The short answer
The best AI for automation in November 2026 is Zapier, because it connects more than 9,000 apps, its Copilot builds automations from a plain-English description, and its AI by Zapier step can reason, call tools and ask for human approval before acting. n8n, from Berlin's n8n GmbH, is the runner-up for technical teams that want self-hosting and any AI model. Make has the best free plan: 1,000 credits a month with Make AI Agents included.
| Our pick | Zapier: 9,000+ apps, Copilot, AI by Zapier step with per-tool approvals |
|---|---|
| Runner-up | n8n: cloud or self-hosted, n8n Assistant, n8n Agents (preview), any model you have a key for |
| Best free option | Make free plan: 1,000 credits a month, routers free, Make AI Agents included |
| How you pay | Zapier counts tasks, Make counts credits, n8n counts executions (one agent turn is one execution) |
| Let ChatGPT or Claude act in your apps | Zapier MCP and Make MCP server on every plan; n8n supports MCP connections |
| Training on your data | Zapier: Enterprise opted out, others by form; Make: Maia data not used to train; n8n: Assistant data not used to train |
The short version: which AI automation tool to pick
If your team is not technical and you want to connect the apps you already use, start with Zapier. It has the most apps, the gentlest learning curve and an AI step that can make a decision in the middle of a workflow with a human approval where it matters. If you have someone technical and want full control, including hosting it yourself, choose n8n. If you want to start for free with real AI decisions in the flow, the Make free plan is the most generous.
All three added serious AI this autumn. Zapier folded its standalone Agents product into the AI by Zapier step. n8n introduced n8n Assistant on 9 September and n8n Agents on 25 September 2026, and n8n 3.0 entered beta on 27 October. Make's Waves '26 event on 20 October 2026 upgraded its Maia assistant with a Plan mode, memory and error fixing.
Our pick: Zapier
Why it wins
Zapier's strength is reach. It connects more than 9,000 apps, so whatever your sales, support or finance tools are, they are almost certainly there. You do not even have to build the automation yourself: Copilot builds Zaps, Tables, Forms, Agents, Chatbots and Custom Actions from a description, can ask for confirmation before it acts, and accepts an image or sketch of the flow you want.
The AI part is now one step. An AI by Zapier step inside any Zap can reason over the input, call tools you pick from a tools picker, return structured fields and use knowledge sources. You choose a Standard, Advanced or Premium model tier, or a named model, per step. Two safety features matter for real work: a per-tool Require approval before running toggle, so the AI must get a person's OK before, for example, sending an email; and a per-run task limit that pauses a run for review if it goes over (75 tasks by default, up to 500). Zapier also publishes agent templates such as Meeting Prep Agent, Lead Enrichment Agent and Support Email Agent, and AI Guardrails by Zapier can check for personal data inside a workflow.
Zapier MCP, included in every plan, lets Claude, ChatGPT, Cursor and other AI tools take actions in those 9,000+ apps, which makes your chat assistant much more useful.
Who it suits
Non-technical teams in sales, marketing, operations and support, and anyone who wants automations working this week rather than this quarter. Team adds up to 25 seats, SAML SSO and shared connections.
The catch
Everything costs tasks. AI steps bill in tasks, with higher multipliers for stronger models, and each MCP tool call uses two tasks, so heavy AI use can push you up a plan quickly. The Free plan allows 100 tasks a month and two-step Zaps only, with no Filters or Paths, which means an AI decision between a trigger and an action needs a paid plan. On training, Enterprise is opted out automatically; other plans opt out through a form linked from Zapier's legal page. See the Zapier guide, Zapier vs Make and Zapier alternatives.
Runner-up: n8n
Why it is excellent
n8n is a workflow automation platform you can use in the cloud or host yourself, with a visual editor plus code wherever you need it. n8n Assistant plans, builds, runs, diagnoses and fixes workflows from a plain-language request, asks for confirmation before it accesses credentials or activates anything, and produces a normal editable workflow. n8n Agents, created in the Agents tab, can be reached in Slack, Telegram, Linear and Discord, run on a schedule or be called from a workflow, and sensitive tools wait for approval. You can use any model you have credentials for.
The billing model is a big advantage at volume: n8n counts executions, and one agent turn is one execution, however many tools it calls.
Who it suits
Technical teams, companies that must keep workflow data on their own servers, and anyone running complex or high-volume AI workflows. Business (self-hosted) adds SSO and LDAP, and companies under 20 employees can check eligibility for a Startup plan with 50% off Business.
The catch
Power comes with upkeep. Self-hosting means you own updates, security and backups, and the move to n8n 3.0 is one-way with breaking changes, with a stable release expected in early December 2026. There is no permanent free cloud plan, only a trial (1,000 executions, no card needed) and the free self-hosted Community Edition. n8n Agents are still in preview. See the n8n guide, Zapier vs n8n and Make vs n8n.
