Korumia

Chatbot vs assistant vs agent — the differences that actually matter

A plain-English breakdown of chatbots, AI assistants, and AI agents — what each one actually does, where the lines blur, and why Korumia is multi-agent.

Why this distinction matters

"Chatbot", "assistant", and "agent" get used interchangeably in marketing copy, which makes it genuinely hard to tell what a product actually does. This page unpacks the real differences, with examples of each category and a direct answer to why Korumia calls itself a multi-agent system. If you are deciding whether a tool is the right shape for your problem, the category matters more than the feature list.

The core spectrum is about agency — who decides what happens next. A chatbot decides nothing; its path is scripted. An assistant decides nothing; you decide every turn. An agent decides a lot — which tool to call, when to search, when to stop, when to hand off. Everything else — memory, tool use, multi-step workflows — flows from how much decision-making the system holds versus how much it hands back to you.

The table — chatbot vs AI assistant vs AI agent

DimensionChatbotAI AssistantAI Agent
Who decides the next stepThe scriptYou, turn by turnThe agent itself
Conversation scopeNarrow, rule-based flowsAny topic, one task at a timeA goal that spans many steps
Uses external toolsRarely; usually read-only lookupsIf you ask, and often one at a timeYes — picks tools, chains them, reads results
Memory across sessionsNone or simple variablesUsually short-term per chatPersistent, structured, reusable
Multi-step reasoningBranching decision treeMostly single-turn answersPlan → act → observe → revise
Can it disagree with youNoSometimes; tends to be politeYes — especially multi-agent setups
Collaboration with other AINoneNone (single persona)Yes in multi-agent systems
Typical failure mode"I did not understand that"Confidently wrong on one turnWrong plan, multiplied over steps
Cost shapeCheap per interactionCheap per turnCan be expensive per goal
Best forSupport routing, FAQs, formsWriting, quick research, Q&ADeep research, coding, strategy, ops

What a chatbot actually is

A chatbot is a conversational interface over a scripted decision tree or a narrow retrieval model. It picks from a finite menu of answers based on your input. Traditional chatbots live on support pages, insurance quote forms, and airline websites. They are cheap, reliable within scope, and catastrophically bad outside it. The moment you ask something the script did not anticipate, the chatbot says "I did not understand that" and hands you to a human or a FAQ link.

Examples in the wild. The bank support bot that walks you through "lost card or new card". The pizza chain's order bot. The insurance quote widget. The legacy rule-based bots that predate the LLM era and many that still run on top of pattern-matching rather than generative models.

Chatbots are not going away — they are the right tool for bounded, high-volume, low-stakes interactions. They are the wrong tool for anything that requires judgement, synthesis, or open-ended reasoning.

What an AI assistant actually is

An AI assistant is a conversational interface over a general-purpose language model, driven turn by turn by you. You ask, it answers. You follow up, it refines. It can write, summarise, translate, code, brainstorm, roleplay, and explain — but within any single turn, it is one voice responding to one prompt, and the next step waits for your next message.

The defining quality is that you are the orchestrator. The assistant is extraordinarily capable but not self-directed. If the first answer missed the mark, you steer it back. If the task needs three steps, you drive the three steps. Tools, when available, are often invoked on your explicit ask rather than selected autonomously.

Examples in the wild. Classic ChatGPT, Claude.ai, Gemini, Copilot's chat surface. Grammarly's rewriting assistant. Notion AI inline. GitHub Copilot's old chat pane before agent mode.

Assistants are the sweet spot for help-me-do-this-one-thing workflows: write this email, summarise this doc, refactor this function, brainstorm these names. They are weaker when the task is help-me-reach-this-goal-over-many-steps, because you end up doing the orchestration yourself.

What an AI agent actually is

An AI agent is an AI system that plans toward a goal, selects and calls tools on its own, observes results, and iterates until it finishes or hits a stopping condition. The user says "figure out why our conversion rate dropped last month" or "draft a Q2 messaging brief informed by our three biggest competitors"; the agent decides the sub-steps — read memory, search the web, open files, generate drafts, revise — without being led through each one.

