In short
- An AI agent is software in which a language model chooses its own steps to reach a goal, working inside your systems. A chatbot answers a question; an agent carries out a task.
- Since late September, OpenAI, Google and Microsoft have released agents that keep working while you are away. In Europe they are only partly available, and often still in preview.
- If you can write the steps down in advance, a fixed workflow is usually the better choice. If you do opt for an agent, settle the owner, permissions, budget and human sign-off first.
An AI agent is software in which a language model decides for itself which steps to take to reach a goal. It uses connections to your systems, such as email or a CRM, and checks the result of each step before moving on. Since late September, OpenAI, Google and Microsoft have announced agents that carry on working while you are in a meeting or have gone home for the day.
The key condition belongs right next to that definition: most business processes do not need an agent. If the steps can be written down beforehand, a fixed workflow with AI at one or two points is usually more reliable, cheaper and easier to check. An agent earns its place when the route differs from case to case and you can clearly judge the outcome. And if it is going to work unsupervised, decide up front who owns it, what it may do, what it may spend and where a person signs off.
How does an AI agent work?
Every agent comes down to three parts:
- A goal. An instruction from a colleague, or a trigger from a system such as a new customer enquiry.
- Tools. Connections that let the model fetch information or act: looking up a customer in the CRM, preparing a draft email.
- A loop. The model picks a step, carries it out, inspects the result and picks the next one. This repeats until the goal is met, a stopping condition is hit or the model asks a person for a judgement.
Take a customer who emails to ask when their order will arrive. In this fictional example, an agent finds the customer, retrieves the order, spots that delivery is delayed and prepares a reply with the new date for a colleague to check and send.
Agent, workflow or chatbot: what is the difference?
Anthropic, the company behind Claude, draws the line like this: in a workflow, the model and tools follow predefined paths; in an agent, the model directs its own process (Anthropic). Plenty of products sold as agents actually follow a fixed route. That is fine, as long as you know which one you are buying.
| Type | Who decides the steps? | Suits |
|---|---|---|
| Chatbot or assistant | The user, per question | Answering questions, drafting text |
| Workflow with AI | Set in advance | Recurring work in a fixed order |
| AI agent | The model, per case | Tasks whose route is unknown in advance |
What is agentic AI?
Agentic AI is the umbrella term for AI that carries out a task using tools and several steps, rather than just answering a question. Anthropic calls these agentic systems and counts both workflows and agents among them. So the label tells you nothing about how much freedom a product has. For anything sold as agentic, ask: does the model choose the steps, or does it follow a route set in advance?
Which AI agents keep working while you are away?
OpenAI, Google and Microsoft announced four such agents within three weeks. This is the position on 15 October 2026:
| Agent | What it is | Netherlands/EU, status |
|---|---|---|
| OpenAI dots | Always on, with its own cloud computer and browser | Business Premium; beta in Enterprise. Not on Pro. Rolling out |
| Gemini app, tasks mode | Toggle from chat to tasks, on Gemini 4 Argon | AI Pro and Ultra, not in the EEA or the UK. Since 13 October |
| Gemini agent for Workspace | An agent for work, optionally with its own account | Selected customers. Private preview, no date |
| Microsoft Copilot Autopilot | An agent that keeps working on tasks on its own | Private preview, announced for late September |
OpenAI dots run on GPT-6 Astra and keep going when your own computer is switched off. Pro subscribers in the European Economic Area, the UK and Switzerland cannot use them, and in Enterprise they are a beta that an administrator has to switch on (ChatGPT documentation, admin documentation). Data residency and cost are covered in OpenAI dots: what can you do with them in Europe?
Tasks mode in the Gemini app is a toggle in the chat window. Gemini then keeps working across your Google apps and more than 40 connected apps while you are away (Google). It only works with a personal account on AI Pro or Ultra, not with a work or school account, and it is not available in the European Economic Area, Switzerland or the United Kingdom (Gemini Apps Help).
The Gemini agent for Workspace is Google's agent for work, summed up as "You give it objectives, not instructions." A coworker agent gets its own account with an email address, and every action is logged against the agent rather than a person (Google Cloud). It is in private preview, according to VentureBeat. More in Gemini agent for Workspace.
