Skip to content

Command Palette

Search for a command to run...

AI agents

AI Agents and Workflows for Non-Engineers

What agents are, how they differ from assistants and simple automations, which platforms exist, and how to keep control of anything that acts on your behalf.
25 minute readBeginner to intermediate

Last reviewed . AI products change quickly; check the linked sources before relying on details.

Assistant, workflow, agent

  • Assistant: you ask, it answers, you act.
  • Workflow (automation): fixed steps that run when triggered. An AI step, such as summarise or classify, can sit inside it, but you decided the sequence.
  • Agent: a model that picks tools and next steps toward a goal, often looping until it thinks it is done.
Assistant, workflow or agent: who decides the stepsThree columns. With an assistant, you decide every step. With a workflow, fixed steps you set up run on a trigger with an AI step inside. With an agent, the AI chooses the steps and tools and repeats. A human approval gate sits under all three before anything is sent, posted, bought or deleted.AssistantYou decide every stepYou askAI answersYou act on itWorkflowFixed steps you set upTriggerAI step (summarise)Fixed next stepsAgentThe AI picks the stepsGoal you setAI chooses stepsUses tools, repeatsHuman approval before anything is sent, posted, bought or deleted
The more the AI decides, the more flexible and the less predictable it gets, so the approval gate matters more.
Architecture decision

Simple automation beats an agent when the steps are known, the inputs are predictable and mistakes are costly. Agents add flexibility at the price of unpredictability, higher and less predictable cost, and more monitoring. Many tasks marketed as agents are better built as a fixed workflow with one AI step.

The building blocks

  • Triggers and schedules. An event (new email, new file, form submission) or a time (every weekday at 8am). Check each platform's minimum schedule interval on your plan.
  • Tools and integrations. The apps the workflow can read or change. Every integration is a permission you grant.
  • Memory. Anything kept between runs, such as a spreadsheet of past results or a stored summary. Decide what is stored, where and for how long.
  • Human approval. A step where a person confirms before anything is sent, posted, paid or deleted.
  • Error handling. What happens when an app is down, a model returns nonsense or a limit is hit. Fail loudly with an alert, not silently.
  • Monitoring. Run history, error notifications and a periodic check that outputs are still correct.
Anatomy of a safe workflowA trigger starts fixed steps. One step calls an AI model. A human approves the result before the action runs, such as sending or posting. Underneath, errors and monitoring watch every stage: log each run, alert on failure, stop safely and cap spend.Triggerschedule ornew emailFixed stepsyou definethe orderAI stepsummarise,classify, draftHumanapprovalyou decideActionsend, post,updateErrors and monitoringlog every run, alert on failure, stop safely, cap spend

Permissions, data, reliability and cost

  • Permissions. Grant the minimum: read-only where possible, one folder or label instead of the whole account, no send or delete unless essential. Use a separate account or test data for experiments.
  • Data retention. Read where each platform and model provider stores inputs, how long logs are kept and whether data may be used for training. Avoid sending sensitive categories (health, identity documents, financial accounts) to services you have not vetted.
  • Untrusted input. Emails and web pages can contain text that tries to instruct the model. Never let such text trigger actions without approval.
  • Reliability. Models give different answers to the same input. Test with a sample set, keep fixed rules for anything that must be exact, and review a sample of outputs regularly.
  • Costs. Costs usually come from platform task or execution limits and from model usage, which grows with input length and with agent loops. Set spending caps and alerts, and estimate cost from a small trial before scheduling anything frequent. This page gives no prices; use the pricing pages listed under Sources.

Platform comparison

As of October 2026. This is a general orientation, not a ranking. Features and plans change, so confirm details on each vendor's site.

