How to Choose an AI Automation Tool: Checklist

Short answer

Judge an AI automation tool on six things: what it bills for, where it runs and who holds your data, whether it connects to the apps you use, how it handles failures, whether a person can approve risky steps, and how hard it is to leave. Compare the billing unit first, because vendors count different things.

This page ranks no tools. Nobody tested any tool named here. Every product statement comes from a vendor page read on 2026-10-03, with the page named.

What the sources say, and what to consider

Each criterion below has two parts. "What the sources say" lists documented facts with the source and its date. "What to consider" is our assessment.

1. Billing unit: what exactly is counted?

What the sources say

Vendors bill in different units. These are the vendors' own words:

Tool	Billing wording	Source
Zapier	"A task is any successful action that runs in Zapier." Zapier Agents use "a separate activity quota."	Zapier Help, updated August 21, 2026
Make	"Credits replaced operations as the term for Make's billing unit."	Make Help Center, Credits
n8n	"Pricing based on monthly workflow executions, regardless of complexity."	n8n pricing
Activepieces	Plans list credits per month, for example "1,000 credits a month" on Free, with a credit reset period per plan.	Activepieces pricing
Lindy	"Credits measure the work Lindy does."	Lindy pricing
Gumloop	Plans list "Included Credits" and an "Orchestration Fee."	Gumloop pricing

For Relevance AI, the pricing page we read lists "Custom Actions" and "Custom Vendor Credits" under Enterprise. We did not find a definition of either term on that page, so we state no billing unit for it.

Details in the Zapier article show how unit definitions work. Triggers, filters, paths and failed actions do not count as tasks. Steps in an error handler path do count. Steps that rerun during a full replay count again. Each successful Zapier MCP tool call counts as 2 tasks. On Make's credits page, a non-AI app operation equals 1 credit. With a custom AI provider connection you pay your provider directly for tokens. With Make's own AI provider, credits are based on tokens and operations.

What to consider

• Write down your workflow as steps and runs per month, then count using each vendor's definition. A fixed number of steps per run makes this easy. An agent, which picks its own steps, makes it harder to forecast.
• Ask what happens at the limit: does the tool stop, or charge more? Pages we read for this article did not answer that for every vendor, so read the plan's terms.
• Ask whether AI model usage is bundled or billed separately. Make's credits page says that with a custom AI provider connection you pay your provider directly for tokens.
• Do not compare prices across tools until the units match.

2. Hosting and data control: where does it run, and who sees the data?

What the sources say

• n8n's documentation says you can self-host on your own infrastructure, on-premises or in a private cloud. Without a license key, a self-hosted install runs as the free Community edition. (n8n docs: Hosting, read 2026-10-03) The n8n GitHub page calls the project "fair-code" distributed under the Sustainable Use License and the n8n Enterprise License. (n8n on GitHub) The same page lists "Source Available: Always visible source code" among its points. Read the license text if your use is commercial.
• Activepieces' pricing page shows "Cloud or on-prem" in the hosting row of every plan, and lists "Audit logs and secrets" under its Ultimate plan and "SSO and standard roles" under Team. (Activepieces pricing) Its documentation says you can run it on your own infrastructure and decide "which models see your data." (Activepieces docs)
• Lindy's pricing page lists "HIPAA compliance & signed BAA" among Enterprise features. (Lindy pricing)
• We did not fetch a self-hosting statement for Zapier, Make, Gumloop or Relevance AI. Absence in this article is not evidence that none exists.

What to consider

• Decide what data the workflow touches: customer names, invoices, health or HR data. Then find the vendor's privacy policy, terms, data processing terms and trust page. Look for who stores the data, how long, and whether it is sent to an AI model provider.
• Check which plan includes the security features you need. Several pages above put audit logs, SSO or compliance terms on higher tiers.
• Self-hosting moves responsibility to you: updates, backups and security. It does not remove the need to read the license.
• This article gives no compliance verdicts. Whether a tool meets your legal obligations is for you and your advisers to decide from the vendor's documents.

3. Integrations: does it connect to your apps and do what you need?

What the sources say

Vendors publish integration counts, and the counts vary by page and date. On 2026-10-03, the n8n GitHub repository description said "400+ integrations," while the README text on the same page said "1500+ integrations." The Activepieces GitHub README said "280+" pieces for MCP use and "over 200" services, while its pricing page referred to "760+ integrations" for the Embed offering. (n8n on GitHub, Activepieces on GitHub, Activepieces pricing)

What to consider

• Do not rely on a count. Search the vendor's app directory for each app you need.
• Then check the specific trigger and action, not only the app. A listed app may support reading data but not the write action you need.
• Check for a generic HTTP or API option as a fallback. It needs more setup.
• If you rely on a niche app, test a trial before you commit to a plan.

