Learn

AI agents

AI workflow examples: 12 business processes that run on their own

By Faisal SaeedUpdated 6 min read

In short

An AI workflow is a business process written as steps that runs on its own every time something happens, with an AI model or agent doing only the steps that need judgement. Good examples are qualifying a website lead, drafting replies to support e-mails, thanking customers who paid, sending a weekly report and onboarding a new hire. Each one has a clear start, a few fixed steps, and a person who approves anything that leaves the team, moves money or changes access.

An AI workflow is a process that runs on its own every time something happens. It does fixed steps in order, and asks an AI model or agent only for the steps that need reading and judgement. Below are 12 real business processes, grouped by team. For each, we list what starts it, its steps, and where a person approves.

If you are not sure whether a job is a workflow or an agent, read AI agent vs workflow first.

What makes a good AI workflow?

  • A clear start. A form is sent, an e-mail arrives, a payment comes in, Friday comes.
  • Steps you can write down. If you can explain the process to a new colleague, it can be a workflow.
  • Judgement in one or two steps. Sorting, scoring, summarising or drafting is where AI helps.
  • A person where it matters. Anything that leaves the team, moves money or changes access waits for an approval.

AI workflow examples for sales

1. Qualify a new website lead

  • Starts: the website contact form calls the workflow.
  • Steps: look up the company; score the lead from 1 to 10 with a short reason; add a row to the leads sheet; post leads that score 7 or more to the sales channel in Slack.
  • Person approves: nothing. Everything stays inside the team.

2. Follow up after a sales call

  • Starts: every weekday at 6 pm, on a schedule in the rep’s time zone.
  • Steps: read that day’s sales calls from the rep’s Google or Outlook calendar; for each one, find the call notes in Google Drive or OneDrive; an agent drafts a follow-up e-mail with the next steps that were agreed; send the e-mail.
  • Person approves: the rep approves the e-mail before it goes to the customer.

3. Sort the website contact form

  • Starts: the website contact form calls the workflow.
  • Steps: check every field; sort the message into a lead, a customer issue, a question, a partnership or spam; skip spam; post the rest to the right Slack channel with a label.
  • Person approves: nothing. Spam is skipped, the rest goes to people.

AI workflow examples for customer support

4. Answer a support e-mail

  • Starts: a new e-mail arrives in the support inbox in Gmail or Outlook.
  • Steps: search the help articles; the support agent drafts the reply with the sources it used; send the reply.
  • Person approves: a support lead approves each reply, at least until the team trusts the drafts.

5. Triage a new Jira issue

  • Starts: a new issue is created in a Jira project.
  • Steps: sort it into bug, feature request or question; suggest a priority with a reason; find similar past issues in the knowledge; post a summary to the team’s Slack channel.
  • Person approves: nothing for the summary. Changing the issue’s owner or priority in Jira can wait for the team lead.

6. Handle a refund request

  • Starts: the refund form on your website calls the workflow.
  • Steps: look up the order through your order system’s API; an agent checks the request against the refund policy; ask a person “Refund $90 to Ana?” with Approve and Reject; issue the refund; e-mail the customer.
  • Person approves: always, before any money moves.

AI workflow examples for finance

7. Thank a customer who paid

  • Starts: Stripe calls the workflow when an invoice is paid.
  • Steps: keep only paid invoices; write a one-line thank-you; send it; log it in the payments sheet.
  • Person approves: nothing, or the first few messages while you check the tone.

8. Check a supplier invoice

  • Starts: a new e-mail with “invoice” in the subject arrives in the finance inbox.
  • Steps: pick out the supplier, amount and order number; find the matching purchase order in your finance database (read only); add a row to the invoices sheet, marked matched or not matched; post the ones that do not match to the finance channel.
  • Person approves: the workflow never pays. A person in finance decides what is paid.

9. Send the monthly spend summary

  • Starts: a schedule, on the first of every month at 9 am in your time zone.
  • Steps: collect last month’s spend by team from your finance database; an agent writes a one-page summary with the largest changes; save the report as a file; e-mail it to the finance lead.
  • Person approves: nothing. The report goes to one internal person.