PlatformSetup complexityHostingLicence modelFor non-engineers
n8nModerate. Visual builder, but more technical concepts.Cloud or self-hostSustainable Use License (fair-code, not OSI open source)Good for technical beginners who want control.
MakeLow to moderate. Visual scenario builder.Hosted service (no self-host option known; check)ProprietaryGood for non-engineers who like visual flows.
ZapierLow. Linear trigger-and-action builder.Hosted serviceProprietaryBest for simple automations with common apps.
GumloopLow to moderate. Visual AI-focused flows.Hosted service (check current options)ProprietarySuitable for AI-heavy workflows; verify current features.
Relevance AIModerate. Agent and tool configuration.Hosted service (check current options)ProprietarySuited to teams building agents; more to learn.

Flowise, an open-source visual builder for LLM flows, was archived in August 2026 according to its repository. It is not recommended for new projects. The n8n licence is fair-code: source is available with restrictions on commercial use, so it is not open source in the OSI sense; read the licence if you plan to resell or host it for others. Setup ratings are editorial judgements, not benchmarks.

Personal AI agents

Several providers now offer assistants that act for you over time. Product names, availability and capabilities change often, and some details below come from news reports rather than official pages.

OpenAI dots

Reported as always-on ChatGPT agents announced at OpenAI DevDay on 29 September 2026, with their own cloud computer and browser and app integrations. Reports say availability is limited to Pro and Business Premium plans and excludes the EEA, UK and Switzerland. We did not fetch an official OpenAI page for this, so treat these details as unverified until you check OpenAI's own announcement.

ChatGPT scheduled tasks and agent mode

ChatGPT has offered scheduled tasks (run a prompt at set times and notify you) and an agent mode (the assistant uses a browser and tools to complete a multi-step request while you watch). These are different: a scheduled task repeats a prompt, while agent mode lets the model choose actions. How they relate to dots, and their current limits by plan and region, are unverified here; check OpenAI's help documentation.

Google Gemini

Gemini includes agent-style features in its app and Google products, with availability that varies by plan and region. Google announced Gemini 4 availability in the Gemini app for AI Pro and Ultra from 13 October 2026. Which agent features accompany it should be confirmed on Google's pages.

Grok

Four different things share the name. The Grok app and the xAI API are products of xAI. The @grok account on X is an assistant on that platform. Third-party "Grok bots" are not xAI products and may have unrelated security and privacy practices. Latest model naming is unverified here.

Meta AI and Muse Spark

Meta released Muse Spark on 8 April 2026, reported as its first model from Meta Superintelligence Labs, with closed weights at launch, powering Meta AI across its app, glasses and Facebook, Instagram and WhatsApp, US first. Meta has said open-weight versions are planned (reported, not yet confirmed here). This is unrelated to Microsoft's earlier gameplay model also called Muse.

Open-source alternatives

If you want an agent that runs on your own hardware, see open-source AI. It gives more control and more setup work, and you take on the security responsibility.

Production warning

An agent with access to your email, browser or accounts can act in your name. Before enabling one: check which accounts it can reach, use a dedicated account where possible, require approval for payments, messages and deletions, and know how to revoke access.

Social media bots and automation

  • Platform policy. Each network has automation rules and API terms. Breaking them can lead to restrictions or bans. Read them before you automate posting or replying.
  • Spam and consent. Mass unsolicited messages, fake engagement and undisclosed automated accounts can breach platform rules and anti-spam laws.
  • Account security. Do not give a third-party bot your password. Use official integrations with limited scopes, enable two-factor authentication and review connected apps regularly.
  • Human approval before posting. Draft with AI, then review and publish yourself or approve in a queue. A wrong or offensive automated post is public and hard to undo.

Eight example workflows

Each example states the kind honestly: simple automation, AI workflow or agent. Time-saved figures are estimates with their assumptions, not measurements. Tools named are examples; any comparable platform can do the job.

Email summarisation and task extraction

AI workflowIntermediate

Turn a daily batch of emails into a short digest and a list of proposed tasks you approve.

