You can build an AI agent today in Claude, ChatGPT, Microsoft Copilot, Google Gemini, or a tool like n8n, and by now all four major vendors call their offering an "agent." The difference isn't in the name, it's in how much memory, file access, and tool use you get without extra work. I built my entire workforce of eight AI employees on Claude. What follows sorts the five approaches by what the vendors document and by what I have noticed working with the tools myself.
I've collected all the fundamentals on agents, chatbots, and assistants on the AI agents page. How I actually build an agent in Claude, in five steps, is in Build an AI Agent With Claude: A Guide; this piece is about the choice that comes before that: which tool for which case.
What the five approaches mean by "agent"
In Claude, an agent is a project: a workspace with its own instructions and a knowledge base that every conversation inside it can access, plus tools such as web search, file handling, and optional integrations.
In ChatGPT, the comparable unit is a Custom GPT: your own assistant with instructions, optional knowledge, and optional actions that call external services. Microsoft calls its builder Copilot Studio, aimed by its own documentation at organizations already running Microsoft 365 and Dataverse who want to connect their data directly. Google bundles its version under Gemini, with customizable personas and a tie-in to Workspace services. n8n isn't a chat assistant at all; it's an automation platform, where an agent is a node in a flow that triggers other services through their interfaces, with a lot more technical setup but also more control over fixed processes.
Five different starting ideas, in other words, not five interchangeable versions of the same thing.
Tools compared
The table below puts the documented product features side by side and adds what stands out to me from working with the tools. There is no comparison test behind it. I only give prices where I can back them up.
| Tool |
Files and knowledge |
Memory across sessions |
No-code start |
Pricing |
| Claude (project) |
own knowledge base per project |
yes, project-wide |
yes |
Free 0 dollars, Pro 20 dollars a month, Max 100 or 200 dollars, Team from 25 dollars per seat (Claude Pricing and Plans Explained, as of September 2026) |
| ChatGPT (Custom GPT) |
knowledge uploadable per GPT |
limited, mostly per conversation |
yes |
free entry tier, full feature set tied to a paid plan |
| Copilot Studio |
connects to Microsoft 365 data |
yes, via Dataverse |
more configuration, less no-code |
tied to Microsoft 365 and Copilot licenses |
| Gemini |
connects to Google Workspace services |
partial, depends on the integration |
yes |
free entry tier, full feature set tied to a paid plan |
| n8n |
via database and file nodes |
none out of the box, you build it yourself |
no, needs technical setup |
self-hosted is cheaper, cloud plan is paid |
The row that matters most is the second one: memory across sessions. An agent that starts from zero in every new conversation isn't an employee, it's a tool you have to brief again every single time.
Olaf in ChatGPT or Copilot: what you would have to check
Olaf is my lead scout: he finds and rates potential customers against a description of what a good-fit client looks like, plus a running list of cases already checked. In Claude, he gets a task like "check this company and tell me if it fits our profile," works with what's stored in his working folder, and writes the result back into that same knowledge base, so the next check builds on it.
If I moved the same task into a Custom GPT in ChatGPT or into an agent in Copilot Studio, the product descriptions point to two things I would want to check first. One is memory beyond a single conversation: a Custom GPT works from stored instructions and knowledge files, not from a continuously updated result list, so you would have to establish how a growing list of already-rated companies is maintained there at all. The other is data storage in Copilot Studio, whose strength is the connection to Microsoft 365 and Dataverse, which in turn assumes the prospect list is neatly sorted into a Microsoft data source first. Those are the two questions I would answer in a small trial run before moving anything, not from the product pages.
Olaf still runs on Claude for me, because the knowledge store sits right next to the role and there is nothing extra to maintain. For a company already running entirely on Microsoft 365, Copilot Studio can be the more obvious choice, because all the customer data already lives there. That is a read along the requirements, not a comparison test I ran.
Why my workforce runs on Claude, downsides included
The main reason is that role and knowledge sit together in one place for me: every AI employee has its own working folder in the file system, holding a personnel file, rules, skills, and learning and memory files. Whatever I write in there once applies on every later start. The full reasoning, downsides included, is in Claude Over ChatGPT: My AI Employees.
To be fair, there's plenty Claude doesn't do better. Recurring, scheduled processes across many different services, say checking several inboxes every morning automatically and routing results into three different systems, are n8n's strength, not Claude's: with Claude you start a conversation, in n8n the flow just runs on a schedule on its own. If you already work entirely inside Google Workspace, Gemini gets you to your own files faster; if you live in Microsoft 365, the same is true for Copilot.
Building the AI agent in four stages
The process stays the same regardless of the tool. First, pick a task that repeats and whose result you can judge quickly. Second, choose the tool based on where your data already lives, not on whatever is being advertised loudest right now. Third, write instructions the way you would brief a new hire on day one, and attach the knowledge the agent needs for the job; how that looks in Claude specifically, with example instructions and a knowledge base, is spelled out step by step in Build an AI Agent With Claude: A Guide. Fourth, test on a real case, name the mistakes instead of fixing them yourself, and feed every correction that should apply more than once back into the instructions.
If money is what's holding you back before you commit: what actually works on the free tiers, I've gone through separately for Claude, ChatGPT, and n8n in Build an AI Agent for Free: What Works?. And if you're torn more fundamentally between a project approach like this one and a workflow platform like n8n, the full reasoning for that choice is in Building AI Agents Without Code: My Way.
Frequently asked questions
Can I build an AI agent without code?
Yes, all five approaches compared here let you start without code: a Claude project, a Custom GPT, a Gemini persona, and the basic features of Copilot Studio can be set up through forms and plain text. Only if you build custom flows in n8n between services that have no ready-made node do you really need some technical grounding.
Is a Custom GPT an agent or an assistant?
Both, depending on how it's equipped. Without actions that call external services, a Custom GPT stays an assistant with its own knowledge. Only once it has connected tools does it become an agent in the stricter sense, one that does things on its own instead of only answering.
Is an AI agent GDPR-compliant?
That depends less on the tool than on what you feed into it and where the provider processes the data. The ground rules I apply myself, such as which customer data is allowed into which system, are in AI and Privacy: What the AI Gets to See.
Can I switch from one tool to another later?
The instructions themselves usually carry over with little effort, they are text in the end. What doesn't carry over is the history in the knowledge store; you have to rebuild that in the new tool or export it cleanly before you switch.
How many agents does a one-person business need?
I run eight, each with exactly one job. To start, one is enough: take the task that costs you the most time and build there, instead of setting up a whole workforce right away.
Where to go from here
Start with the tool where your data already lives, not the one being advertised the loudest, and pick a single recurring task as your first test case. Ready-made blueprints for eight specific roles, from lead scout to bookkeeping, are in my community Claude Practitioners; that's also where I show how I actually run Olaf and the other seven.