Best free option: Make
Why it is the best free plan
Make's free plan gives you 1,000 credits a month. Routers and error handlers are free on every plan, and Make AI Agents, which reason and call tools inside a scenario, are included with Make's AI Provider on every plan. That combination means a flow with a trigger, an AI decision and an action can fit inside the free plan, which Zapier Free cannot do. You build on a visual canvas, and the agent's Reasoning panel shows exactly why it chose a tool, which is excellent for checking its work. Maia, Make's assistant that builds and fixes scenarios with you, is a 30-day trial on the free plan.
Who it suits
People who want to learn automation without paying, small teams with branching workflows, and anyone who likes seeing the whole flow on one screen.
The catch
Credits are easy to burn: most module actions cost 1 credit, built-in AI can cost more, and agents and Maia are still in beta, so pricing and behaviour may change. Make has a steeper learning curve than Zapier and fewer native apps, and Maia requests are processed in the US even if your organisation uses an EU data zone. Read the Make guide.
How to choose: three questions
- Who will build and maintain it? A non-technical team should start with Zapier. A team with a developer can get more control and lower costs at volume from n8n. Make sits between them.
- How many runs a month, and how much AI per run? Count your expected runs. Zapier charges tasks per action and more for stronger AI tiers; Make charges credits per module; n8n charges per execution, with an agent turn as one execution.
- Where must the data stay? If workflow data must stay on your own infrastructure, self-host n8n. If you need SSO and audit logs, compare Zapier Team or Enterprise, Make Enterprise and n8n Business or Enterprise.
A workflow: an AI automation from trigger to result
Example: every new enquiry from your website form gets sorted, enriched and answered with a draft, and the right person approves the reply. The steps use Zapier's names; the same design works in Make or n8n.
- Write the rule in one sentence first. "When a form is submitted, classify it as sales, support or other, look the company up in our CRM, draft a reply and send it after a human approves."
- Ask Copilot to build it. Paste the sentence into Copilot and ask it to confirm the plan before it builds.
- Give the AI step clear rules. In the AI by Zapier step, describe the categories with examples, list the fields it must return (category, urgency, summary) and pick the tools it may use.
- Put a person in the loop. Switch on Require approval before running for the send-email tool, so nothing reaches a customer unchecked.
- Test with real but harmless data. Run five past enquiries through it and compare the AI's categories with what you would have chosen.
- Set limits and watch costs. Keep the per-run task limit low at first and check task usage after the first week.
- Review monthly. Look at what the AI got wrong and add those cases as examples to its instructions.
Prompts to copy for AI automation
Use the first prompt and the last two in Zapier Copilot, Make's Maia, n8n Assistant or any chat assistant to plan, build and fix workflows. Use the four in the middle as instructions inside an AI step or agent. Replace the [placeholders].
Prompts to copy
Build a workflow from one sentence
Build an automation: when [trigger, for example a new row is added in Google Sheets], do [steps in order]. Use [app names]. Before you build anything, show me the plan step by step, list the connections I need, and ask me about anything unclear.
You get: A step-by-step plan and list of connections to confirm before the builder creates the workflow.
Sort incoming email or enquiries
You sort incoming messages for [team]. Classify each message as exactly one of: [category 1], [category 2], [category 3], other. Return JSON with category, urgency (low, medium, high), a one-sentence summary and the customer's company name if present. Examples: [paste two or three short examples with the right category]. If you are not sure, choose other.
You get: Consistent, structured fields the next step of the workflow can route on.
Draft a reply that waits for approval
Draft a reply to this customer message: [message]. Use our tone: [tone]. Answer only from this information: [policy or FAQ text]. If the answer is not in the information, say a colleague will follow up within [time]. Keep it under 120 words. Do not send it; return the draft for approval.
You get: A grounded draft reply that a person approves before it is sent.
Weekly report from your tools
You receive this week's and last week's [metrics] from [apps]. Write a short report for [audience] with: the three numbers that changed most compared with last week, one likely reason for each based only on the data, and one suggested action. Keep it under 150 words and return only the report text.
You get: A plain-English summary of what changed and what to do, ready for the next step to post on a schedule to Slack or email.
Enrich a new lead
For this new lead [name, email, company], find the company's website, industry, approximate size and country using only the tools you have been given. Return these fields plus a confidence score from 1 to 5 for each. If you cannot find a field, leave it empty rather than guessing.
You get: Enriched lead fields with confidence scores and no invented data.
Diagnose a failed run
This workflow run failed: [paste the error or describe the step]. Explain in plain English what went wrong, the most likely cause, and the smallest change that fixes it. Tell me how to test the fix without sending anything to customers.
You get: A plain explanation of the failure and a safe way to test the fix.
Find what to automate first
Here is a list of tasks my team does every week, with roughly how long each takes: [list]. For each, say whether it is a good fit for automation, which parts need a human decision, and the apps involved. Rank the top three by hours saved and how easy they are to set up.
You get: A ranked shortlist of automations worth building first.