The job is changing faster than the job description

The consensus forming across the industry is comfortable and a little vague: marketers will become "conductors of an AI-powered orchestra," setting strategy while agents execute.

The most valuable marketing skills will no longer be about technical proficiency in a specific tool but about the ability to think critically, ask the right questions, and effectively manage a team of AI agents.

It's a tidy story, and it's mostly right about the destination. It's wrong about what the work actually is.

Managing AI agents is being sold as an orchestration skill—learning to delegate, supervise, and approve. The harder, less glamorous reality is that an agent is only as good as what it knows. Point a capable agent at messy data and stale brand rules and you get fast, fluent, confidently wrong output. The marketer's real job in an agentic team isn't directing traffic. It's curating the context agents reason from, then judging whether the result clears the bar.

That distinction matters because it changes what teams should invest in. If managing agents were mostly about workflow design, the winning move would be buying the slickest agent builder. If it's about context, the winning move is fixing the data and brand foundations agents depend on—work that's far less exciting and far more decisive.

"Orchestration" is the part everyone gets right and overweights

There's real substance to the orchestration view. A new operational layer is emerging to run agents at scale.

"Just as DevOps reshaped software deployment in the 2010s, AgentOps will reshape AI operations in 2026," said Joao Moura, CEO of CrewAI

, describing a function that

will sit between engineering and operations and be responsible for managing fleets of AI agents, monitoring cost, reliability, and compliance.

Large enterprises are reportedly building internal environments to design, test, and launch multi-agent workflows while keeping governance in place.

Guardrails are part of this too.

Guardrails are the rules and constraints that ensure agents operate within brand guidelines, compliance requirements, and budget limitations—marketers can set frequency caps, define approval thresholds, and establish "no-go" zones so that even as agents operate autonomously, they do so in a safe and controlled manner.

Approval thresholds, frequency caps, eligibility rules—these are genuine management levers, and teams will use them daily.

But notice what guardrails and orchestration assume: that the agent already has good raw material to work with. A frequency cap stops an agent from over-messaging. It does nothing about whether the agent knows the customer just churned, or whether the creative it generated used last season's logo. Orchestration governs how agents act. It says nothing about whether they're acting on the right information. That second question is where most agent deployments will quietly fail.

Why most agent programs stall in the same place

The evidence that capability isn't the bottleneck is already in.

Nearly 90 percent of CMOs are experimenting with AI use cases across the marketing process, but less than 10 percent have captured value across end-to-end workflows.

The gap between experimentation and value is rarely the model. Models are abundant and improving. The gap is the substrate they run on.

Two failure modes recur. The first is bad data. Most teams are sitting on plenty of it, but it's

scattered, messy, and hard to use—and humans can't parse all of it fast enough to drive personalised decisions at scale.

An agent inherits that mess. If customer records are fragmented across tools with no resolved identity, the agent's "decision" is a guess dressed up as precision.

The second failure mode is the one CMOs name most often about generative output. In conversations with dozens of marketing leaders, the same complaint surfaces repeatedly:

general-purpose AI gets colors wrong, hallucinates products, and just doesn't meet the brand bar.

A model that writes beautifully but invents a product you don't sell, or renders your brand in the wrong palette, isn't a productivity gain. It's rework. The output is fluent and unusable at the same time.

These two failures point at the same root cause from different directions. Data without brand knowledge produces output that's accurate about the customer but off-brand. Brand knowledge without data produces output that's on-brand but aimed at the wrong person. Either way, the marketer ends up fixing what the agent should have gotten right—which is not what "managing agents" was supposed to feel like.

What managing agents actually requires: two foundations, not one

If the real work is supplying context, then the practical question becomes: what context, and where does it live? The useful answer has two parts, and skipping either one is why programs stall.

The first foundation is unified, governed customer data. Agents need a complete, identity-resolved view of each customer to make individual-level calls. Approaches built on the data warehouse address this directly. A warehouse-native, or composable, customer data setup keeps customer data in the company's own cloud warehouse and activates it from there, so the warehouse stays the single source of truth rather than spawning another copy in a vendor's silo. This matters for agents specifically because they reason best against fresh, complete data—not a partial profile that was synced hours ago into a proprietary store.

The second foundation gets less attention and may matter more: operational brand knowledge. Brand guidelines, approved claims, voice rules, visual standards, product catalogs, and legal constraints have to exist as something an agent can query in real time, not as a PDF in a shared drive. Hightouch, which helped define the composable CDP category and now positions itself as an agentic marketing platform, describes this as a brand context layer—a structured store agents reason against so generations stay on-brand.

This brand context layer integrates with a company's existing creative assets in DAMs, ad platforms for past campaigns and performance, brand guidelines, and more.

