An AI agent does more than take an instruction and return an answer. It works out the steps on its own and gets the job done by operating tools and outside services. It has drawn attention as a way to automate business work, and the number of products has grown fast. But products called "AI agents" differ widely in what they actually are, from platforms for building your own agents without code to general-purpose agents that browse the web and research things when you ask.
This article is for people in IT, business transformation, and sales planning who want to use AI agents to automate work. We sort the major AI agents into four types and compare them, laying out the pricing units, what you can try for free, and what to check before you adopt one.
This article does not cover business plans for generative AI chat such as ChatGPT or Claude. Autonomous agents that keep running around the clock (such as OpenAI's dots and OpenClaw) are compared in 6 Autonomous AI Agents Compared. Prices and limits are only what each vendor's official site listed as of October 20, 2026.
Bottom line: recommendations by goal
- You want to build internal agents inside Microsoft 365: Microsoft Copilot Studio
- You want to build your own agents in depth without code: Dify, n8n
- You want to add AI to automations that connect the SaaS tools you already use: Zapier, Make
- You want to hand off research and document creation: Manus
- You want to hand off one specific area such as sales, customer service, or development: Salesforce Agentforce (CRM), Devin (software development)
Types of AI agents
Before comparing, let's sort out the types of AI agents. This article uses the following four.
- No-code builders: products for building your own agents and chatbots through on-screen operations (Copilot Studio, Dify)
- Workflow automation: products that build AI decisions into automations connecting SaaS tools (n8n, Zapier, Make)
- General-purpose agents: products that take a goal and handle everything from planning to execution on their own (Manus)
- Task-specific agents: products that work inside a specific job or system (Agentforce, Devin, Glean)
The difference is that builders and workflow automation need someone to build things, while general-purpose and task-specific agents are often ready to use right away.
How to choose an AI agent
Who builds it, and who uses it
Even with a no-code builder, designing an agent takes the skill to break a business process into steps. If you have builders in-house, a builder platform is a candidate. If not, look at general-purpose or task-specific agents that are closer to finished products.
Whether it connects to your existing systems
If the agent can't operate your internal data and SaaS tools, people end up copying the results over by hand. Check whether integrations exist for the chat, storage, CRM, and other tools you use. Support for MCP (a common way to connect agents with outside tools) is another point to check.
How much you hand off, and where people check
The controls you need depend on whether the agent stops at drafting or goes as far as sending and registering. If you let it send email or update outside systems, check whether there is a way for a person to approve before it runs (human-in-the-loop).
Permissions and records
What data an agent can read has to match the permissions of the person using it. Whether a record (audit log) remains of who had it do what and when is also something IT reviews will always ask about.
Pricing units
AI agent pricing is often based on usage such as runs, credits, or actions, on top of the number of users. With usage-based products, if you sign up without estimating how much you'll use in full operation, costs can grow beyond what you expected.
AI agent comparison table
| Tool | Type | Free plan / trial | Paid plan (approx.) | Pricing unit | Highlights |
|---|---|---|---|---|---|
| Microsoft Copilot Studio | Builder | Trial license available (publishing not allowed) | ¥29,985/month (excl. tax) for a pack of 25,000 credits | Credits per tenant | Build agents that run in Teams and SharePoint |
| Dify | Builder | Sandbox (free) / Community edition (self-hosted, free) | Professional $590/year / Team $1,590/year | Per workspace | Build agents and RAG without code |
| n8n | Workflow automation | Free Cloud trial / Community edition (self-hosted) | Starter €20/month / Pro €50/month (billed annually) | Monthly executions | Unlimited users and workflows on every plan |
| Zapier | Workflow automation | Free (100 tasks/month / Agents: 400 activities/month) | Pro $19.99/month (billed annually, 750 tasks) | Tasks / activities | Connects to many SaaS tools. Agents are counted separately |
| Make | Workflow automation | Free (1,000 credits/month) | Could not confirm on the official site | Credits | Build scenarios on screen |
| Manus | General-purpose agent | Free (credits refilled daily) | Pro $20/month (example: 4,000 credits/month) | Credits | Handles research and document creation on its own |
| Salesforce Agentforce | Task-specific (CRM) | Salesforce Foundation (free) | Flex Credits ¥60,000 per 100,000 credits | Actions / conversations / users | Works inside your Salesforce data |
| Devin | Task-specific (development) | Free | Pro $20/month / Teams full seat $40/month (minimum $80) | Per person + credits | Hand off software development |
| Glean | Task-specific (enterprise search) | Could not confirm on the official site | Contact sales | Contact sales | Searches across internal information and works with agents |
Highlights of the main tools
Microsoft Copilot Studio
Microsoft's tool for building AI agents. The agents you build can be used in Teams, SharePoint, websites, and more. Pricing is ¥29,985 per month (excl. tax) for a pack of 25,000 Copilot Credits per tenant. There is also pay-as-you-go billing for what you use, and a prepaid plan with discounts of up to 20%. A Microsoft 365 Copilot license (equivalent to ¥4,497 per person per month, billed annually) includes some Copilot Studio features, so you can build internal agents at no extra cost. It suits companies that use Microsoft 365 company-wide.
Dify
An open-source platform for building AI agents and chatbots without code. The cloud version starts with the free Sandbox plan (200 message credits, 1 member, 5 apps). The paid Professional plan is $590 per year and Team is $1,590 per year, both priced per workspace. The Community edition, which you can run on your own servers, is free. It suits companies that want to build agents that use their own data in depth. If you self-host, designing operations and security is your own responsibility.
n8n
A workflow automation tool where an AI agent can be built in as one of the steps. Pricing is based on monthly workflow executions: Starter is €20 per month (billed annually, 2,500 executions per month) and Pro is €50 per month (billed annually, 10,000 executions per month). Every plan has unlimited users and workflows. You can try the Cloud version with a free trial that needs no credit card (up to 1,000 executions). There is also a self-hosted Community edition. It suits companies with engineers who want to build things in fine detail.