AI workflow examples for HR

10. Onboard a new hire

  • Starts: an HR agent, when someone asks it to onboard a new colleague.
  • Steps: create the first-week plan from the onboarding checklist; book intro meetings on the calendar; send the welcome e-mail; tell the team in Slack.
  • Person approves: the HR lead approves the calendar invites and the welcome e-mail before they are sent.

11. Announce a policy change

  • Starts: a file changes in the Policies folder in SharePoint or Google Drive.
  • Steps: an agent compares the new version with what the handbook said before and writes what changed in plain words; post the summary to the company announcements channel in Slack.
  • Person approves: HR approves the summary before it is posted.

AI workflow examples for operations and IT

12. Send the weekly engineering status

  • Starts: a schedule, every Friday at 4 pm.
  • Steps: read the week’s issues from Jira or Azure DevOps; an agent writes what shipped, what is blocked and who owns each blocked item; post it to the engineering channel in Slack.
  • Person approves: nothing for the summary. Closing or moving issues would ask.

How do these examples compare?

Workflow Starts on Judgement step Person approves
Qualify a new lead Website form Score 1 to 10 No
Follow up after a call Every weekday evening Draft the e-mail Before it is sent
Sort the contact form Website form Sort the message No
Answer a support e-mail New e-mail Draft the reply Before it is sent
Triage a Jira issue New issue Sort and prioritise Only for changes
Handle a refund Website form Check the policy Always
Thank a paying customer Stripe Write the thank-you No
Check a supplier invoice New e-mail Match to the order Payment stays manual
Monthly spend summary Schedule Write the summary No
Onboard a new hire An agent Plan the first week Before invites and e-mail
Announce a policy change File change Say what changed Before it is posted
Weekly engineering status Schedule Write the summary No

The pattern is the same as for agents: reads and internal posts run freely, anything that leaves the team asks, and money always asks. See human in the loop AI for how to choose approvers.

How to build these workflows in promptev

Each of these can be said to the promptev workflow builder in your own words, such as “when the refund form is sent, check the order, ask me before refunding, then e-mail the customer”.

  • The builder drafts it. It checks what your project can use, names any app you still need to connect, and shows the workflow as a diagram with what one run costs.
  • Every start in the list is real. A schedule in your time zone, a new e-mail in Gmail or Outlook, a calendar event, a file in Google Drive, OneDrive or SharePoint, a Jira issue or Confluence page, a call to the workflow’s own web address (signed by Stripe, GitHub or promptev), a regular check, or one of your agents.
  • Judgement from your agents. A step can ask one of your published agents, with its own knowledge and tools, or ask a model for an answer in a set shape, such as a score from 1 to 10, that is checked before the next step uses it.
  • A person decides. An approver named for a step decides by e-mail or in promptev, and a step can ask a person a question with the answers you choose.
  • Test it, then publish. A test is one real run that shows what each step did. Each run after that has its own page, with the version it ran and what each step got and gave back.
  • Pay for the work. A run costs 1 credit to start, 1 for each agent or model step and 1 for each tool call, at $0.005 a credit, with a credit limit on every run. New workspaces get 1,000 free credits for a 30-day trial.

For the full reference, see the documentation on workflows.

Frequently asked questions

What is an example of an AI workflow?

A support e-mail workflow: a new e-mail arrives, the workflow searches the help articles, an AI agent drafts the reply, a person approves it and the workflow sends it. The steps are the same every time; only the drafting needs judgement.

What starts an AI workflow?

Something that happens: a schedule, such as every Friday at 5 pm; an event in an app, such as a new e-mail or a changed file; another system calling the workflow, such as a website form or a payment; a regular check for new items; or an AI agent that starts it when a request is that workflow's job.

Which steps in an AI workflow need a person's approval?

Steps that send messages outside the company, move money, change access or change records that are hard to undo. Steps that only read, sort or draft for internal use usually run without approval.

How many steps should an AI workflow have?

As few as the process needs. Most useful workflows have three to six steps: a start, one or two steps that need judgement, an approval where it matters, and the action. A long process can often be split into two workflows.

What is a good first AI workflow?

One that happens often, has a clear start and only posts or saves something inside the team, such as sorting website form messages into Slack or a weekly summary. It is useful from the first day and low risk, because nothing leaves the company.