Three things characterise a real agent, versus an assistant with tool-use buttons:

  1. Planning. The agent breaks the goal into sub-steps itself, rather than waiting for you to prompt each one.
  2. Tool selection. The agent chooses which tool to call at each step — search, file reading, image generation, database query, memory lookup — based on what the sub-step needs.
  3. Iteration. The agent reads the result of its own action, decides whether the goal is met, and loops back into more work if not.

Examples in the wild. Claude Code and Cursor Agent (coding agents that plan, edit, run tests, and iterate). Devin (autonomous software engineer-agent). OpenAI Deep Research and Perplexity Research (research agents that run multi-source queries). Korumia's advisory agents — CEO, Marketing, Finance, Operations — each planning inside a business-decision thread, each calling tools, each handing off to another agent via @-tag.

Single-agent systems are already useful; multi-agent systems — where specialised agents with different roles collaborate on a thread — are where the pattern starts to feel qualitatively different from an assistant.

Why Korumia is a multi-agent system, not an assistant

Korumia could have shipped as a single AI assistant with a long system prompt that says "pretend to be my CEO, Marketing, Finance, and Ops advisor depending on what I ask." That is the shape most AI products take. We chose not to because a single persona switching hats loses three things that matter for business decisions:

  • Identity persistence. The Marketing agent remembers it is the Marketing agent. Its past takes inform its future takes. A single-persona assistant flipping between roles every turn resets that continuity each time.
  • Genuine disagreement. When your Finance agent and your Marketing agent look at the same pricing decision, they surface different tensions — margin vs growth, urgency vs brand — and you see the disagreement in the thread instead of getting a single smoothed-over answer. Single-assistant setups tend to average their voices.
  • Tool autonomy per role. Each agent decides what it needs — the Finance agent reaches for memory about your ARR and unit economics; the Marketing agent reaches for web search on competitor positioning. You do not orchestrate which tool fires when; the agents do.

That is the "multi" in multi-agent: multiple named agents, each with persistent identity and its own tool-selection logic, collaborating inside one thread on your behalf.

What Korumia's agents can actually do

Inside a Korumia thread, any agent you tag can:

  • Search the web in your language, pulling current sources for competitor research, market data, or context the model does not already know.
  • Read your company memory — the structured record Korumia extracts automatically from every conversation about your ARR, ICP, product, team, and past decisions.
  • Read files you upload — PDFs, spreadsheets, docs, images — and work from their contents.
  • Generate images for moodboards, landing-page mocks, or marketing concepts.
  • @tag another agent to bring a different perspective into the same thread, passing full context automatically.
  • Remember across conversations. What you discussed with the CEO agent last week is visible to the Marketing agent today, without you re-explaining.

What they deliberately do not do yet: send emails from your account, publish posts to your channels, pay invoices, or act on the external world without your explicit confirmation. The agency is concentrated on the thinking and drafting, not on irreversible outward actions. That split is a design choice, not a limitation we are embarrassed about.

Where the lines blur

Be honest about where these categories overlap, because real products straddle them.

  • Assistants with tool buttons look agentic. Classic ChatGPT with Browse + Code Interpreter + Image Gen feels close to an agent when you click through the tools in sequence. It becomes genuinely agentic only when the model itself decides which tools to chain without your prompt.
  • Agents with narrow goals look like assistants. A coding agent asked to "fix this typo" will probably just edit one line and return. The same agent asked to "triage and fix this issue" will plan, edit, run tests, and iterate. Same system, different mode depending on goal shape.
  • Chatbots with LLMs underneath look like assistants. Many modern support bots now run on an LLM plus a retrieval layer. They feel like assistants inside their domain but collapse to chatbot behaviour outside it.

The honest way to evaluate any product that claims to be agentic: ask what decisions the system makes without you, and what decisions it still hands back. The more decisions it makes while still landing on good outcomes, the more genuinely agentic it is.

Which one do you actually need

  • You need a chatbot if your workflow is high-volume, narrowly-scoped, and low-stakes — support routing, form fills, FAQ answering.
  • You need an assistant if your workflow is "I know what I want, help me produce it faster" — writing, summarising, translating, casual research, coding help that you stay closely in the loop on.
  • You need an agent if your workflow is "here is a goal, figure out the steps" — deep research, multi-file coding tasks, strategic synthesis that spans memory + current data + multiple perspectives.
  • You need multi-agent if the goal benefits from more than one named viewpoint — strategic decisions where you want the CEO, Marketing, and Finance takes on the same question, not a single averaged answer.