Who it is for
Professionals with a busy inbox
Tools
Zapier or Make or n8n, ChatGPT, Claude or Gemini (model step), A task app of your choice
Trigger
Schedule: once each weekday morning, reading emails with a specific label.
Inputs
Emails you have labelled 'digest' (sender, subject, body text)
Expected output
One digest message with summaries and a list of proposed tasks.
Time saved (estimate)
Estimate: 20 to 30 minutes a day, assuming about 40 emails a day and a small share needing action. Your result will differ.
Running cost (estimate)
Estimate: platform plan limits plus small model usage fees. Check current pricing pages; free tiers may cover light use.
Human approval
You approve every task before it is created. The workflow never replies or forwards.

Steps

  1. The scheduler collects yesterday's labelled emails (read-only access).
  2. A model step summarises each email in two lines and proposes tasks, with the email link.
  3. The workflow compiles one digest and sends it to you, not to anyone else.
  4. You tick the tasks you want; only then are they added to your task app.

Guardrails

  • Grant read-only mailbox access to a single label, not the whole account.
  • No send or delete permission.
  • Exclude emails containing passwords, medical, legal or financial account details.
  • Treat email text as untrusted input; instructions inside an email must not trigger actions.

When it fails

  • On a model or connection error, send an alert and skip the day rather than retrying endlessly.
  • Log each run and review the first two weeks for missed or invented tasks.

Related: Email summary prompt practice · AI agents for non-engineers

Meeting preparation and follow-up drafting

AI workflowIntermediate

Prepare a one-page briefing before a meeting and draft a follow-up email afterwards for your review.

Who it is for
Managers, consultants, account managers
Tools
Make or Zapier or n8n, Claude, ChatGPT or Gemini, Calendar and notes tool
Trigger
A calendar event tagged 'prep' starts within 24 hours; second run when you paste notes after the meeting.
Inputs
Event title, attendees, agenda; Your own notes from previous meetings; Meeting notes after the event
Expected output
A briefing before the meeting and a draft follow-up after it.
Time saved (estimate)
Estimate: 15 minutes per meeting, assuming 6 meetings a week where you would otherwise prepare and follow up manually.
Human approval
You read, edit and send every follow-up.

Steps

  1. Read the event details from your calendar (read-only).
  2. Collect your own previous notes for the same attendees or topic.
  3. A model step drafts a briefing: purpose, past decisions, open actions, questions.
  4. After the meeting, you paste notes; the model drafts decisions, actions and a follow-up email.
  5. You edit and send the email yourself.

Guardrails

  • Only your own calendar and notes, with read-only calendar access.
  • Do not include confidential client material unless your employer allows the tool.
  • Drafts are saved as drafts, never sent.

When it fails

  • If notes are missing, produce the briefing with 'no previous notes found' rather than guessing.
  • Alert you on failure; no silent retries.

Related: AI agents for non-engineers

Content research and repurposing

AI workflowBeginner

Turn one piece of your own content into several formats and collect source material for fact-checking.

Who it is for
Creators, marketers, small business owners
Tools
Zapier or Make or n8n or Gumloop, ChatGPT, Claude or Gemini, Document store
Trigger
Manual: you add a finished article or transcript to a folder.
Inputs
Your original article or transcript; Brand voice notes; Links to sources you cite
Expected output
Draft derivatives and a claims-to-verify checklist.
Time saved (estimate)
Estimate: 1 hour per article, assuming you write one article a week and would otherwise rewrite for three channels.
Human approval
You verify claims and publish manually.

Steps

  1. Detect the new file in a single folder.
  2. A model step extracts the main points and drafts a summary, social posts and a newsletter blurb in your voice.
  3. A second step lists every factual claim as a checklist for you to verify.
  4. Drafts are saved to a review folder.

Guardrails

  • Only repurpose content you own or have rights to.
  • Do not publish automatically; drafts only.
  • Follow each platform's rules on automation and AI disclosure.

