AI Pulse Check
← Back to blog

AI Workflows: What They Are, 10 Examples and How to Build One (2026)

By Ethan Halfhide · Updated October 19, 2026

An AI workflow is a set of automated steps, started by a trigger such as a new email or form entry, where one or more steps use an AI model to read, sort, summarize or draft, and the rest follow fixed rules. For a small business, the usual first one is a single AI step added to an automation you already run.

What is an AI workflow?

IBM defines an AI workflow as "the process of using AI-powered technologies and products to automate tasks and streamline activities within an organization", where AI systems "perform, coordinate or enhance processes", either on their own or alongside people. Its simplest example: "a language model classifying incoming support ticket[s]."

Anthropic, which makes Claude, draws a sharper line. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths", while agents "dynamically direct their own processes and tool usage." In plain terms: in a workflow, you decide the order of the steps and the AI does the reading and writing inside them. In an agent, the AI decides what to do next.

Zapier's pricing page describes the same idea as a spectrum, from a "deterministic workflow" where "every step fires in a fixed order", to an AI step in a workflow where "the path is fixed, but one step asks an AI model to draft, summarize, or extract data", to an "agentic workflow" where "AI decides what happens next."

Fixed automationAI workflowAI agent
Who sets the stepsYou, every oneYou set the path; AI handles steps that need reading or writingThe AI chooses the next step toward a goal
Handles messy input (free text emails, PDFs, call notes)Poorly: needs exact fields and rulesYes, inside the AI stepYes
Same input, same result?YesMostly: the path is fixed, but AI output can varyLess so
Typical small business useCopy a form entry into a spreadsheetSort incoming emails and draft replies for reviewWork a support queue, deciding when to look things up or hand off

Most small businesses get the most value from the middle column. Anthropic recommends "finding the simplest solution possible, and only increasing complexity when needed", and says workflows "offer predictability and consistency for well-defined tasks." If you want the right-hand column, our guide to AI agents for business compares agent products and their prices.

The parts of an AI workflow

Every AI workflow, whatever tool you build it in, is made of the same pieces.

  1. Trigger. The event that starts a run. Zapier defines a trigger as "an event that starts a Zap", for example a new lead arriving in your form tool. Triggers can also be a schedule or a button you press.
  2. Context. The data the AI needs: the email text, the customer record, your price list or policy document.
  3. AI step. One call to a model with a clear instruction, such as "classify this message as sales, support or billing." Power Automate calls this action "Run a prompt"; Zapier calls it AI by Zapier.
  4. Rules. Ordinary logic around the AI step: filters, branches, formatting. Zapier's Paths, for example, "use conditional, if/then logic."
  5. Human check. An approval before anything leaves the building. Microsoft documents adding a "Start and wait for an approval of text" step after a prompt so a person can review the AI's text.
  6. Action. The result goes somewhere: a CRM update, a drafted reply, a chat message, a row in a spreadsheet.
  7. Log. A record of each run so you can see what the AI produced. Zapier keeps a Zap history and n8n keeps execution logs.

Five common AI workflow patterns

Anthropic's guide to building agents names five workflow patterns. In small business terms:

PatternWhat it doesSmall business example
ChainingEach AI step works on the output of the one beforeDraft a product description, then shorten it into a social post
RoutingAI sorts the input, then rules send it down the right pathSort inbound email into sales, support and billing (Anthropic's own example routes "general questions, refund requests, technical support")
Parallel checksSeveral AI steps look at the same input at onceCheck a draft reply for tone, policy accuracy and missing details
Review loopOne AI step writes, a second grades it, the draft is revisedRewrite a quote email until it covers every required term
OrchestratorA lead AI step splits a job into parts it decides on the flyResearch a list of suppliers when the number of searches is not known in advance

The first two cover most small business needs. Anthropic notes that chaining trades speed for accuracy "by making each LLM call an easier task": a good rule for prompts is one job per AI step.