The mechanism is what fixes the "wrong colors, hallucinated products" problem at the source rather than catching it in review.

Put plainly: managing agents means owning these two layers. A marketer who controls the customer data foundation and the brand context layer can hand an agent a goal and trust the result. A marketer who controls neither is reduced to proofreading. The skill that separates them isn't prompt-writing. It's curation.

What the work looks like day to day

Consider a concrete loop, because the abstraction can hide how different this is from running campaigns by hand. In lifecycle marketing, an agent might decide that

a lapsed customer should receive a winback 10 percent off offer, but only if they haven't already re-engaged elsewhere, only on SMS based on prior response, and only in a time window when they're historically likely to convert.

That single decision draws on resolved identity, behavioral history, channel preference, and eligibility rules—context, not cleverness.

This is the model behind Hightouch's AI Decisioning, which sits inside its Lifecycle Marketing Studio.

Marketers define the options, constraints, and goals; reinforcement-learning agents continuously experiment across those options to determine what performs best for different customer situations—who to contact, with what content, in which channel, at what time, and when not to send at all; decisions execute in existing tools; and every decision is measured against a control or holdout group.

The marketer's role is upstream and downstream of the agent, not inside its loop: set the goal, define the guardrails, supply the content, then read the lift.

The management surface is deliberately explicit about control.

Marketers stay in full control by authorizing what actions the agent can take or not—defining what's allowed, what content to use, and setting thresholds to balance performance with send volume, so AI optimizes within the brand's strategy.

That's the day-to-day reality of managing an agent: you're not writing the messages, you're defining the space the agent is allowed to operate in and judging the results against a number.

It also reframes the feedback loop everyone hand-waves about. The point isn't that the system vaguely "gets smarter." It's specific:

decisioning uncovers hidden patterns and correlations that humans can't detect and feeds those insights back to marketers so they can continuously optimize messaging, timing, and outcomes.

A managed agent surfaces what's working and why, and the marketer decides what to do with it. The learning loop runs through a human on purpose.

What good looks like

The payoff for getting the foundations right is measurable, and the numbers cluster around the same theme: the agent does the execution volume, the human keeps the judgment. One retailer using decisioning agents on a loyalty program saw results fast—

PetSmart, with more than 70 million loyalty members, used AI Decisioning to increase incremental salon bookings by 22% within just three weeks.

The marketing team didn't hand-build those journeys; they set a goal and let the agent optimize against it.

The replacement of manual work is equally telling.

One customer replaced 60 manual marketing journeys with an agentic lifecycle system that outperformed previous efforts by more than 30%.

Sixty journeys is a lot of human orchestration to retire. What replaced it wasn't a smarter org chart of agents—it was a goal, a set of constraints, and the data and content for the agent to work from.

On the creative side, the same pattern holds once brand context is in place. One brand

reduced campaign production from four weeks to one, increasing click-through rate by 13% and conversions by 15%.

The acceleration came from generating on-brand variations at volume—which is only possible because the brand rules were structured for an agent to use, not buried in a style guide. Across these cases the marketer's value moves to where it's hardest to automate: taste, judgment, and deciding what's worth doing.

The skill to build is curation, not just supervision

The honest version of "every marketer becomes a manager of agents" is that the management is mostly about inputs. McKinsey's account of the emerging division of labor is precise about it:

human workers focus on prompting and managing agents, reviewing output, and enhancing ideas with instincts and insights drawn from years of industry and market experience.

Reviewing output and enhancing ideas are downstream skills. Curating what the agent knows is the upstream skill that determines whether the downstream review is light or brutal.

For buyers evaluating platforms, this suggests a sharper set of questions than "does it have agents." Ask where customer data lives and whether agents work from a resolved, current view or a stale copy. Ask whether brand knowledge is queryable context or a document someone pastes into a prompt. Ask whether outcomes flow back into the system fast enough for agents to learn, or whether the feedback loop runs through so many disconnected tools that learning is effectively offline. Those questions test the foundations. The agent features are the easy part to demo and the hardest part to make trustworthy.

The marketers who do well in this shift won't be the ones with the most agents or the cleverest prompts. They'll be the ones who treated context as the product—who built a clean customer data foundation and a structured brand layer, then let agents run inside well-defined goals.

The future of marketing won't belong to teams chasing shiny new tools; the teams that win with agentic AI will build readiness: organized data, clear processes, well-mapped workflows, and a culture that embraces testing and learning.

Managing AI agents, in the end, is mostly the discipline of deciding what they're allowed to know—and that work starts well before the first agent is ever switched on.

For a deeper look, the composable CDP is a useful starting point.