4. Error handling: what happens when a step fails?

What the sources say

• n8n: you can set an error workflow that runs when an execution fails, for example to send an email or Slack alert. (n8n docs)
• Make: its error handlers page (updated 08 Sep 2026) lists Skip, Retry, Resume, Commit and Rollback handlers. Retry stores incomplete executions and enables automatic or manual retries. (Make Help Center)
• Zapier: failed actions are not counted as tasks, but steps in an error handler path and steps that rerun in a full replay are. (Zapier Help)

What to consider

• Ask four questions. Does the tool alert you? Can you retry from the failed step? Does a retry cost more? Can the failure leave a half-finished change, such as an invoice created but not sent?
• For AI steps, add a check on the output, such as a required format, before the next step uses it.
• OpenAI's guide names exceeding failure thresholds as a trigger for human intervention, and suggests limits on retries. (OpenAI guide, PDF) Set a retry limit.

5. Human approval: can a person stop a risky step?

What the sources say

• n8n's documentation says you can require human approval before an AI Agent executes a specific tool. The workflow pauses, and the person approves or denies. (n8n docs)
• Activepieces' GitHub README says its pieces include one to delay execution or require approval, plus human input triggers such as a chat interface and a form interface. (Activepieces on GitHub)
• OpenAI's guide says sensitive, irreversible or high-stakes actions should trigger human oversight until confidence in reliability grows. (OpenAI guide, PDF)
• We did not fetch an approval-step page for the other tools in this directory.

What to consider

• List the actions that cannot be undone: sending, paying, deleting, publishing. Require approval for each.
• Check that approval works where your team already works, such as email or chat, so reviewers actually see it.
• For agents, check whether approval attaches to a tool, to a step or to the whole run. See AI agents vs AI automation for why this matters.

6. Exit cost: how hard is it to leave?

What the sources say

• n8n's documentation says workflows are saved as JSON, and you can export them as JSON files and import them elsewhere in n8n. (n8n docs: Export and import)
• Activepieces credit allowances reset daily, monthly or yearly depending on the plan, per its pricing page. (Activepieces pricing)
• We did not fetch export documentation for the other tools.

What to consider

• An export file lets you keep a record of what you built. It does not mean the workflow runs in another tool. Each platform has its own step types and app connections, so expect to rebuild.
• Ask where prompts, connections, logs and history live, and whether you can download them.
• Check the cancellation terms: notice period, annual commitments, and whether unused credits expire.
• Keep a written spec of each workflow outside the tool: trigger, steps, rules and the prompt text. That is your real exit plan.

The checklist

Copy this into your own notes and fill it in for each tool.

[ ] Billing unit named in the vendor's own words: ________
[ ] Monthly runs x steps counted with that definition: ________
[ ] Cost of failed runs, retries and replays checked
[ ] AI model usage: bundled or billed separately: ________
[ ] Hosting options and where data is stored: ________
[ ] Privacy policy, terms and data processing terms read
[ ] Security features needed (SSO, audit logs) and the plan that has them: ________
[ ] Each required app found in the directory, with the exact trigger/action
[ ] Failure alerts, retry limit and partial-failure behavior understood
[ ] Approval step available for irreversible actions
[ ] Export format and what it contains: ________
[ ] Cancellation terms and credit expiry: ________
[ ] Workflow spec kept outside the tool

Limitations

• We did not test any tool. The checklist is our assessment, built from the sources named.
• Evidence is uneven. Some tools have documented answers on a criterion in this article and some do not, because we only used pages we fetched.
• Vendor pages change quickly, including plan names and units. Re-check on the day you decide.
• The checklist covers selection, not security review. It gives no compliance conclusion.

When this is not the right choice

• A formal checklist is overkill if you want to try one low-risk workflow. Start with a free trial or free plan and read the billing unit, then revisit.
• It is not enough if you handle regulated data. Take the vendor's documents to whoever owns compliance in your organization.
• If you want a shortlist, use the automation and agents category and then the profiles, such as Zapier, Make and n8n. If you are new to the topic, read AI automation for beginners first.