Roundups
AI agent vs chatbot: what actually counts as an agent
An AI chatbot converses within a defined scope and answers questions; an AI agent perceives its environment, reasons about a goal, decides which tools to use, and executes multi-step actions across systems with the LLM directing its own process. Gartner estimated in 2025 that only around 130 of the thousands of vendors marketing "AI agents" were genuinely agentic by that definition — meaning roughly 95% of the market was, by Gartner's own label, rebranded chatbots, RPA, or simple assistants.
Last updated Aug 21 · 18 min read
The core distinction: autonomy over process, not just tool access
Multiple vendor sources — Aisera, Make.com, and Zendesk among them — converge on the same core distinction despite writing for different audiences: a chatbot operates within a defined conversational scope, responding to inputs and answering questions, while an agent perceives its environment, reasons toward a goal, selects among available tools, and carries out multi-step actions across systems on its own (Aisera; Make.com; Zendesk).
The word doing the real work in that definition is "autonomy," not "tools." A chatbot can absolutely call a function or a tool — looking up an order status, checking a database — and still be a chatbot, because a human or a predefined script is still deciding when and how that tool gets called. What separates an agent isn't that it has tool access; it's that the LLM itself is making the decision about which tool to use, in what order, and when the task is actually done. That's the distinction practitioner debate on Hacker News lands on as well: tool access alone doesn't make something an agent — autonomous decision-making over that tool use does (Hacker News).
How Anthropic defines agents vs. workflows
Anthropic's own engineering guidance, "Building Effective Agents," gives the most precise and widely cited technical definition in this space. It draws a sharp line between workflows and agents: workflows are systems where LLMs and tools are orchestrated through predefined code paths — the sequence of steps is fixed by the developer, even if an LLM is doing some of the reasoning inside each step. Agents, by contrast, are systems where the LLM dynamically directs its own process and tool use, maintaining control over how it accomplishes a task rather than following a script someone else wrote for it (Anthropic).
Anthropic's guidance also lays out five specific workflow patterns that stop short of full agentic autonomy but are often confused with it: prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer. All five involve LLMs and tools working together in genuinely sophisticated ways — they're not "just a chatbot" in the simple sense — but they all still run on predefined paths a developer designed in advance, rather than an LLM freely deciding its own next step. This is a big part of why the agent-vs-chatbot line gets blurry in marketing: a well-built orchestrator-workers system looks impressively autonomous from the outside while remaining, by Anthropic's own definition, a workflow rather than an agent.
Anthropic's practical recommendation follows directly from this: start with the simplest solution — typically a workflow — and only add genuine agency when the added flexibility is worth the added latency, cost, and compounding-error risk that comes with letting a model direct its own multi-step process (Anthropic; analysis via Mer.vin). Most production systems, per this guidance, don't actually need full autonomy — which is itself a useful signal for evaluating vendor claims: a genuine need for agentic autonomy should be the exception a vendor can justify, not the default marketing frame every product reaches for.
What "agent washing" is and how big the gap actually is
Gartner named the phenomenon "agent washing" in 2025 and put a specific, striking number behind it: of the thousands of vendors marketing products as "AI agents," Gartner estimated only about 130 were genuinely agentic by a reasonable technical definition — meaning roughly 95% of self-described "AI agent" products were, in practice, rebranded chatbots, robotic process automation (RPA), or simple assistants wearing new marketing language (Xpert.digital).
A related concern that commentators flag specifically: retrieval-augmented generation (RAG) systems being marketed as "autonomous knowledge systems" is called out as a textbook example of agent washing, since retrieval on its own doesn't imply autonomous decision-making or the ability to take action — it's a lookup capability, not agency (Particula.tech). This matters because RAG-based products are extremely common and genuinely useful; the issue isn't that they're bad products, it's that calling them "agents" overstates what they actually do.
