AI agents have become everyday work buddies for many marketers.
Last month, I joined 30 Days of AI in Demand Gen, a peer cohort organised by Eric Linssen. What I enjoyed most was comparing notes with marketers who were already using AI in their actual jobs, with real data, real restrictions and real consequences.
Some of the builds were seriously impressive.
Blake Cohlan vibe-coded, for his startup, his own version of 6sense: a real-time account journey view that brought signals from Salesforce, AdRoll ABM and HubSpot together.
Angela Ferrante walked us through a workflow designed around a “no lead left behind” principle, finding and investigating leads that had fallen through the cracks before a failed handoff could make them disappear.
But at the same time, I also heard from many peers that they are only using AI in very narrow parts of their work: automating research, enriching leads, preparing reporting, drafting content or solving one recurring operational problem.
That difference was rarely about how sophisticated someone was with AI.
Sometimes their role was specialised, so one part of the funnel was exactly where automation created the most value. Sometimes privacy policies and system permissions limited what an agent could touch. Sometimes AI adoption had simply moved faster than leadership's willingness to give it access or operational authority.
My own situation gives me unusual freedom.
As the solo marketer behind Smartify Marketing, I can redesign my working environment end to end. I decide which systems agents can access, what context they share, how work passes between them, where review happens and which decisions must come back to me.
So I started asking a different question:
What would happen if I designed my AI agents as a marketing team?
My answer is let's try it.
And after rounds and rounds of iteration (together with uncountable frustration and breakdowns), my system now looks like this: AI agents handle substantial parts of my research, editorial planning and data analysis. They review one another, hand approved work into the next stage and bring decisions back to me when human judgement is required.
There is a fashionable term for some of this: harness engineering. I tend to think about it in a less technical way.
It feels like managing a small, but extremely fast team.
There is one caveat tho. Marketing runs on trust, judgement, taste, commercial context and emotional connection. In my own workflow, I would estimate that at least half of the work is deliberately human. Client conversations, final creative judgement, publishing, campaign authority, budget decisions and decisions involving private data all stay with me.
The interesting problem is how to make artificial intelligence and human intelligence work together without wasting either one.
At the beginning, I created a completely dysfunctional workplace. I was the control-freak manager, and the agent was the employee who said yes to everything.
I gave it a task, watched the output, corrected individual mistakes and added another instruction whenever something went wrong. The prompt got longer and I stayed involved in almost every step.
That works for a one-off task. It scales badly once several agents need to move work from research to planning to review.
The change came when I started asking the questions I would ask when designing a human team:
Who owns the job?
What decisions can they make themselves?
What context do they need?
What are they never allowed to do?
Where do they need me?
That sounds obvious, but it changed how I build agents.
The first rule is clear ownership.
In my workflow, I mainly hand over three areas of work to AI: market research, editorial planning and data analysis.
So I delegate 3 lead agents who are responsible for taking the work from a defined brief to something ready for human review.
They all have clear ownership. The research agent owns the evidence-gathering process. The editorial agent works against the brand knowledge base and content calendar. The analytical agent writes and runs programs against advertising, demand, CRM and Deal data.
The second rule is to brief the outcome more carefully than the method.
There is a strange temptation to respond to disappointing AI work by making the prompt longer and longer. One exception becomes a paragraph, the paragraph becomes a policy, and soon the agent is carrying an instruction manual before it even sees the task.
OpenAI's current model guidance says to “favor leaner prompts”: remove repeated instructions, state each instruction once and keep the tools and context relevant to the task. Its own agent-first engineering practice describes the same idea as “enforcing invariants, not micromanaging implementations”.
That is close to how we brief a capable employee. We explain the outcome, provide the context they cannot reasonably infer, define the safety boundaries and describe what a finished deliverable must contain. Then we leave room for them to choose the route.
For agents who can't judge the boundaries by themselves, I now apply that principle through role design.
Instead of making one agent carry every concern, I split an important job into specialised roles. The lead agent concentrates on delivery. Other sub-agents handle narrower responsibilities around it.
Each gets the context relevant to its own job.
The brief fixes the destination and the boundaries. The route remains open to the agent.
For example, for important tasks, I usually divide the work between three roles.
The worker owns the job. It receives the operating context, relevant tools, the outcome I need and a clear definition of what a good deliverable should look like.
Within those boundaries, I give it considerable freedom.
The guard has a much narrower remit. It watches the things I do not want delivery pressure to override: privacy, permissions, sensitive information, unsupported claims and actions that cross an authority boundary.
The reviewer judges the finished work against the brief, evidence and quality standard.
I deliberately keep the reviewer separate from the worker. It does not need the worker's entire history or every intermediate decision. I want a cold read from something that has no reason to defend the route already taken.
Anthropic describes a similar evaluator-optimizer pattern, where one model produces work and another evaluates it against defined criteria.
In my workflow, the reviewer sends its findings back to the worker. The worker has more operating context, so I let it assess the criticism instead of automatically accepting every suggestion.
