketing & research at Phaedon, flips the question to the brand’s side. McKinsey has forecast that agents could mediate $3 trillion to $5 trillion of consumer commerce by 2030. If a traveler’s assistant can’t read their points or status, the program’s value never enters the decision. The brand loses the sale and has no abandoned cart or lapsed-member signal to show for it.
Programs must be legible to agents, Holt says, which means readable, trusted and real-time without a custom integration for each assistant. She points to open standards such as the Model Context Protocol as one route. Her closing point is that structure is for the machine and recognition is for the human, and brands need both.
What the seven voices agree on
The experts disagree on tactics, from CDPs to agency operating systems to loyalty standards. They agree on where the work sits. The model is rarely the problem. The inputs, the limits and the visibility around it decide the outcome.
As teams plan their next agent rollout, five questions run through the collection:
- Does your agent work from live, linked data or from last night’s batch?
- Are its limits written as brand policy and enforced in the data layer?
- Is your inventory, audience or loyalty value described richly enough for an outside agent to act on?
- Are you set up to compare more platforms and options, now that execution costs are falling?
- Do your metrics show what the agent resolved, not just what it clicked?