The gap between buying enthusiasm and real deployment adds context to why this mislabeling is so widespread right now. Gartner's 2026 CIO survey found only about 17% of organizations had actually deployed AI agents, while over 60% expected to within two years — a large gap between intent and reality that gives vendors strong incentive to apply the "agent" label to whatever they're already selling, ahead of the market actually catching up to genuine agentic deployment (DigitalApplied). Gartner itself goes further on this point, predicting in a June 2025 release that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls — a direct consequence of unclear value and mislabeled capabilities purchased under the "agent" banner not holding up once organizations try to operationalize them (Gartner).
How to tell a real agent from a relabeled chatbot
Practical buyer scorecards converge on a small set of concrete questions rather than a single test. Does the system genuinely decide, on its own, which tool to use next based on the state of the task — or does it follow a fixed sequence a developer wrote? Can it act across multiple systems (not just retrieve and respond within one interface)? Is it working toward an open-ended goal, or answering one bounded question at a time (DigitalApplied; Entagl)?
There's also a security dimension worth taking seriously in this evaluation, and it's not a marketing nuance. A chatbot that produces a wrong or unsafe piece of text is dangerous in a bounded way — the damage is limited to bad information reaching a user. A genuinely agentic system that can take real actions — deleting data, calling paid APIs, executing transactions — changes the entire threat model, because a reasoning error doesn't just produce bad text, it produces a bad action that already happened (Hacker News). This is a real, practical reason "is this actually an agent" is worth answering precisely before deployment, not just a semantic argument about labeling.
Practical examples
– Real, per Anthropic's own guidance: a customer support system that combines a familiar chatbot-style interface with genuine tool-integrated autonomy underneath — Anthropic specifically names this pattern as a promising, real agent use case, showing the interface alone (looking like a chat window) isn't what disqualifies or qualifies something as an agent; the decision-making loop behind it is (Anthropic).
– Real, per Gartner's framing: a RAG-based document search tool marketed as an "autonomous knowledge agent" — genuinely useful for retrieval, but not agentic by the autonomy definition, since it doesn't independently decide on and execute multi-step actions toward a goal.
– Illustrative, not a documented case: picture a scheduling tool that a vendor calls an "AI agent" but that, on inspection, only ever follows one fixed sequence — check calendar, propose three times, send email — with no branching decision-making based on what it finds. That's a workflow by Anthropic's definition, however capable it looks in a demo.
Data and evidence
– ~130 out of thousands of vendors marketing "AI agents" were assessed by Gartner as genuinely agentic in 2025 — roughly a 95% mislabeling rate across the category (Xpert.digital).
– Only ~17% of organizations report having actually deployed AI agents, versus over 60% expecting to within two years, per Gartner's 2026 CIO survey (DigitalApplied) — a large intent-to-deployment gap.
– Five named workflow patterns — prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer — are documented by Anthropic as predefined-path systems that fall short of full agentic autonomy, despite often looking sophisticated (Anthropic).
– Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls — tied specifically to unclear value and capabilities mislabeled at purchase time (Gartner; DigitalApplied).
– Evidence not sufficiently verified: there is no single, universally agreed technical certification or test suite that formally determines "real agent" status across the industry — the criteria above are converging practitioner consensus and named vendor/analyst frameworks, not a codified standard.
Comparisons: agent vs. chatbot vs. workflow
System type: Chatbot · Who decides the next step: A human, or a fixed conversational script · Can it act across multiple systems?: Rarely, and usually only via a single predefined tool call · Typical marketing label risk: Low — usually honestly labeled
System type: Workflow (Anthropic's five patterns) · Who decides the next step: A developer-defined code path, even with an LLM reasoning inside each step · Can it act across multiple systems?: Sometimes, but only along pre-built paths · Typical marketing label risk: High — often marketed as "agentic" despite being predefined
System type: AI agent · Who decides the next step: The LLM itself, dynamically, based on the state of the task · Can it act across multiple systems?: Yes, in principle, across whatever tools/APIs it's given access to · Typical marketing label risk: Should be low if genuinely agentic, but frequently applied to workflows or chatbots too
The practical risk sits squarely in the middle row: workflows are sophisticated enough to demo convincingly as "agents," which is exactly the gap Gartner's agent-washing estimate is describing.
