Most "AI agents" are chatbots with a database

The Turing Test We Should Be Running

Alan Turing asked: can a machine convince a human it’s human? That question shaped 70 years of AI research and produced exactly zero useful operational software.

Here’s a better test: can the machine do the work without being asked?

We don’t care if your AI passes the Turing test. We care if it follows up with the prospect you forgot about. We care if it catches the broken deployment at 3 AM. We care if it noticed that your AWS bill spiked 40% this month and flagged it before the CFO did.

That’s the difference between a chatbot and an agent. And right now, 90% of the products claiming to be “AI agents” fail this test completely.

The Chatbot Costume

Let’s describe a product we’ve seen at least thirty times in the last year.

You sign up. There’s a slick onboarding. You connect your data sources, CRM, email, maybe a database. Then you get a chat interface. You type: “Show me deals that haven’t been contacted in 30 days.” It returns a list. Maybe it even drafts emails for you. You feel productive.

Then you close the tab and go to lunch. When you come back, nothing has happened. The stale deals are still stale. The emails are still drafted but unsent. The AI is sitting there, cursor blinking, waiting for your next prompt.

That’s a chatbot. A fancy one, with RAG, function calling, maybe even some tool use. But a chatbot nonetheless. It did nothing you didn’t explicitly ask for.

Now let’s describe an agent.

You configure it once: “Monitor my pipeline. When a deal goes 14 days without contact, draft a follow-up based on the last conversation. If it’s a high-value deal, escalate to me for review. Otherwise, send it.”

You go to lunch. You go home. You sleep. When you wake up, three follow-ups were sent to medium-priority deals. One high-value deal is flagged in your inbox with a drafted message and the agent’s reasoning for why it chose this particular angle. Your pipeline didn’t stagnate overnight because the agent was working while you weren’t.

That’s the difference. And it’s not a spectrum. It’s a binary.

The Three Requirements for Agency

After building twelve agents at Apollo Space, we’ve distilled what separates agents from chatbots into three non-negotiable capabilities:

1. Continuous Observation

An agent must have eyes. Not metaphorical eyes, actual, persistent connections to data sources that it monitors without being asked.

Apollo Space’s Observability agent doesn’t wait for you to ask “is the site up?” It’s connected to your infrastructure. It watches response times, error rates, deployment events, and log patterns continuously. When it detects something anomalous, it’s already halfway through diagnosing the issue before you even know there’s a problem.

Most “AI-powered monitoring tools” are still chatbots. They let you query your logs in natural language. That’s useful, but it’s not observation. Observation means the system is always watching, even, especially, when you’re not.

2. Autonomous Decision-Making

Observation without decision is just logging. An agent must evaluate what it observes and decide whether and how to act.

This is where most products fall apart. Decision-making requires a policy, a set of rules, priorities, and thresholds that govern when action is appropriate. Building this is hard because it requires encoding operational judgment.

Apollo Space’s SDR agent has a policy engine. When it identifies a stale deal, it doesn’t just flag it. It evaluates: How stale? What’s the deal value? What was the last touchpoint? What’s the prospect’s engagement history? Based on these factors, it decides: follow up now, wait, escalate, or deprioritize. Different decisions for different contexts, without human intervention.

This is fundamentally different from a chatbot that says “here are your stale deals, what would you like to do?” The chatbot put the decision back on you. The agent made the decision and is asking for confirmation at most.

3. Direct Action Execution

An agent must be able to act on its decisions. Not suggest actions. Not draft actions. Actually execute them.

Apollo Space’s QA agent doesn’t send you a message saying “you should run the test suite on staging.” It runs the test suite. It compares screenshots. It creates tickets for failures. It assigns those tickets to the right developer based on who last touched the relevant code. By the time a human sees the ticket, the diagnosis is already done.

This is the scariest capability for most organizations, and it’s why genuine agents are rare. Giving AI the ability to act, to send emails, create tickets, deploy code, modify infrastructure, requires trust, guardrails, and careful scope management. Most companies aren’t willing to do this work, so they build chatbots instead and call them agents.

Why the Market Is Confused

The confusion isn’t accidental. There are structural reasons why the market is drowning in chatbots dressed as agents.

Reason 1: Chatbots are 10x easier to build. A chatbot needs: an LLM, a prompt, and maybe RAG over some documents. You can build one in a weekend. An agent needs: persistent observation infrastructure, a decision engine, action APIs, safety guardrails, rollback mechanisms, logging, and a way for humans to override decisions. That’s months of engineering.

Reason 2: Chatbots are easier to demo. In a live demo, a chatbot looks magical. You type a question, you get a smart answer. The audience is impressed. An agent, by definition, does its best work when nobody is watching. How do you demo something that runs at 3 AM? You show dashboards of what it did, which looks less impressive than a live chat interaction.