Autopilot is part of the new Microsoft Copilot: an agent that carries on with tasks independently, even when the user is away from their computer. On 25 September Microsoft announced that Autopilot would move into private preview at the end of that month. According to Microsoft, Cowork, Code and Autopilot all run on usage-based billing (Microsoft). That usage is charged in Copilot Credits, but the admin documentation does not yet list Autopilot among the services it covers (Microsoft Learn). How to cap that spending: Microsoft Copilot and Copilot Credits.
All four are new, most are in preview or still rolling out, and pricing is largely unknown. Do not build a critical process on any of them yet.
What to settle before an agent works while you are away
An agent that works unsupervised has access to systems and data, and nobody is watching in real time. So settle four things first:
- An owner. One person who answers for what the agent does, maintains its instructions and can switch it off. A name, not "the team".
- Permissions. Start with reading and suggesting, and connect as few apps as possible. With dots, a check before any action that touches your accounts or shares information decides whether the agent may proceed, needs your approval or hands the step back to you (ChatGPT documentation).
- A budget. More steps means more spend. Google lets you set a hard spending cap per project, after which the agent pauses. In Microsoft you set limits for Copilot Credits. Do this before the first task runs.
- Human sign-off. Anything that reaches customers, money or publications goes past a person first. Keep a record of what the agent did, and agree who reads it.
We turned this into five questions in Using AI agents safely.
Why most organisations should start with a workflow
Anthropic's advice is sober: find the simplest solution that works. Agentic systems trade speed and cost for better results on open-ended tasks, and because they take several steps on their own, errors can compound. For a quote or an invoice you want the same rules every time, a clear view of where things went wrong and costs you can predict.
An example from our own work. For a European supplier of fitness equipment we built a content portal that fills 82 text fields per product in six languages. AI checks product benefits against the specifications, writes the copy and translates it. The client's team manages the writing instructions and reviews everything before it is written back to the product information system. That is a workflow: the order is fixed, and at every point you can see what AI does and what the editor decides.
When is an AI agent the right choice?
Use these four questions as a decision rule:
- Can you write the steps down in advance? Yes: build a workflow. No, because the route depends on what turns up along the way: then an agent comes into view.
- Does the software need to look things up or act inside your systems? No: an assistant or chatbot will do.
- Can you tell whether the result is good? A resolved ticket or a passing test can be checked; "a good strategy" cannot. Without a success criterion, the agent cannot tell when it is done, and neither can you.
- What happens if it goes wrong while nobody is watching? The bigger the potential damage, the more steps a person should approve.
Anthropic sees the most value in customer support and software development: conversation combined with action, clear success criteria and room for human oversight. If you recognise that in one of your processes, it is a serious candidate.
Where do you start?
Pick one recurring process that eats time. Write out the steps and mark every point where someone makes a choice. If those choices keep following the same rules, you have a workflow. If they genuinely differ per case and the result can be checked, trial an agent with tight permissions and review after two weeks what you had to correct. Not sure yet where your organisation stands with AI? Start with the self-assessment in AI agents in SMEs.
Want to learn to build this yourself? In the workshop AI workflows and agents you break down one of your own processes and turn it into a working flow.
Want to know whether a process in your organisation needs an agent, a workflow or neither? Book a conversation.
Sources
- Anthropic, Building effective agents, 19 December 2024
- OpenAI, ChatGPT documentation: dots, accessed 15 October 2026
- OpenAI, ChatGPT Enterprise: cloud and local access, accessed 15 October 2026
- Google, Gemini 4 availability, 13 October 2026
- Gemini Apps Help, Manage your tasks & workflows in Gemini Apps, accessed 21 October 2026
- Google Cloud, Welcome to Gemini at Work 2026: Introducing the Gemini agent, 8 October 2026
- VentureBeat, Google Cloud unveils persistent Gemini Agents for long-running tasks, 8 October 2026
- Microsoft, Introducing the new Copilot with Home, Code and Autopilot, 25 September 2026
- Microsoft Learn, Usage-based billing and cost management for Copilot Credits, accessed 15 October 2026