Zapier
A tool that automates work by connecting many SaaS tools. The Free plan allows up to 100 tasks per month and is limited to two-step automations. Pro starts at $19.99 per month (billed annually, 750 tasks per month), or $29.99 billed monthly. The AI agent feature (Zapier Agents) is counted separately: 400 activities per month on Free and 1,500 activities per month on Pro (from about $33.33 per month billed annually). It suits companies that want to add AI decisions to their everyday SaaS operations.
Make
An automation tool where you build the flow of steps (a scenario) on screen. Pricing is counted in credits, and most steps use 1 credit per run. Steps that use AI may consume more credits. The Free plan includes 1,000 credits per month. We could not confirm the prices of the paid plans (Core, Pro, Teams) in the text of the official pricing page. Annual billing is cheaper than monthly billing.
Manus
A general-purpose agent that takes a goal and handles everything from planning through research and document creation on its own. Pricing is credit-based, and the official help gives the example of 4,000 credits per month on the $20-per-month Pro plan. The Free plan has credits refilled each day, with a monthly consumption cap of 1,500 credits. There is also a Team plan. It suits situations where you want to hand off research or document creation entirely. Credit consumption varies with the task, so it helps to get a sense of it during a trial.
Salesforce Agentforce
An AI agent that runs inside Salesforce. The tools for building agents are available to start with the free Salesforce Foundation. Pricing is ¥60,000 per 100,000 Flex Credits (one action consumes 20 credits), or ¥240 per conversation under conversation-based billing. Per-user add-ons for sales and service are $125 (¥15,000) per person per month. It suits companies that keep their customer data in Salesforce.
Devin
An AI agent you hand software development to. There is a Free plan; Pro is $20 per month and Max is $200 per month (both per person). On Teams, a full seat is $40 per person per month, with a $80 monthly minimum and up to 200 people. Usage beyond the limit is paid with on-demand credits. It suits companies that want to hand off working through the development team's backlog or fixing code, and does not suit work outside development.
Glean
A product that searches across internal information and then uses agents to get work done based on the results. You can build agents without code. It supports access control based on each user's permissions and tracking of questions, answers, and actions. Pricing is not published; you start by requesting a demo. It suits large companies with many internal tools and documents that first want to cut the time spent searching.
How AI agents differ from RPA and chat AI
RPA repeats fixed steps accurately. It stops when screen layouts or steps change, but it is good at handling large volumes of the same work. AI agents can rethink the steps to fit the situation, but they don't always produce the same result every time. It becomes easier to think about if you assign routine, high-volume work to RPA or workflow automation and work that needs judgment to AI agents.
Chat AI stops at answering questions and writing text. Entering the results into internal systems, or creating and sharing files, is left to people. AI agents differ in that they carry on through that second half of the work. In return, you need to decide what to hand off and where people check.
How to adopt without failing
- Narrow it down to one job to hand off: start with work whose steps are clear and where mistakes can be undone
- Decide where people check: put an approval before anything that affects the outside, such as sending, registering, or paying
- Try it small and measure: record the time it took, the number of redos, and the cost of usage
- Check permissions and records before expanding: before adding more departments, confirm the permission settings and how audit logs are kept with IT
Samurai App as an option
Samurai App's AI agent reads your internal knowledge and tables, and gets work done by operating the browser and the outside services you connect. You always work with the same single agent, and the agent itself picks the skill that fits the request. There's no need to build separate agents for each task.
When information is missing, it asks a clarifying question, and for work with many steps it shows an execution plan before going ahead. Before a sensitive action it shows an approval card and does not go further until you approve. It can only work within the permissions and visible scope of the person who asked. You can choose the AI model in the input box, and you can also bring your own API key for a model you provide (BYOK).
Pricing combines one base fee per company with credits based on usage, and it doesn't grow with the number of people you invite. There is a free trial with no credit card needed.
Frequently asked questions
What is the difference between an AI agent and chat AI?
Chat AI stops at returning answers to questions or text. An AI agent works out the steps on its own and gets work done by operating outside tools and services. That is why the scope of permissions and a way for a person to approve before execution become important.
Do I need programming to use AI agents for business automation?
Some products, such as Copilot Studio, Dify, and Zapier, let you build agents and automations through on-screen operations. You still need to break down and design the business process, though. If you have no one to build, it's realistic to start by trying general-purpose or task-specific products.
Can I try AI agents for free?
Dify, n8n, Zapier, Make, Manus, Devin, and others offer free plans or free trials. However, runs and credits are capped, so using them for real work assumes a paid plan.
Is there work I shouldn't hand off to an AI agent?
Operations that can't be undone if something goes wrong (payments, contracts, mass sends to outside parties, and so on) should not be handed off without an approval step. It's safest to start with work where people can check the results, such as drafting and research.
Sources
- Microsoft Copilot Studio pricing
- Microsoft Copilot Studio licensing (Microsoft Learn)
- Dify pricing
- n8n pricing
- Zapier pricing
- Make pricing
- Manus credit consumption and how to obtain credits (help)
- Salesforce Agentforce pricing
- Devin pricing (self-serve plans)
- Glean pricing
Prices and specifications were confirmed on each vendor's official site as of October 20, 2026. Check each vendor's site for the latest information.