AI Agents are Becoming More Popular
There is a terminology problem developing around artificial intelligence, and it is starting to become a business problem.
Nearly every software company seems to be talking about agents. Products that were described as copilots six months ago are now agentic. Automation platforms have agents. CRMs have agents. Customer service tools have agents. Marketing platforms have agents. In some cases, the label is justified. In others, the product has changed less than the language around it. Click the link to see our Marketing Super Agent.
That would be mostly harmless if “agent” were just another marketing term. It isn’t. The distinction matters because there is a meaningful difference between software that gives you information and software that can take action inside your business.
I was reminded of that while reading Chapter 1 of Sam Schreim’s Beyond the Prompt. The chapter spends a lot of time separating actual agentic systems from the growing pile of products being marketed under the same label. The point that stuck with me is simple: once AI can move from recommending an action to actually performing one, the conversation changes.
At that point, we are no longer talking only about productivity. We are talking about permissions, accountability, integration, oversight, and risk.
A chatbot gives you an answer. An agent moves a process forward.
Most of us learned generative AI through a prompt-and-response model. We ask a question, request an analysis, or give the system a piece of work to produce. It responds. We decide what to do with the answer.
That model is still extremely useful. I use it constantly. But it keeps a human in an important position: between the AI’s output and the real-world consequence.
If an AI assistant drafts a poor customer email, someone can edit it before it goes out. If it makes a weak recommendation, someone can ignore it. The system may influence the decision, but it does not necessarily execute the decision.
An agent changes that relationship. Instead of being asked only for an output, it can be given a goal and allowed to work through some of the steps required to reach it.
Think about the difference between asking AI to “write a follow-up email for yesterday’s sales demo” and asking a system to “make sure every qualified prospect from yesterday’s demo receives the appropriate follow-up.”
The first request produces text. The second potentially creates a workflow.
To accomplish the second goal, a system might need to identify attendees, check each account in the CRM, review previous conversations, determine whether the opportunity is qualified, select the appropriate next step, prepare a message, create a task, update a record, and perhaps send something once the right conditions are met.
The email is only one part of the work. The sequence is where the value starts to become interesting.
That is also where the risk begins to change.
The three capabilities I design for for in a real AI agent
Reasoning means the system can determine what should happen next.
Traditional automation is very good at following predefined instructions. When a specific condition occurs, the software performs a specific action. That can be incredibly reliable, and for many business processes it is exactly what you want.
Agentic systems become more interesting when the path cannot be completely predetermined.
Imagine a subscription business trying to identify customers at risk of leaving. A conventional AI assistant can explain common churn signals and write a retention email. An agentic workflow could go further by reviewing account activity, checking recent support interactions, looking at billing behavior, and preparing a recommended intervention based on what it finds.
The important question is not whether the final recommendation sounds intelligent. The more important question is whether the system can work through the steps required to arrive at that recommendation without someone manually directing every move.
This is one reason polished demos deserve a little skepticism. A demo usually shows the successful path. Real businesses run on exceptions.
What happens when the customer record is incomplete? What happens when two data sources disagree? What happens when the expected field is missing or a previous exception changes the normal rule?
That is where you begin to see the difference between something that can genuinely navigate a workflow and something that simply produces an impressive response.
Memory means the system has access to the context the job requires.
Reasoning without context is not very useful.
A person handling an account might know that the customer complained last week, received a special pricing exception six months ago, recently changed products, and was promised something unusual by a salesperson. All of that context can affect the correct next step.
An AI system doing meaningful work needs some way to retrieve the relevant information when it matters. Connecting a model to a folder full of documents is not the same thing as giving it useful organizational memory.
The practical questions are more specific. What can the system access? Where does that information come from? How current is it? Can we tell which information influenced a recommendation? What happens when sources conflict? Who can change the information the system relies on?
Those questions may sound technical, but they quickly become management questions. If an agent is making decisions based on company knowledge, then the quality and provenance of that knowledge matter.
A bad piece of context used once can create a mistake. A bad piece of context used automatically across thousands of decisions can create a pattern.