Ten AI workflow examples for a small business

Each row is a pattern you can build in any of the tools below. The human check column is our suggestion for a first version.

WorkflowTriggerAI stepHuman checkAction
Inbox triageNew emailLabel as sales, support, billing or other; flag urgencyNone for labelsApply label, alert the owner on urgent ones
Lead follow-upNew web form entrySummarize the request and draft a first replyApprove the replyAdd to CRM, send reply, create a follow-up task
Support ticket routingNew ticketClassify category and prioritySpot-check weeklyAssign to the right person or queue
Meeting notesMeeting transcript savedPull out decisions and action itemsApprove before sendingEmail the summary to attendees, add tasks
Invoice and receipt captureNew attachment in a mailboxExtract supplier, date, amountReview anything above a set amountSave the file to a folder and log a row in a spreadsheet
Review repliesNew online reviewDraft a reply in your voiceApprove every replyPost or queue the reply
Daily email digestEvery morningSummarize unread emailNoneSend the summary to you
Pre-meeting briefCalendar event in the next hourSummarize attendees, past emails and attachmentsNonePost the brief in chat
Content repurposingNew blog post publishedDraft social posts and a newsletter blurbApprove before postingSave drafts to your scheduler
Quote request intakeCustomer uploads a documentExtract job details into fieldsConfirm the fieldsCreate the job record and notify the estimator

Several exist as ready-made templates: Google lists Workspace Studio templates including "Get a daily summary of unread emails" and "Send email summaries and action items after meetings", and Zapier's pricing page uses "AI ticket triage" as a worked example. If your workflows are mostly campaigns and follow-up, our directory of AI marketing automation tools covers tools built for that.

Where to build an AI workflow (as of October 2026)

These five platforms all let you add an AI step to a triggered workflow. Prices are from each maker's own page on October 19, 2026. For a ranked list, see our guide to the best AI automation tools; this table only covers how each one handles AI steps.

PlatformHow the AI step worksHow usage is billedEntry paid priceFree option
ZapierAI by Zapier step inside a Zap; pick a model tier, optionally let it call toolsTasks. Each AI step uses 1, 3 or 5 tasks by model tier, plus the same again per tool callProfessional from $19.99/monthFree plan: 100 tasks/month, two-step Zaps; AI steps can be previewed but not published. 14-day Professional trial
MakeAI apps (OpenAI, Anthropic Claude, Mistral AI and others) and AI Agents built in the Make canvasCredits. Most actions use 1 credit; some features using Make's AI Provider "may use more"Not shown on the pricing page we could readFree plan: 1,000 credits/month
n8nAI Agent node and AI steps on the workflow canvasExecutions. One full run counts once, however many stepsStarter €20/month billed annually for 2,500 executions (the page showed euro prices)Free self-hosted Community edition; cloud trial with 1,000 executions, no card
Microsoft Power Automate"Run a prompt" action in a cloud flowFlow licence plus AI usage. New customers buy Copilot Credits to run promptsPremium $15/user/month paid yearlyFree licence for cloud flows with standard connectors; 30-day trial
Google Workspace StudioDescribe the automation in plain language and Gemini creates itIncluded in eligible Workspace plansPart of Workspace Business and Enterprise plansNo separate plan

Zapier

Zapier says it connects to "9,000+ apps." Since June 15, 2026, AI by Zapier steps are priced by model tier: Standard (1x), Advanced (3x) and Premium (5x). Zapier's help article, updated October 1, 2026, says a new AI step on a paid plan "defaults to Premium", and gives the formula: tasks per run = (1 x model rate) + (tool calls x model rate). A Premium step that calls two tools uses 15 tasks. Built-in tools such as Formatter, Paths and Filters use no tasks. Each AI step is capped at 75 tasks per run by default; if a live run hits that, "Zapier will pause the step and ask you to approve it before continuing." On the free plan you can preview an AI step in an unpublished Zap, but publishing it needs a paid plan. The Team plan starts from $69 a month for 25 users.