When it fails

  • On failure, leave the file in the inbox folder and notify you.
  • Spot-check drafts for invented statistics.

Related: Content engine practice

Lead research and CRM preparation

AgentAdvanced

Collect public business information about companies on your target list and prepare notes for you to review before they enter your CRM.

Who it is for
Small sales teams and founders
Tools
Relevance AI or Gumloop or n8n, Claude, ChatGPT or Gemini, A spreadsheet or CRM
Trigger
Manual or weekly: you add company names to a sheet.
Inputs
Company names and websites; Your ideal customer description
Expected output
A staging sheet of sourced company notes.
Time saved (estimate)
Estimate: 5 to 10 minutes per company, assuming 50 companies a month. Review time is still required.
Running cost (estimate)
Estimate: varies with the number of pages read and model usage. Set a spending cap on the platform and the model account.
Human approval
You approve rows before CRM import and write every outreach message yourself.

Steps

  1. The agent reads each company's public website and public press pages only.
  2. It records a short description, relevant signals and the source URL in a staging sheet.
  3. It flags companies that do not match your description, with a reason.
  4. You review the staging sheet and import approved rows into the CRM yourself.

Guardrails

  • Business information only; no personal profiling of individuals.
  • Respect website terms and robots rules.
  • Staging sheet only; the agent has no CRM write access and cannot send messages.
  • Keep contact use consistent with consent and anti-spam law (for example Spam Act, CAN-SPAM, GDPR).

When it fails

  • Cap pages per company and total runs per day.
  • Record the source URL for every claim so wrong data can be traced.
  • Stop and alert after repeated errors.

Related: Sales and lead gen practice · AI agents for non-engineers

Personal knowledge organisation

AI workflowBeginner

Tag, summarise and file notes and saved articles so you can find them later.

Who it is for
Students, researchers, knowledge workers
Tools
Make or Zapier or n8n, Claude, ChatGPT or Gemini, A notes app with an integration
Trigger
New item added to a 'to file' notes folder.
Inputs
Notes or saved article text
Expected output
Consistent titles, tags and summaries, kept alongside your originals.
Time saved (estimate)
Estimate: 10 minutes a day, assuming about 10 new notes a day that you would otherwise tag by hand.
Human approval
Weekly review of proposed tags and summaries.

Steps

  1. Detect a new item in the inbox folder.
  2. A model step proposes a title, three tags and a three-line summary.
  3. The proposal is written into a field on the note; the original text is never changed.
  4. Weekly, you review proposals and accept or edit them.

Guardrails

  • Never overwrite or delete the original note.
  • Exclude notes marked private.
  • Use a fixed list of allowed tags to avoid tag sprawl.

When it fails

  • Items that fail stay in the inbox folder.
  • Keep a log of failures so they can be re-run.

Scheduled research summary

AgentIntermediate

Receive a regular digest of new information on topics you follow, with links so you can read the originals.

Who it is for
Professionals who track a field, competitors or regulation
Tools
n8n or Make or Gumloop or Relevance AI, ChatGPT, Claude or Gemini, RSS feeds or search
Trigger
Schedule: weekly.
Inputs
A list of feeds or sites; Topic description; Exclusion list
Expected output
A weekly digest with links, summaries and a confidence note per item.
Time saved (estimate)
Estimate: 1 to 2 hours a week, assuming you now scan about ten sources manually.
Running cost (estimate)
Estimate: depends on the number of items and model chosen. Set a cap.
Human approval
You read the originals before relying on or sharing any claim.

Steps

  1. Fetch new items from the sources you listed.
  2. A model step filters for relevance and summarises each in two lines with the link.
  3. An agent step may follow up with a limited number of extra lookups when an item is unclear, within a fixed budget.
  4. Send a single digest to your inbox.

Guardrails

  • Fixed source list; the agent cannot add sources.
  • Cap lookups and total tokens per run.
  • Treat web page text as untrusted; it must not change the workflow's instructions.