Action is where the stakes change.
This is the capability I would pay the most attention to.
An AI tool can reason about a refund. An agent may be able to issue one. It can recommend changing a CRM record, or it can change the record. It can draft a message, or it can send the message. It can identify an overdue invoice, or it can initiate the next collection step.
Once the system has access to tools, APIs, connectors, and business applications, it can affect the world outside the conversation window.
That does not automatically make the system dangerous. It does mean we should evaluate it differently.
The question is no longer simply, “How accurate is the AI?” It becomes, “What is the AI allowed to do when it believes it is right?”
That is a much better question.
Permissions are probably more important than most AI demos make them look.
We already understand this principle with employees.
A new hire does not normally receive unlimited access to every customer record, pricing system, bank account, ad platform, and corporate social account on the first day. Access grows according to the job, the risk involved, and the level of trust the organization is prepared to place in that person.
AI should not get a free pass on the same logic.
For any action an agent can take, I would want to understand the potential blast radius. Can the action be reversed? Will someone know it happened? Could it affect a customer? Could it create a financial commitment? Does it change one record or potentially thousands? At what point does a human need to approve what happens next?
Those are more productive questions than asking whether we “trust AI.”
Trust is too broad. A business can be perfectly comfortable allowing an agent to categorize inbound leads while being completely unwilling to let the same agent independently alter pricing. Both positions can be rational.
The useful conversation is about the exact action, under exact conditions, with exact limits.
A chatbot, a copilot, automation, and an agent are not the same thing.
Part of the confusion in the market comes from grouping several useful technologies together as though they were interchangeable.
A chatbot primarily communicates. You ask for something and it responds. A copilot usually assists a person while that person remains in control of the work. Traditional automation follows a predefined sequence of rules. An agent, at least in the more meaningful sense of the word, has some responsibility for moving a goal forward and deciding how to proceed within the constraints it has been given.
None of these categories is automatically better than the others.
In fact, one of the mistakes businesses may make over the next few years is using agents where ordinary automation would be cheaper and more reliable. If a process is completely predictable, adding probabilistic reasoning may create complexity without creating much value.
The reverse is also true. A company may think it has adopted agentic AI when it has really added text generation to an existing workflow. That can still improve productivity, but the economic opportunity is different.
Knowing what you bought matters because expectations determine how the technology gets deployed.
“Agent washing” is going to make buying decisions harder.
The phrase “agent washing” has started showing up for a reason. As budgets move toward agentic AI, vendors have an obvious incentive to position their products inside the category.
That does not mean every vendor using the word “agent” is being deceptive. Technology categories genuinely do blur together, and products evolve quickly. But competitive pressure tends to stretch definitions.
A workflow with a language model in the middle can start being described as an autonomous agent. A chatbot with a few integrations may suddenly become a digital employee. A product that still requires a person to execute every meaningful step can be presented as though it owns the workflow.
This is where buyers can get into trouble.
If leadership believes it purchased a “super chatbot,” it may use a capable system for minor tasks and never capture much operational value. If leadership believes it purchased a fully autonomous employee, the opposite problem occurs: expectations jump far ahead of what the process, data, integrations, and controls can support.
Neither mistake is really about artificial intelligence. They are expectation problems caused by unclear definitions.
The integration question may expose more than the AI question.
One of the easiest things to underestimate is how much real work occurs between systems.
A customer record lives in the CRM. Billing information lives somewhere else. Support history is in another platform. A team maintains an important spreadsheet that was never meant to become infrastructure. A legacy system technically has an API, but nobody has touched it in years. One critical exception exists only because an experienced employee knows what to do when a certain customer calls.
Then a vendor shows an agent completing the process end to end in a controlled demo environment.
The AI may be capable of the reasoning. That does not mean your organization is capable of giving it the access and context required to perform the same workflow in production.
This is why I think workflow mapping should come before serious agent procurement.
Before buying anything, document how the work actually gets done. Identify the systems involved, the handoffs, the decisions, the exceptions, the permissions, and the places where people compensate for broken processes without realizing they are doing it.