If the feedback improves the work, it revises.
If the suggestion conflicts with the evidence or another requirement, it can explain why and flag the disagreement.
Then the revised work goes through review again.
By the time something reaches me, the agents have already done the kind of internal challenge I would expect from two strong colleagues.
Once several agents were working together, chat windows became the wrong management interface.
I did not want to watch every tool call. I did not want to be asked for approval every few minutes. I wanted the equivalent of a manager's desk: show me what needs a decision, keep the supporting evidence within reach and tell me when something is ready.
So I vibe-coded an approval interface.
A task reaches me after the worker has completed it and the reviewer has challenged it. I see the proposed deliverable, the evidence I may need, unresolved findings and any disagreement that requires human judgement.
I can approve it, reject it or send it back with feedback.
Approval then changes what the system is allowed to do next.
An approved research pack can move into the knowledge base and then feed the editorial planning. An approved topic and evidence set can move into drafting. An approved analysis can move into a budget recommendation.
The next agent continues from a trusted checkpoint rather than rediscovering the whole problem. The agents have different jobs, but their work connects, and I no longer have to manually carry context between every stage.
My role sits mainly at those checkpoints. I inspect what the team produced, resolve disagreements and decide whether the work is good enough to move forward.
This structure makes me comfortable giving agents substantial autonomy in areas where their capacity genuinely exceeds mine.
Market research is the clearest example.
I define the question, geography, period, source hierarchy, exclusions and evidence standard. The agent can then search a much wider field than I could cover manually, follow useful branches, compare sources, surface disagreements and assemble an evidence set.
I do not tell it which search query to run next.
My work happens around the search. I decide what question is worth answering and, once the evidence comes back, I inspect the important source paths, methodology and definitions before deciding what I believe.
The second workflow is my content calendar.
An editorial calendar looks simple until it has to remember the brand.
I built a governed brand knowledge base.
It distinguishes confirmed first-hand experience, external evidence, published material, drafts and writing rules. My agents can retrieve from those different categories according to the job they are doing.
When a topic enters the editorial process, an agent can pull relevant research, previous claims, examples and related pieces. It can check for topic overlap and structural repetition. It can show me what the brand already knows before I decide what the next article should say.
This is another area where I am happy for the system to work aggressively.
Retrieve everything relevant. Compare it. Organise it. Find contradictions. Tell me when I have made essentially the same argument before.
What I do not outsource is the reason the article should exist.
An agent cannot contribute an experience the brand never had. It cannot decide which personal observation I am willing to put my name behind. It cannot own the relationship between the brand and the people who read the work.
That stays human.
The agent manages memory.
I decide the point of view.
And I still own the final content and publication decision.
Data analysis is the third area I delegate heavily, with a harder execution boundary.
I don't ask AI to calculate, as we know, they are bad at numbers (like sometimes couldn't figure out how many "r" are in the word strawberry). But luckily they are really good at coding, and hard coded programs are pretty reliable for calculation.
So I vibe coded programs that combine advertising, demand, CRM and Deal data, test different views, model trends and return the report to AI agent for preparing recommendations. A perfect loop.
These are good AI jobs because they reward breadth, persistence, memory and repeatable analysis. My attention is more valuable when I define the problem and judge the result.
This is where the employee metaphor reaches its limit.
AI agents can have responsibilities inside my system. They do not have my accountability outside it.
The more consequential the external action becomes, the more likely I am to put a human gate in front of it.
My agents can prepare material for a client conversation, but they do not conduct that relationship autonomously.
They can prepare content, but they can't write and publish them under my name.
They can recommend campaign changes, but they can't move the money.
They can work with sensitive business information inside the permitted environment, but they can't override privacy rules.
These boundaries are built into the system deliberately.
If something eventually reaches a client, an audience, a live account or a budget, a human should always hold accountability.
With this system, I now hand a substantial amount of marketing work to AI agents.
They research. They retrieve. They compare. They organise. They write programs. They run analyses. They review one another. They prepare work for the next stage.
That removes a great deal of labour from my week.
But it has not removed me from the workflow.
It has changed where I spend my attention.
I spend less time collecting material, moving context between systems, repeating analytical work and supervising individual steps. More of my time goes into defining the job, deciding what good looks like, designing the right team, protecting the boundaries and making the calls that carry real consequences.
A surprising amount of what makes this work came from leading human teams.
Capable people need clear ownership, enough context to make good decisions and room to use their judgement. They also need review, escalation paths and a manager who knows which decisions cannot be delegated.
I now use many of the same principles with AI agents.
I build systems like this through Smartify Marketing, so I have a commercial interest in the method. You do not need my exact setup to try the principle.
Start with one recurring workflow. Give the job a clear owner. Define the outcome, context and non-negotiable boundaries. Split delivery and review where the stakes justify it. Decide where human approval becomes mandatory.
Then give the agent enough room to do the job.
That is roughly how I run my AI marketing team today. The agents do far more of the work than they used to, while I spend far less time telling them how to work.