Most "AI Agents" Are Just Chatbots in Costume
Gartner expects 40% of enterprise applications to include AI agents by 2026, up from less than 5% today. That projection has set off a marketing land grab. In the Greek vendor landscape we audit, almost every chatbot has been rebadged as an "agent" in the past six months. Almost none of them are.
The distinction is not pedantic. Real AI agents do work that chatbots cannot do, and they change the economics of an enterprise function in ways chatbots cannot. Buying the wrong thing under the right name leaves you with the productivity ceiling of a 2023 chatbot and the price tag of a 2026 agent.
This piece is about the four-criterion test we apply to every vendor that pitches us "agent" capabilities, what the distinction means for your AI budget, and the questions to ask before signing.
What actually counts as an AI agent
We use four criteria. A system needs all four to qualify. Three out of four is a chatbot with extras.
Autonomy. The system completes a task without a human deciding each step. Where a chatbot answers a question and waits for the next prompt, an agent receives a goal and pursues it, including making intermediate decisions about what to do next.
Tool use. The system can invoke external tools (APIs, databases, calculators, other agents) to accomplish its goal. A chatbot generates text; an agent generates text and triggers actions in the systems where work actually happens.
Memory. The system retains state across interactions. Not just conversation history within one session, but persistent memory of users, prior outcomes, and learned preferences that informs future decisions. Where a chatbot starts each conversation from zero, an agent gets sharper over time.
Multi-step planning. The system can break a goal into subtasks, sequence them, and adjust the plan when a step fails. Where a chatbot answers what you asked, an agent figures out what needs to happen for you to get what you wanted.
Most vendors selling "AI agents" in the Greek market today fail on at least two of these. The most common failure is autonomy: the system requires a human to approve every step, which means it is automating typing, not work.
Why the economic difference is enormous
The conversation about chatbots was always a labour-reduction conversation. Deploy a chatbot, deflect tier-1 support tickets, save the headcount cost of the agents who would have answered them. The numbers were real but bounded. Most enterprise chatbot deployments saved 15 to 30% of tier-1 support cost.
The conversation about agents is a workflow-replacement conversation. Deploy an agent that owns a process end to end, and the cost structure of that process changes. Not 15 to 30% saved. Sometimes 60 to 80%, because the agent does work that previously required coordination across multiple humans.
A concrete example. A Greek enterprise we work with had a procurement workflow that involved a procurement officer collecting quotes from five vendors, a finance reviewer checking budget alignment, an operations lead approving the technical specs, and a procurement officer issuing the PO. Cycle time was four to seven days per request.
The agent version of this workflow gathers quotes via vendor APIs, runs the budget check against the live ERP, surfaces technical fit using the requirements document, presents a recommendation with reasoning, and waits for one human approval before issuing the PO. Cycle time is under two hours for routine cases, under a day for complex ones. The procurement function did not get cheaper by 30%. It got fundamentally smaller, with the existing team handling roughly four times the request volume.
That is the economics gap. The two products live on different price points, with different ROI math behind them and different organisational implications when deployed. A chatbot helps a human do their job faster. An agent does the job.
What the Greek market is actually selling
We audit a lot of vendor pitches. The pattern across the past six months is consistent. Vendors fall into three buckets.
Chatbots relabelled as agents. The actual product is a RAG-powered Q&A interface over the company's documents. Useful, but not autonomous, no tool use beyond document retrieval, no memory across sessions, no multi-step planning. Pricing has gone up 30 to 60% since the relabel.
Workflow tools with an LLM bolted on. A traditional workflow automation platform (think Zapier-style) that now includes an LLM step for text generation. The LLM does not plan, it just fills in templates. The "agent" framing is generous. The product is fine for what it is, but the ROI looks like classic workflow automation, not agentic AI.
Actual agents from non-Greek vendors. A small number of international platforms (mostly US-based) ship genuine agentic capability. The Greek market reseller channels are starting to pick these up, but the localisation work for Greek language, Greek regulatory context, and Greek enterprise data structures is mostly absent.
The result is that a Greek enterprise looking to buy "an AI agent" today will see twenty pitches, of which maybe two qualify as actual agents, and of those two, neither is fully ready for the local context. This is not an indictment of the vendors. It is a description of where the market is in early 2026.
Why 2026 is the multi-agent shift year
Both Gartner and Forrester have called 2026 the breakthrough year for multi-agent systems, where the interesting capability is not a single agent doing a single job, but multiple agents coordinating to solve a larger problem.
The procurement example earlier was a single-agent case. The multi-agent version is more powerful and more complex. A negotiation agent talks directly to vendor agents to settle pricing, while a compliance agent monitors the negotiation against your procurement policies and flags anything that breaches your supplier-risk thresholds. The procurement officer becomes a supervisor, not an executor.
Most Greek enterprises are not ready for multi-agent systems. The honest reason is that they have not yet completed the data integration and process mapping that single agents require. The good news is that the design pattern for multi-agent systems is the same pattern, applied recursively. Every learning from a single-agent deployment carries forward. None of the work is wasted.
The risk for Greek enterprises is buying multi-agent marketing without the single-agent foundation. We have started seeing pitches for "multi-agent platforms" that are essentially three chatbots in a UI. The four-criteria test applies recursively here too. Each agent in the system has to qualify on its own.
Five questions to ask any vendor pitching you "agents"
These are the questions we use in vendor evaluations for our clients. They sort the real agents from the rebadged chatbots in about ten minutes.
Question 1: Show me the system completing a task without me clicking anything. If the demo is "ask it a question, see the answer," it is a chatbot. If the demo is "give it a goal, watch it work," it is closer to an agent.
Question 2: What tools does it have access to, and what happens if I add a new one? Real agents have an extensible tool layer. Adding a new tool should be a configuration change, not a code release.
Question 3: How does it remember things between sessions? The answer should involve specific memory architectures (vector stores or structured user profiles) and not "we keep the last 10 messages in context."
Question 4: Show me a case where its initial plan failed and it recovered. Real planning agents have demonstrable replanning behaviour. Demos with happy-path-only execution are red flags.
Question 5: What does the autonomy boundary look like, and who configures it? A serious agent product gives you fine-grained control over what the agent can do alone versus what requires approval, with hard-coded boundaries beyond that. If the answer is binary (agent on / off), the product is not designed for enterprise deployment.
If a vendor cannot answer four out of five of these crisply, you are either looking at a chatbot or at an agent that is not ready for production. Both might still be useful purchases. Just not at agent prices.
Where this fits with the rest of the AI strategy
The agent-versus-chatbot question is downstream of two more fundamental ones.
The first is governance. We covered this in our piece on the Absorption Gap. Adopting agents without the governance function in place compounds the cultural debt that already exists. An agent that can act without a human checking it must have an owner who can be held accountable when it acts wrongly. Most enterprises do not even know what AI tools their employees are already using, a problem we map in detail in our Shadow AI piece.
The second is investment economics. The agent-versus-chatbot distinction maps directly to the build vs. buy vs. hire decision every Greek enterprise is now making. A chatbot buying decision is a tactical purchase. An agent buying decision is a strategic capital allocation, with workflow-redesign costs that the next CFO scrutiny round will surface fast.
If you are about to walk into a vendor pitch and you are not sure which category you are looking at, the four criteria are the test. The five questions are the diagnostic. Use them in that order.
We help Greek enterprises evaluate AI agent vendors and design deployments that survive both the four-criteria test and the CFO scrutiny that follows. If the latest vendor pitch left you uncertain whether you were looking at an agent or a chatbot in a costume, talk to us. Get in touch at inbusiness.gr.