There's a growing opinion that AI agents are overhyped, unreliable and, in many cases, simply not very useful.
I don't fully agree with that.
I use AI every day, I build workflows around it, and I think agents will become a normal part of how companies operate. But I also think we are moving past the stage where "we built an AI agent" is interesting on its own.
The more I use these systems in real marketing work, the more obvious one thing becomes: automating a task and being able to do that task well are not the same thing.
And this is where I think a lot of the current AI-agent hype gets complicated.
Not every task needs an agent
The original promise of agents was very attractive.
Instead of doing something manually, you give an AI a goal, connect it to the necessary tools and let it figure out the rest.
Research a market. Analyze an ad account. Generate content. Send emails. Make recommendations. Update a CRM.
Ideally, the work happens almost by itself.
Technically, we can already automate a surprising amount of this.
But the question I increasingly ask is not whether AI can perform a task. It is whether AI is actually the right way to perform it.
There are tasks where the rules are clearly defined and the expected result should be predictable. If I need to calculate a metric, check whether a customer meets a threshold or trigger an email after a specific event, I usually don't need an LLM to "think" about it. Traditional software is often better.
Then there are tasks where interpretation is part of the work. Research, qualitative analysis, finding patterns in customer feedback, generating hypotheses or exploring possible creative directions are much more probabilistic by nature.
This is where AI becomes genuinely useful.
The strongest systems, in my view, will probably combine both approaches rather than trying to make everything agentic.
I still don't trust AI 100%
There are very few tasks where I currently take AI output and use it without checking.
One interesting exception is meeting summaries.
Today, if I give ChatGPT a good meeting transcript, the resulting summary is often almost exactly what I need. I used to edit these summaries heavily. Now I sometimes barely touch them.
And I think there is a reason for that. The source of truth already exists. The conversation happened. The transcript contains the information. AI is mostly compressing, organizing and structuring something that is already there.
The situation changes completely when I ask it to create something that requires professional judgment.
If I use AI for strategy, Meta Ads, Klaviyo, creative concepts, copy or data interpretation, I still check everything.
Not because every answer is bad. Many are useful.
But useful is very different from reliable enough to implement without thinking.
Marketing is a good example of the problem
I see this constantly in performance marketing.
You can ask an LLM to build a Meta Ads strategy and receive something that sounds completely professional.
It will talk about awareness, consideration, conversion, retargeting, creative testing, audience segmentation and full-funnel strategy. All the vocabulary is there.
But sometimes the logic behind it makes very little sense in practice.
For example, you can ask for a performance strategy focused on generating purchases and receive a recommendation to launch an Awareness campaign as a major part of acquisition.
Could I ever run an Awareness campaign? Of course. But then there should be a reason for doing it.
Maybe we are supporting a launch. Maybe we deliberately want to build reach in a specific market. Maybe we have a defined hypothesis about how that audience will later be used.
In that case, I also want to know how long we are running it, what comes next and how we will evaluate whether the additional step contributed anything to the business.
Without this context, "use awareness to warm up the audience" is not much of a strategy. It is a phrase that sounds like one.
And this is the kind of difference that is extremely easy to miss if you haven't actually worked inside ad accounts.
The internet has a lot of marketing advice and surprisingly little marketing context
I think part of the problem comes from the information AI has available to learn from.
There is an enormous amount of marketing content online. There are SEO articles, platform documentation, agency blogs, LinkedIn posts, case studies, YouTube videos and endless lists of best practices.
What is much harder to find is the complete context behind actual business decisions.
Very few companies publicly show their full acquisition costs, margins, repeat purchase rate, cash constraints, inventory situation, CRM performance, customer mix and everything they had already tested before deciding what to do next.
We might see that a company changed its campaign structure and ROAS increased. We rarely see the complete unit economics behind that decision.
We might read that a brand increased its investment in top-of-funnel activity. We don't necessarily know how mature the brand was, how much organic demand already existed, what its repeat purchase rate looked like or whether the company was optimizing for short-term contribution margin or long-term market share.
And those details completely change whether the same recommendation makes sense for another business.
This is why I think a lot of so-called marketing "best practices" are simply context-dependent decisions that became detached from their original context.
AI is extremely good at synthesizing all of this information. But if the underlying information is incomplete, generic or outdated, synthesis doesn't magically turn it into practical experience.
Fast-changing industries make this even harder
Marketing platforms themselves are also changing constantly.
Meta uses more automation and AI in its advertising products every year. Targeting, delivery, creative optimization and campaign structures continue to evolve.
The way I might structure an account today is not necessarily the way I would have structured the same account several years ago.
That creates another problem for AI systems. A recommendation can be technically valid according to older marketing knowledge and still be a poor recommendation today.
This is especially visible when I ask AI to generate hypotheses for Meta Ads.
It can easily produce twelve creative angles. Usually, a few will be useful.