Make

Make suits owners who think in flowcharts. Its pricing page says most actions "consume 1 credit, while some advanced features leveraging Make's AI Provider may use more." Its AI Agents are "available on all plans", built in the same canvas as your scenarios, and Make says you "can set clear rules, add manual approvals, or stop the Agent at specific points." The free plan includes 1,000 credits a month. Our zapier vs make comparison covers the two side by side.

n8n

n8n bills by execution: "a single run of your entire workflow. It doesn't matter how many steps are in the workflow." That makes long AI workflows easier to budget. Self-hosted without a licence key, n8n "runs as the free Community edition." Its human-in-the-loop feature lets you "require human approval before an AI Agent executes a specific tool", with requests sent through Slack, Microsoft Teams, Gmail, Outlook, WhatsApp and others. Hosted plans store data in Frankfurt, Germany. The trade-off: n8n pitches itself at technical teams. See n8n vs zapier for the comparison.

Microsoft Power Automate

Power Automate fits a business already on Microsoft 365. Microsoft says basic Power Automate use "is included at no cost for work or school accounts" in a Microsoft Entra tenant, limited to standard connectors. AI steps are the complicated part. Microsoft's licensing page says "new customers must purchase Copilot Credits to run AI Builder features", and that from November 1, 2026 new or renewed licences no longer include seeded AI Builder credits. Its rate table lists a standard-model prompt at 1.5 Copilot Credits per 1,000 tokens. Copilot Credits come in packs of 25,000 for $200 a month, or pay-as-you-go through an Azure subscription.

Google Workspace Studio

Workspace Studio is Google's tool for AI-powered automations inside Workspace; Google calls it "an evolution of the capabilities we released in alpha as Google Workspace Flows." You describe what you want, for example "every Friday, ping me to update my tracker", and Gemini creates it. Google lists it for Business Starter, Standard and Plus and Enterprise Starter, Standard and Plus, with connections to Asana, Jira, Mailchimp and Salesforce, plus webhooks. Google says Studio data "is not used to train Google's general AI models outside the customer's domain." It fits businesses that run on Gmail, Drive and Sheets.

If nobody on your team wants to touch a visual editor, our list of the best no code AI automation tools ranks options for non-technical owners.

What an AI step costs: worked examples

These are our own sums from the published rates above, as of October 2026, for an inbox triage workflow that runs 300 times a month.

Platform and setupUsage per monthHow we got there
Zapier, AI step on Standard tier, then 2 ordinary action steps900 tasks300 runs x (1 for the AI step + 2 actions)
Zapier, same workflow on the Premium default2,100 tasks300 x (5 + 2)
Zapier, Premium AI step that calls 2 tools, plus 2 actions5,100 tasks300 x (5 + 10 + 2)
n8n, any number of steps300 executions300 runs x 1 execution each

The Zapier rows show why the model tier matters as much as the number of runs. Switching the AI step from the Premium default to Standard cuts usage by more than half in this example, and Zapier describes Standard models as suited to "simple tasks like classification, extraction, and short summaries." For Make and Power Automate the cost depends on how many credits or tokens each AI call uses, which you only see after running it on real data.

How to build your first AI workflow

  1. Pick a repeated job that involves reading: sorting email, writing up calls, copying details from documents.
  2. Write the steps down by hand first. Trigger, what the person reads, what they decide, where the result goes. Each decision that needs reading becomes one AI step; everything else is a rule.
  3. Start where your data lives. On Microsoft 365 or Google Workspace, check Power Automate or Workspace Studio before adding a subscription.
  4. Give each AI step one job and a fixed output, such as a label from a short list or named fields, so rules downstream can act on it.
  5. Use the cheapest model that does the job. On Zapier that means changing the Premium default to Standard for simple sorting.
  6. Keep a person in the loop at first. Add an approval before anything is sent to a customer or changes a record.
  7. Run it on old data, then watch the log. When corrections become rare, consider removing the approval for low-risk steps.

