How do you use agentic AI in advertising without losing the human connection that builds trust? Let agents run the plumbing and keep people on meaning. Agents can plan, pace, optimise and report within limits you set. People should approve the consequential decisions, own what the brand says, and plan from how real audiences actually talk.
This is no longer theoretical. Digiday reports that holding companies and independents are launching AI agents for planning, buying and measurement. Dentsu's Carat is building a planning and buying agent for upfronts with NBCUniversal, FreeWheel and Newton. Tinuiti has launched a Bliss Point MCP server that opens its measurement layer to AI agents. Clients, meanwhile, are wary, citing reports of rogue agents and hacks. WPP Media's answer: "governance and human accountability are central to agentic media buying."
What is agentic AI advertising?
Agentic AI advertising uses AI agents that carry out multi-step tasks, such as planning a schedule, buying inventory, shifting budget or compiling measurement, with a degree of autonomy. That is different from rules-based automation, which follows fixed instructions, and from generative AI, which produces content when asked.
The difference matters because an agent acts. It does not just suggest a budget shift; it can make one. Digiday's examples show the direction of travel: Carat's upfronts agent, Tinuiti opening its measurement to agents, and Stagwell relaunching its Assembly agency as Stagwell Media around an agentic system called Machine OS.
Why are clients worried about rogue agents?
Because the fear has a reasonable core. Anything that can spend money can spend it wrongly, and an agent can do it quickly. Digiday notes that client caution is fed by reports of rogue agents and hacks, and by a limited understanding of how AI works.
The industry's answer so far is reassurance through oversight. "Everything still requires human oversight," Dentsu's Rebekah Shalit told Digiday. FreeWheel's David Dworin described "this very strong need to have a human in the loop". WPP Media's Lauren Wetzel said financial or activation decisions require "explicit controls, approval thresholds, accountable human oversight."
There is a quieter fear beneath the headline one. It is not that the agent overspends. It is that the brand gradually stops sounding like anyone.
Why it keeps happening: meaning is made between people
Symbolic interactionism, a school of sociology, offers a useful lens. Its core idea is that people respond to what things mean rather than to the things themselves, and that meaning is built through interaction. A wedding ring is a band of metal until two people agree what it stands for.
Advertising works the same way. An ad is a gesture in a long conversation between a brand and its audience, and what it means depends on context, timing and tone. A bid adjustment is not a gesture. Nor is a budget reallocation. Agents can handle those well. What they cannot do is decide what the brand's next gesture should mean to these people, at this moment.
That is the insight in one line: authenticity thrives when technology takes a backseat to human connection. Automate the mechanics. Keep humans on the meaning.
Who should do what in an agentic media workflow?
The simplest way to reassure a client, and your own team, is to draw the line in advance.
| Task | Agent can own | Human must own |
|---|---|---|
| Audience planning | Clustering conversations into personas and sizing them | Choosing which tension to speak to |
| Budget allocation | Recommending shifts and pacing within limits | Approving moves above the agreed threshold |
| Bidding and optimisation | Real-time adjustments inside guardrails | Setting the objectives and the guardrails |
| Creative | Resizing, versioning and QA | The idea, the message and the tone |
| Reporting | Assembling data and flagging anomalies | Explaining what it means to the client |
| Brand safety | Flagging risks for review | The final call and the accountability |
How do I keep humans in the loop without slowing everything down?
- Set thresholds by consequence, not by task. Routine pacing can run freely. New audiences, new messages and budget moves above an agreed level need a person's sign-off.
- Write the guardrails in plain English. If a client cannot read what the agent is allowed to do, they cannot trust it.
- Give every agent a named owner. Accountability belongs to a person, not a system. Clients should know who that person is.
- Log every action and its reason. What changed, when and why. An agent you cannot audit is an agent you cannot defend.
- Keep the message human. Agents can version and resize creative. People approve what is said and how it sounds.
- Plan from real conversations. Start from how real audiences talk about the category, not from demographic boxes. Agents optimise towards whatever you point them at, so point them at people.
- Loosen controls as trust is earned. Review weekly at first. Widen thresholds when the logs show the agent behaving as expected.
How do I explain agentic media to a nervous client?
Start with roles, not technology. Stagwell's Slavi Samardzija told Digiday that client confidence depends on explaining the orchestrator's role, and that agencies will have more senior people guiding agents, in "teams whose shape has changed".
A simple structure works. Say what the agents do. Say what they are not allowed to do. Name who signs off, and on what. Show how the client will see every change. Then show them the audience you are planning for, in that audience's own words.
That last step is where human connection becomes practical. Defy Monitor, for instance, clusters audience content into personas and tensions automatically, so the plan starts from real conversations rather than demographics, and agents working from that plan optimise towards something recognisably human. You can see how it connects to agents and analysis across the Defy platform.
Waiting is not risk-free either. Carat's Carrie Drinkwater told Digiday that "if we don't start now, we'll be behind." The agencies that win will not be the ones with the most autonomous agents. They will be the ones clients trust to hold the controls.