Reason 3: Chatbots are safer. A chatbot that gives a wrong answer is embarrassing. An agent that takes a wrong action is dangerous. It could send a bad email to a customer. It could create a false-positive incident. It could make a decision that costs real money. Companies default to the safer option, which is asking humans to make the final call on everything, which defeats the entire purpose.

Reason 4: The term “agent” is being abused for marketing. In early 2025, “AI agent” became the hot term. Every startup pivoted their messaging. Products that were chatbots on Monday became “agent platforms” by Friday. No architecture changed. Just the landing page.

Sequoia Capital published an analysis in late 2025 noting that of 200+ startups claiming to build “AI agents,” fewer than 15% had products that could take actions autonomously. The rest were chatbots with better UX.

The Observe-Decide-Act Loop

At Apollo Space, we think about agents in terms of a continuous loop: Observe → Decide → Act → Learn.

This is not a request-response cycle. It’s an always-on loop. The agent is always observing. When it detects something worth acting on, it decides. When it decides, it acts. When it acts, it observes the result and learns whether the action was effective.

Here’s how this looks concretely with Apollo Space’s Competitor Watch agent:

Observe: The agent monitors five competitor websites, their blogs, their changelog pages, their pricing pages, and their social media accounts. It runs checks every six hours.

Decide: On Tuesday at 2 AM, it detects that Competitor X updated their pricing page. The agent evaluates: Is this a significant change? (Yes, they dropped their enterprise tier price by 20%.) Is this relevant to current deals? (Yes, three prospects in the pipeline are evaluating Competitor X.) Should this be escalated? (Yes, the magnitude of the change and its relevance to active deals crosses the escalation threshold.)

Act: The agent sends a briefing to the sales team with the specific change, its implications, and suggested talking points. It also updates the deal intelligence on the three relevant prospects to include this competitive development.

Learn: It tracks whether the sales team used the talking points and whether those deals progressed. Over time, it calibrates what types of competitive changes are worth escalating and which are noise.

No human prompted any of this. The agent did its job because its job is to watch, think, and act. Continuously.

The “AI Feature” Trap

We see this pattern constantly in SaaS companies adding AI to their products:

1. Take an existing product (CRM, project management tool, analytics platform)
2. Add a chat interface powered by an LLM
3. Give the LLM access to the product’s data
4. Call it “AI-powered” or “with AI agents”

The result is always the same: a chatbot that can answer questions about your data. It’s a better search bar. It’s not an agent.

The trap is that this feels like progress. Users get excited. Demos go well. But adoption data tells a different story. Intercom reported in 2025 that AI chat features in B2B SaaS products have an average weekly active usage rate of 12%. That means 88% of users who have access to the AI feature don’t use it in any given week.

Why? Because a chat interface requires the user to know what to ask. And if the user knew what to ask, they probably already knew where to find the answer. The bottleneck was never “I can’t query my data in natural language.” The bottleneck was “I don’t have time to look at my data at all.”

Agents solve the actual bottleneck. They don’t wait for you to ask. They surface what matters, when it matters, with recommended actions.

How to Tell the Difference

If you’re evaluating AI products and trying to separate real agents from chatbot theater, here’s a simple framework:

Ask: “What does this product do while I’m sleeping?”

If the answer is “nothing,” it’s a chatbot.

If the answer is “it monitors, it decides, it acts, and it sends me a summary of what it did”, that’s an agent.

Ask: “How many actions does this product take per week without me initiating them?”

If the answer is zero, it’s a chatbot.

Ask: “Can this product surprise me?”

Not with a wrong answer, with a right action. Can it surface something I didn’t think to ask about? Can it solve a problem I didn’t know existed yet? Can it do something useful that I never explicitly told it to do?

A chatbot can only tell you what you ask about. An agent can tell you what you didn’t know to ask.

The Honest State of the Market

We’ll be blunt: most of what the market calls “agents” today, including many well-funded startups, are chatbots. Some are very good chatbots. They use sophisticated RAG pipelines, multi-step reasoning, and tool calling to deliver impressive responses.

But responses aren’t actions. And prompts aren’t autonomy.

The companies that will win the agent era are the ones building the hard infrastructure: persistent observation systems, policy engines, action frameworks, safety mechanisms, and human-in-the-loop escalation paths. This is plumbing, not magic. It’s not as demo-friendly as a slick chat interface. But it’s what separates software that works for you from software that waits for you.

We built Apollo Space as an agent system from day one. Not because it was easier, it was dramatically harder. But because we’d lived the alternative. We’d used the chatbot CRM features. We’d tried the AI-powered dashboards. And every time, we ended up back at the same place: staring at a screen, doing the work ourselves, because the AI couldn’t take the next step without us.

The next step is what matters. And only agents take it.