Real-world use cases
Anthropic's own published guidance names customer support as a genuinely promising real-world agent use case specifically because it combines a familiar interface (a chat window customers already know how to use) with underlying tool-integrated autonomy — checking order systems, issuing a refund within defined limits, escalating when the task exceeds its authority — that goes beyond a fixed script.
The RAG-labeled-as-agent pattern flagged by Particula.tech is a second, very common real-world case worth naming directly: countless production retrieval systems genuinely help users find information inside documents or databases, and are legitimately useful — the issue specifically flagged is the "autonomous" or "agent" label applied to a system that is fundamentally a lookup tool, not a decision-making one.
Common mistakes
– Assuming tool access equals agency. A chatbot that calls a function to check an order status is still a chatbot if a human or a script decided when to call it — the deciding factor is who's directing the process.
– Buying based on the word "agent" in a product name or pitch deck, rather than asking the specific autonomy questions (who decides the next step, can it act across systems, is it goal-directed) that separate a real agent from a workflow.
– Treating a well-built workflow as inferior to "a real agent." Anthropic's own guidance recommends workflows as the right default for most production systems — the mistake is mislabeling one as the other, not choosing one over the other.
– Underestimating the security implications of genuine autonomy. Evaluating a system's "agent" claim only on capability, without asking what happens if it takes a wrong action rather than just generates wrong text, misses the threat-model shift that comes with real agency.
– Assuming a RAG-based product is automatically an agent because it's "AI-powered" and does something impressive. Retrieval and generation, however sophisticated, aren't the same as autonomous multi-step decision-making.
Best practices for evaluating an "AI agent" claim
1. Ask directly: who decides what happens next — the model, or a predefined path a developer wrote? This single question, more than any feature list, separates an agent from a workflow per Anthropic's definition.
2. Ask whether the system can act across multiple systems/tools autonomously, not just retrieve information or respond within one bounded interface.
3. Start with the simplest architecture that solves the actual problem, per Anthropic's own recommendation — don't reach for agentic autonomy by default just because a vendor's pitch assumes you should.
4. Weigh the added latency, cost, and error-compounding risk of genuine agency against the flexibility it buys you, explicitly, before committing to an "agent" architecture over a workflow.
5. Treat the security implications of real autonomy as a first-order evaluation criterion, not an afterthought — a system that can take real actions needs a different risk review than one that only generates text.
6. Use a buyer's scorecard (autonomy, goal-orientation, adaptive tool selection, cross-system action) rather than trusting the vendor's own label, given the documented ~95% mislabeling rate Gartner found across the category.
Key takeaways
– The core distinction between an agent and a chatbot is autonomy over the process — who decides the next step — not whether tools are involved.
– Anthropic's own framework defines agents as systems where the LLM dynamically directs its own process, versus workflows that follow predefined code paths; most production systems are better served by the latter.
– Gartner estimated only about 130 of thousands of self-described "AI agent" vendors were genuinely agentic in 2025 — roughly a 95% mislabeling rate.
– A genuinely agentic system changes the security threat model, since a reasoning error can produce a real, already-executed action rather than just bad text.
– Only about 17% of organizations report actually deploying AI agents despite over 60% expecting to within two years — a gap that helps explain why agent washing is so widespread right now.
Relevant tools.scult.in resources
The AI engineering prompt library is a practical place to see the difference between workflow-style prompting (fixed sequences) and genuinely agentic prompt design (dynamic, self-directed tool use) in practice, rather than just in theory.
If you're trying to figure out whether your own use case genuinely needs an autonomous agent — with the added latency, cost, and risk that comes with real autonomy — or whether a well-built workflow would actually serve you better and faster, that's precisely the kind of scoping question worth bringing to a team that builds AI agents and automation rather than resolving from a vendor's pitch deck alone.
For a related, free starting point, try the AI Visibility Checker.
Frequently asked questions
What is the fundamental difference between an AI agent and a chatbot?
A chatbot converses within a defined scope; an agent perceives inputs, reasons toward a goal, selects tools, and executes multi-step actions with the LLM directing its own process, per converging vendor and Anthropic definitions.