Korumia sits in that last bucket on purpose. If your question is "help me write this email", a general-purpose assistant is a better fit. If your question is "should we move from monthly to annual pricing, and how would we message the transition", you want the multi-agent version.

Related reading

For the specific comparison with ChatGPT, the vs ChatGPT page goes deeper. For what each agent inside Korumia actually does, start with the AI CEO Agent and AI Marketing Agent pages. For the wider category, vs consultant and vs advisory board cover how a multi-agent system compares to human equivalents.

How Korumia works

A true multi-agent system

Agents hand off to each other via @mentions, search the web, read and generate files and images, and share a memory system so every agent has the relevant context.

Deep company context

Every agent understands your industry, your customers, your goals. No more explaining your business from scratch every conversation.

Pay only for what you use

No monthly subscriptions or commitments. Add credits when you need them, use them at your own pace. 100 credits ≈ 10 conversations.

Frequently asked questions

A chatbot answers. An assistant helps you do one task at a time when you ask. An agent pursues a goal — it can plan, decide which tools to use, call them, read the result, change direction, and keep going until the goal is done or it hits a limit. The core dividing line is agency: who decides what happens next. In a chatbot, the script decides. In an assistant, you decide turn by turn. In an agent, the agent itself decides and acts. Everything else — memory, tools, multi-step reasoning — follows from that distinction.

Historically ChatGPT started as an assistant — a conversational interface over a general-purpose language model, one task per turn, driven by the user. With tools like Browse, Code Interpreter, and the more recent 'agent mode', it crosses into agent territory for some workflows: when it decides on its own to run code, fetch a page, then write a file, it is acting as an agent. So the accurate answer is: ChatGPT is an assistant that can behave as an agent on specific tasks. The product category is fluid right now and most big-name AI products straddle the line.

Because our agents do three things an assistant does not. First, each agent has a persistent identity and role — CEO, Marketing, Finance, Operations — not a single voice switching hats. Second, they decide what to do inside a task: search the web for a competitor's pricing, read a document you uploaded, generate an image of a landing-page mock, query company memory — without you hand-holding each step. Third, they collaborate with each other — one agent @tags another, hands over context, and the second agent acts on it. That collaboration pattern is the 'multi' in multi-agent, and it is the piece single-assistant products do not offer.

Not today. Our agents are human-in-the-loop — you start a thread, they plan and act within it using tools and memory, and you stay in the driver's seat. We deliberately did not ship fully-autonomous background agents because the failure modes (runaway token spend, shipping actions to your real accounts, making decisions without oversight) are too severe for a strategic-advisory product. The near-term roadmap leans into scheduled agent actions and cross-agent workflows you explicitly kick off, not unsupervised daemons.

Concrete agent examples: a coding agent that reads a bug report, edits files, runs tests, reads the test failures, and iterates until tests pass (Claude Code, Cursor Agent, Devin). A research agent that plans a multi-source query, fetches pages, compares claims, and writes a report (OpenAI Deep Research, Perplexity Research). A Korumia Marketing agent that researches your competitor's positioning, reads your company memory, and drafts a messaging brief you can tag the CEO agent into. The thread running through all three: the agent decides the sub-steps, not you.

Both, and the ratio shifts depending on who is saying it. Used precisely, 'agentic' means the system has real agency — it plans, selects tools, evaluates results, and iterates toward a goal without you hand-holding each step. Used loosely, the term gets slapped on any chatbot with a single tool-use button and becomes meaningless. Our rule of thumb: if you can swap in 'chatbot with buttons' and the marketing still reads the same, it was not really agentic.

No — and this is the subtle point most of the industry is missing right now. More agency means more potential failure modes: an agent that decides for itself is an agent that can decide wrong for itself. For high-stakes workflows (strategy, finance, customer-facing actions), you want a lot of agency on research and synthesis, and very little on irreversible actions. That split — agentic on the thinking, human-gated on the doing — is why Korumia's agents plan, search, draft, and recommend freely but do not send emails from your account.

Ready to build your AI agents?

Join thousands of business owners making better decisions with AI agents who actually understand their business.

200 free credits • No credit card required • Setup in 2 minutes

Chatbot vs Assistant vs Agent — What's the Real Difference | Korumia