A new title on the same broken process changes nothing
The phrase "manager of agents" has become shorthand for where marketing is headed. The premise is appealing and mostly correct: instead of executing every task by hand, marketers set direction, define standards, and judge what agents produce.
Marketers shift from execution to direction, and from doing to deciding.
But most of the conversation stops at the human. It treats the manager of agents marketing operating model as a mindset upgrade — learn to delegate, write better prompts, get comfortable reviewing instead of doing. That framing is incomplete, and the incompleteness is expensive.
Promoting a marketer to "manager of agents" does nothing if the agents have nothing reliable to manage against. A manager of people inherits an org with payroll, role definitions, performance data, and a handbook. A manager of agents who inherits a pile of disconnected tools and a brand guide trapped in a PDF is managing a team with no memory and no shared truth. The title is the easy part. The operating model underneath it is the work.
The market is selling autonomy and quietly skipping the foundations
Scan the current pitches and a pattern emerges. Vendors describe a future where agents execute marketing tasks end to end and the human simply approves. The seductive version goes: stop managing channels, start managing outcomes; stop building workflows, let agents build them. One vendor frames it as the shift from a tech stack to an operating system where
instead of managing tools, you manage strategy; instead of building workflows, AI agents build and execute them; instead of checking dashboards, you get intelligence delivered to you.
It sounds clean. It skips the hard part. Autonomy is the headline; the foundation that makes autonomy safe is the fine print.
Analysts who have looked closely keep landing on the same caveat. McKinsey, studying agentic marketing workflows, found that the binding constraint is rarely the model itself.
Teams must confirm that agents can technically integrate with required systems—data platforms, content repositories, and activation platforms—since system interoperability, not model design, is often the limiting factor.
The same research notes that marketers will have to oversee
data quality and schemas, content metadata, orchestration rules, and API governance that ensures agents operate safely and consistently.
That is not a story about a smarter chatbot. It is a story about plumbing. And the operating-model failure happens when organizations adopt the title without rebuilding the pipes.
The undercurrent: a manager's anxiety about output they can't vouch for
There is a quieter tension beneath the excitement, and it shows up the first time an agent ships something wrong. Any manager of people knows the feeling: you are accountable for work you did not personally produce. With agents, that accountability arrives at machine speed and machine volume.
Two failure modes recur. An agent pulls the right customer data but produces something off-brand — accurate audience, wrong voice, a claim legal never approved. Or an agent stays perfectly on-brand but aims the message at the wrong people, because it never had a trustworthy view of who the customer is. The first embarrasses the brand. The second wastes the budget. Both erode the trust a manager needs to actually delegate.
This is why "be a manager of agents" cannot be answered with training alone. A marketer cannot govern outputs they have no basis to trust. The operating model has to supply that basis. Practitioners describe the same realization from the field: as agents get more capable,
marketers need to think less like task managers and more like system designers, setting goals, constraints, inputs, and evaluation criteria — good prompts matter, but governance matters more.
Governance is not a slide. It is data and rules that an agent can read at the moment it acts.
What actually has to change: two foundations agents reason against
If the title is the easy part, here is the hard part stated plainly. Agents produce good marketing only when they can reason against two things in real time: a trustworthy view of the customer, and a trustworthy view of the brand.
The first is unified, identity-resolved, governed customer data. Not a quarterly export, not a segment someone built last month, but a live, queryable understanding of who each customer is and how they behave. Approaches built around the data warehouse keep that information in the organization's own environment rather than copying it into yet another proprietary store. This matters for an agentic operating model specifically because every copy is a second source of truth an agent can get wrong, and because much of the anxiety about autonomous systems is really anxiety about where sensitive customer data travels.
The second foundation is the one most teams overlook: operational brand knowledge. Brand guidelines, approved claims, voice and visual rules — structured not as a document a human skims but as a queryable layer an agent can check against before it generates anything. A static brand PDF is useless to a machine deciding, in the moment, whether a headline is allowed. This idea is gaining independent traction. Harvard Business Review's prescription for the agentic marketing org centers on what its authors call a
"brand code": a machine-readable knowledge base encoding brand strategy, customer insights, and business rules that both people and AI agents can act on.
Put the two together and the logic is simple. Data without brand knowledge is accurate but off-brand. Brand knowledge without data is on-brand but pointed at the wrong audience. A manager of agents needs both, wired in, before delegation is anything more than a hope.
Evaluation criteria: how to pressure-test a platform before you reorganize
Reorganizing the team is premature until the platform underneath can support delegation. Buyers should pressure-test a few things rather than take the autonomy demo at face value.