What is an AI agent?
A system where an LLM dynamically directs its own process and tool use to accomplish a goal, rather than following a predefined sequence of steps written by a developer (Anthropic).
What is agentic AI?
A general term for AI systems exhibiting agent-like autonomy — reasoning, planning, and acting across multiple steps or tools toward a goal — as opposed to single-turn, scripted, or purely conversational AI.
Is a chatbot with function calling an AI agent?
Not necessarily — if a human or a fixed script decides when and how the function gets called, it's still a chatbot with tool access, not an agent; the deciding factor is whether the LLM itself is directing that decision.
What is agent washing?
The practice of marketing ordinary chatbots, RPA, or simple assistants as "AI agents" without genuine autonomous decision-making — a term Gartner used after finding roughly 95% of self-described "agent" vendors didn't meet that bar.
What percentage of AI agents on the market are genuinely real, per Gartner?
Gartner estimated only about 130 out of thousands of vendors marketing "AI agents" were genuinely agentic — roughly 5% of the labeled category.
How does Anthropic define an agent versus a workflow?
Workflows use predefined code paths to orchestrate LLMs and tools; agents let the LLM dynamically direct its own process and tool use, maintaining control over how it completes a task.
What are the five workflow patterns Anthropic describes before recommending full agents?
Prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer.
When should you build a workflow instead of a fully autonomous agent?
When the task doesn't require the added flexibility to justify the extra latency, cost, and compounding-error risk that comes with full autonomy — Anthropic recommends starting with the simplest workflow solution by default.
Is a RAG system automatically an AI agent?
No — retrieval-augmented generation is a lookup/generation capability; marketing it as an "autonomous knowledge agent" is specifically flagged as a common agent-washing pattern.
Does tool access alone make an LLM system an agent?
No — practitioner consensus (including Hacker News debate) lands on autonomous decision-making over tool use, not mere access to tools, as the deciding factor.
Why does the security threat model change once a system is a true agent rather than a chatbot?
Because an agent can take real actions (deleting data, calling paid APIs, executing transactions) — a reasoning error produces an action that already happened, not just bad text a human can disregard.
What percentage of organizations have actually deployed AI agents?
About 17%, per Gartner's 2026 CIO survey, versus over 60% expecting to within two years.
Why might agentic AI project cancellations rise by 2027?
Gartner itself projects over 40% of agentic AI projects will be canceled by end of 2027, pointing to escalating costs, unclear business value, and inadequate risk controls — the same unclear-value pattern that shows up when capabilities are mislabeled as "agentic" at purchase time.
Is customer support a realistic use case for a genuine AI agent?
Yes — Anthropic specifically names it as a promising application, combining a familiar chat interface with underlying tool-integrated autonomy.
What's the practical risk of buying a "workflow" while believing you bought an "agent"?
You may expect flexibility and autonomous problem-solving the system was never built to provide, since it follows a fixed path rather than dynamically deciding its own steps.
Do vendors have an incentive to mislabel their products as agents?
The large gap between deployment (17%) and buying intent (over 60% within two years) creates strong incentive for vendors to apply the more in-demand "agent" label ahead of actually building agentic capability.
Is "AI agent" a regulated or legally defined term?
No — the definitions cited here come from vendor documentation, Anthropic's engineering guidance, and analyst commentary (Gartner), not a legal or regulatory standard.
Can a chatbot become an agent by adding more tools?
Adding tools alone doesn't create agency; the system also needs the LLM itself deciding when and how to use those tools across multiple steps toward a goal, rather than following a fixed script.
What's the relationship between "autonomy" and "goal-orientation" in defining an agent?
Both are named criteria in buyer scorecards: autonomy is about who decides the next step; goal-orientation is about whether the system is working toward an open-ended outcome rather than answering one bounded query.
How do I tell if a product is really an AI agent?
Ask whether the LLM itself decides the next action dynamically, whether it can act across multiple systems, and whether it's working toward a goal rather than answering one fixed query — a real agent should score yes on all three.
How do I spot agent washing in a vendor pitch?