You may discover that the AI model is not the bottleneck at all.
Four questions I would take into any AI agent demo
Schreim closes the chapter with a practical reality test. I would expand it slightly and use some version of these four questions whenever a vendor describes its product as an agent.
First, what goal does the system actually own? Is it responsible for moving a business outcome forward, or does it wait for a human prompt every time something needs to happen?
Second, what context can it retain and retrieve? Does it have access to the information required to make the decision, and can we understand what information influenced its behavior?
Third, what actions can it take inside our systems? Does it only make recommendations, or can it create, change, send, approve, trigger, or otherwise execute work?
Fourth, can we see what it did and why? This may be the most important question of the four. If something goes wrong, can we reconstruct the decision and understand the action that followed?
A vendor should be able to answer those questions without hiding behind a product name.
If the answers are vague, the safest assumption is that the capability is vague too.
Marketing is a good example of why this distinction matters.
Marketing teams are already heavy users of generative AI, but much of that use is still centered on content production. We use AI to draft emails, summarize research, create ads, brainstorm campaigns, rewrite landing pages, and repurpose content.
Those are useful applications. They are not necessarily agentic.
A more agentic marketing workflow might begin with an objective such as improving follow-up on high-intent leads. The system could monitor new leads, retrieve CRM history, identify the right segment, look at the content or campaign that generated the lead, prepare an appropriate response, route unusual cases for review, update the CRM, and schedule the next action according to predefined rules.
Now the work spans multiple systems and decisions. That is where the potential efficiency becomes more significant, but it is also where controls matter.
The biggest opportunity is probably not replacing an entire marketing department with autonomous software. That makes for a dramatic headline, but it is a poor starting point for most companies.
A better starting point is to find the friction inside an existing workflow.
Maybe qualified leads wait two days for follow-up. Maybe campaign reporting requires someone to pull data from four platforms every Monday morning. Maybe customer reviews are being collected but never routed to the right person. Maybe the sales team repeatedly searches for the same account information before outreach. Maybe lifecycle campaigns depend on manual lists that are always slightly out of date.
Those are real problems. They can be mapped, measured, and improved.
They also make much better candidates for AI than the vague instruction to “automate marketing.”
Start with the workflow, not the tool.
If I were evaluating where agents belong in a business today, I would start with one recurring workflow and put it on paper.
What triggers the work? What outcome are we trying to create? Which information does a person need to make the right decision? Which systems are involved? Where do exceptions occur? Which steps require actual judgment? Which actions are safe to automate? Which actions should require approval? What metric tells us whether the workflow improved?
That exercise is deliberately boring.
It is also more valuable than watching another carefully rehearsed AI demo.
Once you understand the workflow, the technology becomes easier to evaluate because you are no longer shopping for “AI.” You are shopping for a specific improvement to a specific piece of work.
That is the point where the agent conversation becomes practical.
For marketing teams in particular, this is the work I believe is worth doing now: identifying which workflows genuinely benefit from reasoning and action, which ones should remain simple automation, and where a human should stay in the loop. If you are trying to map that out inside your own growth operation, that is also the kind of problem we work through at Get Marketing VIP.
The advantage will come from operating better, not collecting more AI tools.
Most companies will eventually have access to roughly the same class of AI models. The models will continue to improve, and features that feel sophisticated today will become standard software capabilities.
That suggests the durable advantage is unlikely to come from simply having access to AI.
The advantage will come from knowing where to use it, giving it the right context, connecting it to the right systems, limiting its permissions intelligently, and measuring whether the underlying workflow actually got better.
That is why the definition of an agent matters.
If we treat every chatbot, automation, copilot, and workflow as the same thing, we lose the ability to make sensible decisions about cost, risk, and value. If we understand the mechanics, the category becomes much less mysterious.
Before asking whether a business needs more AI agents, I would ask a simpler question:
What work are we actually trying to improve?
Once that answer is clear, deciding what the technology should do becomes much easier.