The rest may be generic, impossible to execute with the assets available, disconnected from the actual positioning of the product, incompatible with the economics of the business or based on ideas that sound better in theory than they perform in a real account.
An experienced marketer can filter them very quickly.
And that leads to what I think is one of the most important questions around AI.
AI works very differently for an expert and for a beginner
For a strong specialist, AI can be an incredible tool.
If I receive twelve hypotheses and ten of them are weak, that is not necessarily a problem. I can discard the ten quickly and work with the two that are interesting. The AI has still saved me time.
It can help me explore more directions, challenge my first idea, structure information faster and get through repetitive intellectual work much more efficiently.
But a beginner sees the same twelve hypotheses. The difference is that they may not know which ten should be rejected.
That creates a strange situation where AI can produce the appearance of expertise faster than it produces expertise itself.
The output looks professional. The terminology is correct. The argument is coherent.
But coherence is not the same as correctness, and correctness is not always the same as something being applicable in a real business.
I think this is why AI is often particularly powerful in the hands of someone who already knows the domain. It amplifies their existing expertise.
For a beginner, I would use it differently.
Instead of asking, "Build me a Meta Ads strategy," I would use AI as part of learning and practice.
Explain why you chose this objective. Give me three possible approaches and the trade-offs between them. What assumptions are you making? What information would you need before deciding? Here is a real account, what would you look at first and why? Now critique my decision.
That is a completely different way of working with AI.
But even then, I don't think you can remove practice from the equation. At some point, you need to see what happens when real money is being spent, real customers behave differently from your assumptions and a strategy that looked great on paper simply doesn't work.
Domain expertise may become more important, not less
This is also why I am skeptical when I see agents being built for specialized industries by people who have very little experience in those industries.
Building the technical workflow is becoming easier.
You can connect an LLM to Meta, Shopify, Klaviyo, analytics tools and a database. You can give it access to company documentation and create a nice interface.
The result can look extremely sophisticated.
But who decided what the agent should look for? Who decided which metrics matter? Who defined when a drop in ROAS is actually a problem?
Who knows whether the company should optimize for acquisition right now or intentionally accept a higher CAC because retention is unusually strong? Who understands that two businesses with exactly the same Meta performance might need completely different decisions because their margins, inventory and cash positions are different?
This is the layer that is much harder to automate.
The hardest part of building a useful agent may not be getting the model to reason. It may be encoding enough real-world understanding into the system to know when that reasoning is good.
Business context matters as much as data
There is another limitation here that I think gets overlooked.
People often say AI needs better data. That is true, but sometimes the problem isn't bad data. It is missing business context.
You can give an agent perfect Meta Ads data and perfect Shopify data and still not give it enough information to make a good strategic decision.
Maybe the founder is deliberately trying to clear inventory. Maybe the company cannot increase spend because of cash flow. Maybe one category has a much higher repeat purchase rate than another. Maybe the next shipment will not arrive for six weeks. Maybe the brand is entering a new market and is temporarily willing to sacrifice efficiency. Maybe a campaign that looks weak in Meta is driving customers who later convert through another channel.
Those are not necessarily problems an agent can discover by simply querying another API.
Sometimes they exist in meetings, experience, internal constraints and things that people inside the company simply know.
A good strategist doesn't only analyze dashboards. They understand the business around the dashboards.
So what should we actually automate?
I don't think the answer is to avoid AI agents.
I think the answer is to be much more precise about what we expect them to do.
There are many parts of my work I am very happy to automate or accelerate with AI: research, meeting summaries, structuring information, analyzing large amounts of qualitative data, finding patterns in customer reviews, producing initial creative directions, generating hypotheses, creating first drafts, and exploring alternative explanations for what we see in the data.
I am much more cautious when AI moves from helping me think to making consequential decisions for me: changing budgets, deciding strategy, interpreting business performance without full context, communicating something sensitive to a client or customer, or making a financial decision.
I don't think humans need to remain in every workflow forever. But today, in specialized work, I still want someone with domain expertise deciding where AI can operate independently and where it needs supervision.
Maybe the future isn't "AI replacing experts"
I think we may have framed the AI conversation incorrectly from the beginning.
We often talk about AI as something that reduces the amount of expertise a person needs. And for simple tasks, it absolutely can.
But the pattern I see in more complex work is almost the opposite. The better AI becomes at producing plausible answers, the more valuable it becomes to know whether those answers are actually good.
A strong specialist can use AI to move much faster because they know what to trust, what to question and what to throw away.
A beginner can also learn much faster with AI, but only if they treat it as a tool for learning rather than an unquestionable source of answers.
And this leaves me with a question I don't think we have fully answered yet.
AI is often sold as a technology that will allow people to understand less about the work they are doing. But what if, as the tasks become more complex, we actually need more expertise to know whether the AI's answer is any good?