AI agent vs AI assistant: a decision table, four yes or no questions and the cut that matters to a buyer: does the tool finish the job or hand it back?
The AI agent vs AI assistant difference is initiative and reach. An AI assistant answers, drafts and summarizes when you ask, then hands the result back. An AI agent takes a goal, plans the steps and uses tools to pursue it. This page settles that in a table you can apply to any tool, then asks the better question: whether the tool finishes the job or hands it back.
AI agent vs AI assistant: the short answer
Both labels describe software built on a language model; the difference is how much of the work it carries after you speak. Run the tool you are evaluating through the five rows below; the vendor pages cited on this page do not answer the last one.
Question to ask | AI assistant | AI agent |
|---|---|---|
Who starts the work | You do, with a prompt, every time | You once, with a goal, or an event or schedule |
How many steps it takes alone | One per request, then it waits for you | A planned sequence, run in order |
Does it act in other systems | Works inside the app it lives in; suggests what to do elsewhere | Calls tools, APIs or a browser to read and write in other systems |
What comes back | A draft, an answer or a summary | A result, a log of what it did, or a deliverable such as a file |
Who finishes the job | You, by editing, sending or filing what it produced | Depends on the tool: some finish, some stop at a draft for you to act on |
The definitions cited below stop at row four. Row five is where a budget gets saved or wasted.
What is an AI assistant
An AI assistant is software you talk to that completes one task when you ask and returns the output to you. IBM’s comparison puts it in one line: “AI assistants are reactive, performing tasks at your request.” Google Cloud adds the buyer’s part: the assistant “can recommend actions but decision-making is done by the user.”
Microsoft’s overview of Microsoft 365 Copilot lists what it does in each app: in Word, “Draft, rewrite, and summarize documents”, and in Outlook, “Draft emails, summarize threads, and use the coaching tips to improve clarity, sentiment and tone”; the same page’s training section calls Copilot Chat “your AI assistant for work”. Gemini in Gmail uses Help me write to “Generate a new email draft” or “Refine existing text in your draft for tone and clarity”; you prompt, click Create, refine, and the draft is yours to send. Gemini in Google Docs shows suggestions in the document and waits for you to click Accept suggestion or Accept all.
Three things are constant: you start it, it does one step, and the result lands in front of you for the next move.
What is an AI agent
An AI agent takes a goal, decides which steps to take and uses tools to take them. IBM: “AI agents are primarily proactive, autonomously planning and taking actions to achieve a defined goal using the tools and permissions available to them.” OpenAI’s developer docs: “Agents can plan and complete tasks using tools, work with other agents, and maintain context across steps.” OpenAI’s practical guide to building agents: “Agents are systems that independently accomplish tasks on your behalf.” Anthropic’s engineering post separates agents from scripted workflows: “Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”
In Microsoft Copilot Studio, event triggers start an agent without a message from a user: “event triggers allow your agent to act autonomously in response to the defined event occurring”, including a Recurrence trigger that fires after “A set amount of time passed”. GitHub’s Copilot cloud agent “can research a repository, plan changes, and implement them in the background” and “handles branch creation, commit messages, and pushing changes”.
Now notice where every definition ends. Autonomy, planning, reasoning, tool use: each describes the mechanism, and none says what you hold at the end or who does the last step. IBM comes closest: “After an initial prompt, AI agents can continue working without further input, reducing the need for human intervention at every stage.” Reducing the need is not removing it, and no definition says where the remaining intervention sits. That gap is why the label alone cannot tell you what to deploy. Comparisons against rule-based tooling have their own pages: AI agent vs chatbot and RPA vs AI agents.
How to classify a tool you are evaluating
Four yes or no questions. Ask them with one of your real jobs in hand, not the demo job.
- Does it start without a person typing? Yes: an event, a schedule or an inbound record kicks it off. No: someone prompts it each time.
- After one instruction, does it choose and sequence its own steps? Yes: it decides the order and number of steps. No: one step per prompt.
- Does it write into systems other than the one it lives in? Creating the CRM record or pushing the commit counts; suggesting that you do it does not.