Look past the word "agent" in the marketing copy and ask for a concrete description of the decision-making loop — if it's a fixed sequence of steps, however sophisticated, that's a workflow, not an agent by Anthropic's definition.
How do I decide between building a workflow and building a full agent?
Start with a workflow by default, per Anthropic's guidance, and only move to a full agent if the task genuinely requires flexibility that a predefined path can't provide, given the added latency, cost, and error risk.
How do I evaluate the security risk of a genuinely agentic system before deploying it?
Assess what happens if the system takes a wrong action, not just what happens if it generates wrong text — the failure mode of a true agent includes real-world consequences a chatbot's failure mode doesn't.
How do I use a buyer's scorecard to catch agent washing?
Score the product against explicit criteria — autonomy, goal-orientation, adaptive tool selection, cross-system action — rather than relying on the vendor's own "agent" label.
How do I build a genuinely agentic customer support system rather than a chatbot with tools?
Design it so the model itself decides which action to take (check order, issue refund within limits, escalate) based on the state of the conversation, rather than following one fixed script for every case — mirroring the pattern Anthropic names as a real agent use case.
How do I explain the agent-vs-workflow distinction to a non-technical stakeholder?
Frame it as "who's driving": in a workflow, a human pre-planned every turn the car takes; in an agent, the model is deciding the route itself as it goes.
How do I audit my own product's marketing claims for agent-washing risk?
Compare your own description against Anthropic's definition (dynamic LLM-directed process control) rather than checking whether your product merely uses an LLM with tool access.
What percentage of RAG products get marketed with agent-washing language?
No specific percentage was found in the research reviewed; the pattern is named as common by industry commentary (Particula.tech) rather than quantified precisely.
Is there a certification or audit process to verify an "AI agent" claim?
No formal, industry-wide certification process was found in this research; the available tools are practitioner scorecards and analyst frameworks (e.g., DigitalApplied's), not a certifying body.
Does adding memory or persistent state to a chatbot make it an agent?
Not by itself — memory affects what the system knows across turns, but agency specifically concerns whether the LLM is directing its own multi-step process and tool use, a separate architectural question.
Is multi-agent orchestration (multiple LLMs coordinating) a distinct category from a single agent?
Anthropic's orchestrator-workers pattern involves multiple LLM calls coordinating on subtasks, but it's still described as a workflow (predefined orchestration path) unless the orchestrating logic itself is dynamically directed by the model rather than fixed code.
How does OWASP's involvement relate to agentic AI risk?
OWASP is referenced in this topic's broader research context as an organization tracking security risks specific to more autonomous, tool-using AI systems, reflecting the threat-model shift discussed above; specific OWASP agentic-AI guidance details weren't independently verified in this research.
Is Salesforce Agentforce a genuine agent by Anthropic's definition?
This research did not independently verify Agentforce's internal architecture against Anthropic's specific autonomy criteria — treat vendor-specific agent claims, including this one, with the same scorecard-based scrutiny recommended throughout this article rather than as pre-verified.
Does an agent need to use multiple tools to qualify as an agent, or can it use just one?
The number of tools isn't the deciding factor — a system using a single tool but dynamically deciding when and how to invoke it toward an open-ended goal can still qualify as agentic; tool count is a separate axis from autonomy.
AI agent vs. AI chatbot vs. LLM workflow — what's the practical difference in production?
A chatbot answers within a bounded scope; a workflow executes a predefined multi-step path (possibly using LLMs at each step); an agent lets the LLM dynamically direct its own steps and tool use — increasing autonomy and increasing latency/cost/risk in that order.
Anthropic agents vs. workflows — which does Anthropic recommend building first?
Anthropic explicitly recommends starting with the simplest workflow solution and only adding agentic autonomy when the added flexibility is worth the tradeoffs.
Agentic workflow vs. autonomous agent — is there a meaningful difference, or is it just terminology?
By Anthropic's technical definition, yes: a workflow (agentic-sounding or not) still runs on a predefined path; an autonomous agent has the LLM directing that path itself — the label "agentic workflow" is sometimes used loosely to blur that line.