If you are still deciding where AI fits at all, start with how small businesses use AI.

Risks and how to keep an AI workflow safe

AI output varies. Zapier's FAQ says the same AI step will not always give the same outcome, because of "non-determinism." Fixed outputs, rules and approvals keep that from reaching customers.

Text the AI reads can steer it. The OWASP GenAI Security Project says prompt injection "occurs when user prompts alter the LLM's behavior or output in unintended ways", and that these inputs "do not need to be human-visible/readable, as long as the content is parsed by the model." Any workflow that reads customer emails or uploaded documents is exposed.

Too much access turns small errors into big ones. OWASP's Excessive Agency entry names the root causes as "excessive functionality; excessive permissions; excessive autonomy." Its example applies directly: a step that summarizes email "may only require the ability to read emails", not to delete or send them.

Costs can climb quietly. Tasks and credits scale with use. Set Zapier's per-step task limit and check what your AI steps use before you go live.

The guardrails, in order: connect only the apps and permissions the workflow needs; put an approval step before anything with outside impact (Power Automate, n8n and Make all document one); keep AI output structured so rules can check it; check the run history weekly; and read your platform's data terms before sending customer data through an AI step.

What we don't know

  • Zapier's billing period. The pricing page shows Professional "starting from $19.99/month" but does not label whether that is the monthly or yearly price, or which task tier.
  • Make's plan prices and limits. Make's pricing page did not show plan prices, scenario limits or scheduling intervals in any version we could load on October 19, 2026, so we left them out. Check make.com/en/pricing directly.
  • Make's AI credit usage. Make says some AI features use more than one credit but does not publish a fixed rate on its pricing page.
  • n8n's US dollar prices. n8n's pricing page showed euro prices to us.
  • Workspace Studio usage limits. Google said per-user limits would follow after launch; we did not find them on the pages we read.
  • Accuracy. How often an AI step labels or extracts correctly depends on your data and prompt.

FAQ

What is an AI workflow?

An AI workflow is a series of automated steps, started by a trigger, where at least one step uses an AI model to read, classify, summarize or draft, and the other steps follow fixed rules. A common example is sorting incoming email by topic and drafting a reply for a person to approve.

What is an AI workflow tool?

Software that lets you connect apps, set a trigger and add AI steps without writing much or any code. Zapier, Make, n8n, Microsoft Power Automate and Google Workspace Studio all do this. They differ mainly in how many apps they connect, how they bill AI usage (tasks, credits, executions or Copilot Credits) and whether you can host them yourself.

How do you build an AI workflow?

Follow the seven steps above: map the job by hand, give each AI step one job with a fixed output, use rules for the rest, and keep an approval in place until the log shows few corrections.

What is the difference between an AI workflow and an AI agent?

In an AI workflow, you set the steps and their order, and the AI works inside them. In an AI agent, the AI decides which step to take next and which tools to use. Workflows are more predictable; agents handle work where the path changes each time.

Can I build an AI workflow for free?

Partly. n8n's Community edition is free if you host it yourself, Make's free plan includes 1,000 credits a month, and Microsoft's free Power Automate licence covers cloud flows with standard connectors. Zapier's free plan lets you preview AI steps but not publish them. Running AI steps at any volume usually means a paid plan or paying for model usage.

Do I need to know how to code?

No. Zapier answers "Nope!" to that question on its pricing page, and Workspace Studio builds automations from a plain-language description.

How this page was put together

We read each platform's own pricing, product and help pages on October 19, 2026, plus IBM, Anthropic and OWASP on the definitions and risks. We did not run these platforms. The examples come from vendor templates and documentation, and the cost table is our own arithmetic from published rates. Prices change often, so check each pricing page before you commit.

Sources