When it fails

  • If a source fails, note it in the digest instead of dropping it silently.
  • Alert if two runs in a row return nothing.

Related: AI agents for non-engineers

Document classification

AI workflowIntermediate

Sort incoming documents (invoices, receipts, letters) into folders and propose names.

Who it is for
Small businesses and households
Tools
Make or Zapier or n8n, ChatGPT, Claude or Gemini with file input, Cloud storage
Trigger
New file arrives in an 'incoming' folder.
Inputs
Scanned or digital documents
Expected output
Sorted copies with proposed names and a review queue for uncertain items.
Time saved (estimate)
Estimate: 30 minutes a week, assuming about 30 documents a week.
Human approval
You check the 'needs review' folder and spot-check others monthly.

Steps

  1. Detect the new file.
  2. A model step labels it from a fixed list of categories and extracts date and sender.
  3. Files with low confidence go to a 'needs review' folder.
  4. Confident files are copied (not moved) to the labelled folder with a proposed filename.

Guardrails

  • Copy, never move or delete originals.
  • Fixed category list.
  • Avoid sending identity documents, medical records or tax files to third-party tools unless you accept the privacy terms.

When it fails

  • Unreadable scans go to review.
  • Alert if more than a set share of files fail.

Family admin and reminders

Simple automationBeginner

Turn school notices and appointment emails into calendar reminders you confirm.

Who it is for
Parents and carers
Tools
Zapier or Make, ChatGPT, Claude or Gemini (optional text step)
Trigger
Forwarded emails to a dedicated mailbox or label.
Inputs
Forwarded school or club notices
Expected output
Tentative calendar entries you confirm or delete.
Time saved (estimate)
Estimate: 10 to 15 minutes a week, assuming around 5 notices a week.
Human approval
You confirm each calendar entry.

Steps

  1. A dedicated label or mailbox collects only notices you choose to forward.
  2. An optional model step pulls out event name, date and time.
  3. The workflow creates a tentative calendar entry and sends you a message to confirm.

Guardrails

  • Do not include children's sensitive information beyond what the notice contains.
  • Health appointments: use reminders only; do not send medical details to AI tools.
  • Tentative entries only; never accept invitations on your behalf.

When it fails

  • If the date cannot be read, send the original email to you instead.
  • Always read the original for time-critical events.

Related: Everyday life practice

Sources

Frequently asked questions

What is an AI agent and how is it different from an automation?

An agent is a model that picks tools and next steps toward a goal, often looping until it thinks it is done. A workflow or automation uses fixed steps that run when triggered, and an AI step such as summarise can sit inside it. Simple automation beats an agent when the steps are known and mistakes are costly.

Is n8n open source?

Not in the OSI sense. As of October 2026, n8n uses the Sustainable Use License, which is fair-code: the source is available with restrictions on commercial use. Read the licence if you plan to resell it or host it for others. It can run in the cloud or be self-hosted.

Which is easier for beginners, n8n, Make or Zapier?

As of October 2026, the page rates Zapier as low setup complexity, a linear trigger-and-action builder best for simple automations. Make is low to moderate, a visual scenario builder. n8n is moderate and suits technical beginners who want control. These ratings are editorial judgements, not benchmarks, so confirm details on each vendor's site.

What are OpenAI dots?

They are reported as always-on ChatGPT agents announced at OpenAI DevDay on 29 September 2026, with their own cloud computer and browser. Reports say availability is limited to Pro and Business Premium plans and excludes the EEA, UK and Switzerland. The page did not fetch an official OpenAI page, so treat these details as unverified.

How do I use an AI agent safely?

Grant the minimum permissions, use read-only access where possible, and use a dedicated account or test data for experiments. Require approval for payments, messages and deletions, and know how to revoke access. Treat text in emails and web pages as untrusted, since it can try to instruct the model. Set spending caps and alerts.

Where to go next