How does a real AI agent compare to traditional RPA (robotic process automation)?
RPA typically follows rigid, rule-based scripts with no model-driven reasoning about what to do next; a genuine AI agent reasons and adapts its actions dynamically — RPA rebranded as an "agent" is one of the specific patterns Gartner's agent-washing framing calls out.
How does a genuine agent compare to a simple AI assistant (e.g., an email autocomplete feature)?
A simple assistant typically performs one bounded, assistive task without independently deciding on a sequence of actions toward a broader goal — it lacks the multi-step, self-directed process that defines agency.
I bought a product marketed as an "AI agent" but it only ever follows the same fixed steps — is that expected?
That's a signal it may be a workflow rather than a true agent by Anthropic's definition; ask the vendor directly whether the model itself decides the next step or whether it's following a predefined sequence.
My team built what we call an "agent" but it can't handle any situation outside its script — did we build an agent?
Probably not in the strict sense — if it can't adapt its own next step based on the task state, it more closely matches Anthropic's definition of a workflow, however useful it may be.
Our "AI agent" keeps taking actions we didn't want because it has too much autonomy — how do we fix that?
This is exactly the security/threat-model tradeoff that comes with genuine agency; the fix is typically narrowing its permitted actions and adding explicit approval gates for higher-risk actions, not necessarily reducing it back to a chatbot.
Customers keep asking if our product is a "real" agent and we're not sure how to answer — what should we say?
Answer honestly against the criteria in this article — who decides the next step, can it act across systems, is it goal-directed — rather than defaulting to marketing language, given how closely buyers are now scrutinizing this claim.
We're getting pushback that our "agent" is just a chatbot with a plugin — is that a fair criticism?
If a human or fixed script still decides when your plugin/tool gets called, that criticism is technically accurate by the autonomy-based definitions in this article — worth addressing directly rather than dismissing.
Should I buy an "AI agent" product or build a simpler workflow myself?
Per Anthropic's own guidance, most production needs are better served by a workflow; only choose a genuine agent product if your use case specifically requires the flexibility that autonomy provides, and you've weighed the added cost/risk.
What questions should I ask a vendor before trusting their "AI agent" claim?
Ask who/what decides the next action, whether it can act across multiple systems autonomously, what happens when it's wrong, and whether they can describe their architecture as something other than a fixed script.
Is it worth paying a premium for a "true" agentic product over a cheaper workflow-based one?
Only if your use case genuinely needs the autonomy — Anthropic's guidance is explicit that added flexibility isn't free, and a workflow solving the same problem at lower cost and risk is often the better choice.
How do I avoid buying (or building) an agent-washed product for my team?
Use a concrete scorecard — autonomy, cross-system action, goal-orientation, adaptive tool selection — rather than trusting a vendor's "agent" label, given Gartner's ~95% mislabeling estimate across the category.
Where can I get help figuring out whether my use case actually needs a real agent or just a well-built workflow?
That's exactly the kind of architecture decision worth scoping with a team that builds agentic and workflow-based automation for a living, rather than guessing from a vendor's marketing page.
Sources
- https://aisera.com/blog/ai-agent-vs-chatbot-differences/
- https://www.make.com/en/blog/AI-agent-vs-chatbot
- https://www.zendesk.com/blog/ai/chatbots/ai-agents-vs-ai-chatbots/
- https://www.anthropic.com/engineering/building-effective-agents
- https://mer.vin/2026/05/when-not-to-build-ai-agents-anthropics-workflow-vs-agent-playbook/
- https://xpert.digital/en/the-label-fraud/
- https://particula.tech/blog/agent-washing-real-vs-fake-ai-agents
- https://www.digitalapplied.com/blog/agent-washing-definition-buyers-scorecard-2026
- https://www.entagl.com/blog/agent-washing-real-ai-agent-vs-rebranded-chatbot
- https://news.ycombinator.com/item?id=44623207
- https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
Need this built into your business?
The free tools and prompts on this site handle the small, solved problems. If what you need is bigger — AI agents & automation, built and maintained for you — that's Scult's day job.