- Does the result arrive finished, or as something you must act on? A sent report is finished; a drafted report is not.
Score questions 1 to 3 together: two or three yes answers make the tool an AI agent by every definition above; zero or one, an AI assistant. Score question 4 separately: both categories can fail it, and it decides your headcount math.
A worked example. The job: every Monday, send each client a one-page summary of last week’s ad spend and leads. Tool A lives in your email client: no to question 1 (you open a compose window and prompt it), no to 2 (one draft per prompt), no to 3 (you paste the numbers in), no to 4 (a draft per client that you check and send). Verdict: an AI assistant; the Monday job stays on your calendar. Tool B runs on a schedule and reads the ads account, so questions 1 to 3 are yes. Question 4 is where to press: does the summary land in the client’s inbox or in your drafts folder, and when the ads login expires on a Monday morning, does the run finish later or come back to you? Ask the vendor to show you that Monday, not the good one.
The cut that matters: does it finish the job or hand it back?
Set two tools side by side. Microsoft’s training material calls Copilot Chat an AI assistant for work, and in Outlook it drafts the email: you read, edit and press send. GitHub’s docs call Copilot cloud agent an agent, and it researches the repository, plans, edits files, runs tests, commits and pushes. Then the docs continue: “You can review changes and request refinements before creating a pull request”, and “Logs do not replace your own review and testing.” The output is a pull request that someone on your team reviews, runs, merges or sends back.
One is labelled an assistant and one an agent, and the labels are not even stable: the Microsoft 365 Copilot overview describes GitHub Copilot as “an AI coding assistant”. Both hand the job back. For code, that handback is correct: the pull request exists so a person reviews before merge. The label told the buyer about mechanism, steps and tools, and nothing about where the work stops. A finance lead who budgets for an AI agent to run month-end reporting learns from the label that it will plan steps and call tools, not whether the reconciliation lands done in the ledger or in her inbox as a spreadsheet to finish.
Vendor documentation gives two honest reasons for handing a job back, and product design gives a third. The step needs access the tool does not hold: Copilot Studio states that “If an agent’s flow is interrupted because it can’t receive information or an action failed, it can’t continue the session.” The action is risky: OpenAI’s guide says actions that are “sensitive, irreversible, or have high stakes should trigger human oversight”. Or the product was built to produce an artifact rather than an outcome. Only the second is a feature; the other two return the job to someone’s calendar.
So the question for a vendor is not whether the product is an AI agent or an AI assistant. It is: when this runs on a Tuesday with nobody watching, what is left for a person to do, and is that step approve or reject, or finish the rest?
Where the AI employee fits
An AI employee is the category built around the fifth row. He takes a multi-step instruction in plain language, does the work in the systems the work lives in, and finishes, with a person approving the sensitive steps rather than completing the task. For the full definition read what an AI employee is; for the agent side, read how an AI employee compares with an AI agent.
Viktor is an AI employee who lives in Slack and Microsoft Teams. He has his own computer in the cloud where he writes and runs code to complete tasks. He connects to 3,200+ tools, most through one-click OAuth and some through API keys, with 27 native integrations, and builds a custom integration when a tool is missing. Money moves, code pushes and customer emails wait for explicit approval in Slack: he prepares the action as a draft and posts an approval card, and the action runs only after someone clicks Approve. Recurring work runs as Tasks, and twice a day he checks recent workspace activity to follow up on blockers. A tool gateway injects API keys at execution time so the model never sees them, and customer data never trains a model. He is SOC 2 Type 1 certified with Type 2 in progress, ISO 27001 in progress, CASA Tier 3, GDPR aligned and CCPA compliant.