Where does customer data live, and does it move? If a platform requires customer data to leave the organization's infrastructure to power its AI, that is both a governance question and a second-source-of-truth problem. A warehouse-native architecture — where the customer data stays in the warehouse and the platform reads from it — sidesteps both. This is the model platforms like Hightouch built their Composable CDP around: unified, identity-resolved data kept in the customer's own warehouse rather than copied out. Is brand knowledge structured or decorative? Ask how the system enforces voice, visual rules, and approved claims at generation time. A useful signal: does the platform reuse approved, existing assets before generating something new? Hightouch's Content Assembly works this way — agents search existing asset libraries first, which keeps output on brand and shortens approval cycles. That is brand knowledge doing real work, not sitting in a folder. Can a marketer stay the manager? The point of the model is that the human directs and decides. The platform should let a marketer describe an outcome, review what comes back with the reasoning attached, and choose what goes live — with creative and legal still in the loop where judgment matters.Once a marketer is happy with a concept, the assets get routed to the creative team for review and editing, either in the platform or within existing workflows.
Autonomy that removes the human from approval is not a feature; it is a liability.
Does it orchestrate across the tools you already run? An agent that can draft a campaign but cannot execute it through your existing email, ad, and CRM systems just moves the bottleneck. Interoperability, as the analysts keep noting, is where these efforts live or die.What it looks like in practice: a churn play, end to end
The abstract argument gets concrete in a single workflow. Consider a retention program — the kind of work that historically required stitching data, logic, copy, and channels together by hand.
In an agentic operating model, the marketer starts with an outcome, not a brief: identify high-value customers at risk of churning and build win-back campaigns for email and push. From there the agents do the assembly. The system reads customer behavior over time and past lifecycle performance, decides which at-risk segments are worth acting on, drafts the audience, assembles on-brand content from approved assets, generates the email HTML, and can orchestrate the campaign through existing tools. This is roughly how Hightouch's Lifecycle Marketing Studio — which houses capabilities including AI Decisioning and Native Delivery — is designed to operate.
Notice what the human does and does not do. The marketer does not hand-build the segment or write five tools' worth of glue. The marketer sets the goal, reviews the proposed campaigns, kills the ones that miss, and approves what ships. That is the manager-of-agents loop working as intended — and it only works because the customer data and the brand rules were trustworthy inputs, not afterthoughts.
The feedback loop is the part that compounds. Each round of human judgment — this concept, not that one; this audience, not that segment — teaches the system the team's standards, so the next batch arrives closer to acceptable. The manager's job is less "fix the output" and more "tune the system that produces it."
What success actually looks like
The payoff of a working manager of agents marketing operating model is not that humans disappear from marketing. It is that human time moves to where judgment creates value.
The volume shift is real.
Teams will test 500 ads instead of 5, run 30 campaigns instead of 3, and do it in hours, not weeks — and they won't want to go back.
McKinsey reports comparable speed gains where the foundations were in place: in some content pilots,
a slow and manual process became a fast, data-driven system that increased end-to-end speed by four times versus traditional workflows.
But velocity is the visible metric, not the meaningful one. The meaningful change is what marketers spend their attention on.
Instead of taking tickets, chasing approvals, and stitching work together across tools, teams focus on the parts of marketing that actually benefit from human judgment: setting direction, defining standards, shaping creative systems, and deciding what's worth putting in front of customers.
Google's analysis lands in the same place — the skill set
must evolve from execution to governance, with marketers defining the goals and constraints for AI agents.
Success, in other words, is a team that has stopped being a coordination engine and started being a direction-setting one. That is only possible when the agents underneath are managing against real data and real brand rules.
The title follows the foundation, not the other way around
The temptation with any operating-model shift is to start with the org chart, because the org chart is the thing leaders control directly. Rename the roles, run a workshop on delegation, announce that everyone is now a manager of agents. None of it sticks if the foundations are missing.
The order is the opposite. First, give agents a trustworthy view of the customer — unified, identity-resolved, and governed, ideally without copying sensitive data out of the warehouse. Second, give them operational brand knowledge they can query at the moment they act, not a PDF nobody reads. Third — only then — reorganize the humans around direction and judgment, because now there is something safe to direct.
The marketers who thrive in this model will not be the ones with the best prompts. They will be the ones whose organizations did the unglamorous work of wiring data and brand knowledge into a system agents can reason against. The title is a consequence of that work, not a substitute for it.
For a fuller picture of how the role and the platform fit together, Hightouch's Agentic Marketing Platform is a useful reference point for what "managing agents" looks like when the foundations are actually in place.