AI agents in marketing: what works?
Agents work when the job has clear inputs, observable outputs and a way to tell whether the work is good. A working list of what works in the wild, and what gives your customers mediocre marketing.
AI agents may be the most overhyped and underexamined idea in marketing right now.
The promise is compelling. Build an AI marketing team. Give agents responsibility for research, content, social, SEO, outbound and analytics. Let them coordinate with one another. Wake up to a marketing machine that has been working while you sleep.
Some of this works.
Some of it creates an extraordinary amount of mediocre marketing very efficiently.
After spending time thinking about where agents actually create leverage, I've come to a fairly simple conclusion:
Agents work best when the job has clear inputs, observable outputs and a way to determine whether the work is good.
The further you move from those conditions, the more human involvement matters.
What works: monitoring for change
This is one of the strongest use cases I've found.
Give an agent a defined universe to watch and a clear definition of what constitutes an important change.
The agent's job is to identify the change, provide the evidence and explain its possible significance.
This works because the task has boundaries.
The agent knows where to look. It knows what matters. A human can quickly determine whether the result is useful.
Competitive-intelligence platforms like Klue watch the rivals for you, and Common Room watches the people. A research layer on top — Perplexity for sourcing, Claude for the read — turns a feed of events into something worth acting on.
The useful version of this is narrower than “research my competitors”:
Tell me when something changes that should cause us to reconsider what we're doing.
What works: finding patterns across information humans don't have time to read
This is probably where I see the most untapped value.
Marketing teams produce enormous amounts of qualitative data:
Sales calls. Customer interviews. Lost-deal notes. Support conversations. Surveys. Reviews. Emails. Campaign responses.
Most organizations analyze a fraction of it.
Agents can repeatedly examine that information for specific patterns.
What objections are increasing. Which problems customers are describing differently. What separates fast-moving opportunities from stalled ones. Which messages are showing up organically in customer language. What customers are asking for that our marketing barely discusses.
An agent can run those questions every month and compare the answers with the previous period.
That gives you movement over time.
What works: preparing humans to perform better
Some of my favorite agents never interact with a customer.
They prepare the person who does.
Before a sales meeting, an agent can assemble the relevant account history, previous conversations, company developments, likely priorities and open questions.
Before a webinar, it can research the audience and identify questions worth addressing.
Before a campaign review, it can synthesize performance, customer responses and sales feedback.
Before a leadership meeting, it can surface the decisions that actually need to be made.
This is an important category of AI leverage because the human still owns the interaction and judgment.
The agent improves the quality of the starting point.
What works: narrow, high-volume classification
AI is very good at jobs that marketers rarely enjoy doing manually.
- Classify these 2,000 companies by use case.
- Read these 500 survey responses and identify themes.
- Categorize these lost opportunities.
- Determine which inbound inquiries fit our ICP.
- Tag these customer quotes by problem.
- Match these accounts against these five buying signals.
- Evaluate these 300 pages against our messaging framework.
These aren't glamorous applications.
They can be incredibly valuable.
They turn information that was previously too expensive to structure into something the business can actually use.
What works: controlled experimentation
Agents can also make experimentation cheaper.
Imagine giving an agent three positioning hypotheses and asking it to create tightly controlled variations for different audiences.
Run small tests. Collect the response. Have the agent analyze what happened. Use the evidence to determine the next experiment.
This creates a much faster learning loop.
The important word is controlled.
The strategy, audience, hypothesis and success criteria should already be clear.
AI increases the number of experiments you can afford to run.
What doesn't work particularly well: “Run our social media”
This sounds fantastic in an agent demo.
In practice, it tends to produce content nobody needed.
The agent can find a topic, write something, create an image, schedule it and report that the work has been completed.
Every step can function perfectly.
The marketing can still be bad.
Publishing was never the hard part. Having something worth saying is.
Agents are useful for monitoring conversations, researching topics, extracting ideas from existing material, repurposing strong thinking and preparing drafts.
Giving an agent a quota of five posts per week usually optimizes for fulfilling the quota.
What doesn't work: autonomous thought leadership
Thought leadership requires a point of view.
That point of view usually comes from experience, pattern recognition, conviction, disagreement and occasionally being willing to say something the market doesn't already believe.
AI can help interrogate an idea.
It can research it. Challenge it. Find evidence. Find counterarguments. Help structure it. Turn a conversation into a draft.
Those are powerful capabilities.
An agent tasked with “developing thought leadership” has a different problem. It can synthesize what has already been said extremely well — which is the opposite of naming something first.
That isn't necessarily leadership.
What doesn't work: fully autonomous outbound
Technically, this is becoming increasingly possible.
That doesn't make it a good growth strategy.
An agent can find prospects, research them, generate personalized messages, send them, follow up and potentially respond.
At scale, a small error in judgment becomes a large reputational problem — and autonomy should track reversibility.
There is also an uncomfortable industry-wide consequence: when everyone can generate personalized outreach almost for free, personalization itself becomes less valuable.
The bar moves toward relevance.
AI is excellent at finding the right accounts, identifying timing signals, researching context and preparing outreach.
I would put significantly more automation into deciding who deserves attention and why than into maximizing the number of messages sent.
What doesn't work: agents evaluating their own mediocre work
This is a subtle failure mode.
AI evaluation is useful, but you need external signals.
Customer response. Conversion. Pipeline movement. Search behavior. Sales feedback. Revenue. Human judgment.
Otherwise, you can create a closed loop where AI continually validates AI.
What doesn't work: building the AI org chart first
I see versions of this everywhere.
CMO Agent. Content Agent. SEO Agent. Social Agent. Research Agent. Analytics Agent. Demand Gen Agent.
Suddenly you've built an impressive diagram containing seven agents without identifying a single business problem.
I would build from the work backward.
The architecture should emerge from useful workflows.
The test I use
Before turning a marketing activity into an agent workflow, I ask five questions:
- Can I define the job precisely?
- Can I give it the information required to do the job?
- Can I tell whether the output is good?
- Can mistakes be detected before they become expensive?
- Does running this repeatedly create meaningful leverage?
The more confidently I can answer yes, the better the agent use case.
That's why monitoring works. Classification works. Research works. Synthesis works. Preparation works. Pattern detection works. Controlled experimentation works.
And it's why I remain much more cautious about handing agents responsibility for brand, relationships, original thinking and consequential external communication.
Where I think this is going
The future of marketing probably includes a lot of agents.
I just don't think most customers will ever know they exist.
The best ones will be running underneath the organization.
Watching. Researching. Connecting information. Preparing people. Identifying patterns. Running analysis. Maintaining context. Surfacing opportunities.
Marketing teams will walk into decisions knowing more.
Salespeople will walk into conversations better prepared.
Leaders will see market changes sooner.
Customer insight will travel further through the organization.
And small teams will be able to maintain capabilities that previously required much larger ones.
That's where I see the real opportunity.
Measure a marketing agent by the decisions around it.
Does the organization make better marketing decisions because this exists?