A case study of a 20-person digital agency shows a team running this way. The page counts 19 of 20 people on Viktor day to day, 44 Spaces built (web apps he builds and hosts) with 38 in production, 726 Tasks run in a two-week window, and 26,000+ threads since March 2026. Its weekly Meeting Agenda Builder assembles the agenda ahead of a client meeting; it runs without a person kicking it off, and the account team reviews the agenda before it goes out. Two dated entries show the finishing part. On March 5, 2026, Viktor pulled two Fathom call recordings and a set of project documents, compiled a 42-item development task list across 14 sections and delivered it as a PDF to the account manager the same day. On April 12, 2026, a JavaScript syntax error broke a carousel component across every service page of the agency’s site; Viktor diagnosed and fixed it, and the fix was live the same day. The page also lists what he got wrong and who caught it; review before client delivery is a fixed part of how the work ships there. That is the fifth row in practice: the person’s step is review and approve, not finishing a draft.
See the tools Viktor connects to →
Give Viktor one job to finish: get started for free →
Which should you deploy?
Decide by the work, not by the label.
Questions, summaries and first drafts. Deploy an AI assistant where your people already write: Word, Outlook, Gmail, Docs. The job was always going to end with a person reading and sending, so the handback costs nothing extra.
Single-system tasks with a clear trigger. Deploy an AI agent inside that system: a record arrives, the agent classifies and routes it, and the result stays in the same tool. Settle the access question first: Copilot Studio’s documentation says event triggers can only use the agent maker’s credentials, so decide whose access the agent acts under.
Multi-step jobs across tools that must end finished. Month-end reporting that reads Stripe and QuickBooks and posts the variance in Slack. Client summaries that pull from the ads account and go out on Monday. Deploy an AI employee, gate the sensitive steps behind approval, and measure by one number: how much of the job came back to a person’s calendar.
Start with one job, not a department: the job nobody has time for, with a done state you can check. Viktor’s pricing is usage-based credits per workspace with no per-seat fee, and a new workspace starts with up to $100 in free credits and no credit card.
Pricing checked October 2026.
Book a demo and bring one job nobody has time for →
Start with up to $100 in free credits. No card required.
FAQ
What is the difference between an AI agent and an AI assistant?
An AI assistant completes one task when you ask and returns the result for you to use. An AI agent takes a goal, plans the steps and uses tools to pursue it. IBM frames it as reactive versus proactive, Google Cloud as less versus more autonomy. Neither label tells you who finishes the job.
Is ChatGPT an AI agent or an AI assistant?
Both, by OpenAI’s own description, depending on the mode. OpenAI introduced ChatGPT in 2022 as a model that “interacts in a conversational way”. Its help center now says ChatGPT includes Chat and Work: “Chat is for fast, conversational assistance and everyday questions. Work is an agent designed for longer, multi-step work and finished deliverables.” The agent mode launched in July 2025 is now listed as no longer available, and in Work you still “approve important actions”.
Can an AI assistant become an AI agent?
Yes, by adding triggers and tools. Microsoft Copilot Studio describes its agents as “AI-powered assistants that understand your business context and take action on your behalf”, and Google Cloud defines AI assistants as AI agents packaged as products that collaborate directly with users. The line moves as soon as a tool gets a trigger and write access; whether the result arrives finished does not.
Which is safer, an AI agent or an AI assistant?
Safety follows permissions, not the label. OWASP’s LLM06:2025, Excessive Agency, covers damaging actions taken on “unexpected, ambiguous or manipulated outputs from an LLM”; its controls limit permissions “to the minimum necessary” and “require a human to approve high-impact actions before they are taken”. An AI assistant that only drafts holds less write access by construction. Anything that writes needs the approval gate. Viktor gates money moves, code pushes and customer emails behind approval in Slack.
Does an AI agent still need a person in the loop?
Yes, and their builders say so. Anthropic writes that agents “can then pause for human feedback at checkpoints or when encountering blockers”, and OpenAI’s guide says a human intervention mechanism lets an agent transfer control when it cannot complete a task. The useful question is what the person does then: approve a prepared action, or finish the task.
What is an AI employee?
Software that takes a multi-step instruction in plain language, works in the systems the work lives in and finishes the job, with a person approving sensitive actions. Viktor is an AI employee who lives in Slack and Microsoft Teams, has his own computer in the cloud where he writes and runs code, connects to 3,200+ tools and runs recurring work as Tasks.