What is an AI workflow?
An AI workflow is a written procedure for a repeated task in which an AI tool does one or more steps and a person does the rest, including the check. 'Every Monday, paste last week's support tickets, ask for themes using our saved prompt, check each theme against three sample tickets, then post the summary to the team' is an AI workflow. 'Use ChatGPT to help with support' is not.
The term is also used for automation: software that passes data between apps and calls an AI model without a person in the loop. That can be valuable, but it is a different project with different risks. This page is about the version most teams need first, which needs no code and keeps a human check in every run. Once a manual workflow has run reliably for a few weeks, you will know whether any part of it is safe to automate.
What does a good AI workflow contain?
A good AI workflow names six things: the trigger, the inputs, the prompt, the check, the sign-off and the output. If any of them lives only in one person's head, the workflow breaks the first time that person is on holiday.
Trigger
When the workflow runs: every Monday, when a new supplier contract arrives, when a ticket is escalated.
Inputs
Exactly what material goes in, and what must be removed first. 'Ticket text only, customer names replaced with [CUSTOMER]' is specific; 'the tickets' is not.
Prompt
The saved prompt, with the role, the audience, the format and the rules, such as 'only use facts in the material; mark anything missing as [CHECK]'.
Check
What the reviewer verifies and how: facts against the source, figures recalculated, promises compared with policy, tone read aloud.
Sign-off
Who decides the output is fit to use, and what happens if it is not.
Output
Where the finished work goes and in what format, so the next person knows what to expect.
How do you build an AI workflow step by step?
Build the workflow from a task you already do well by hand, not from a tool's feature list. Doing the task manually first tells you what a good result looks like, which is the standard the AI version has to meet.
1. Pick one repeated task
Choose something weekly or daily, with source material and a reviewer, where an error is caught internally.
2. Write down how you do it now
List the steps, the inputs and what you check. This becomes the skeleton of the workflow.
3. Mark the step AI could draft
Usually the first draft, the summary or the extraction. Keep judgement steps with the person.
4. Write and test the prompt
Run it on three past examples where you know the right answer. Fix the prompt until the drafts need only small edits.
5. Write the check
Turn what you corrected during testing into a short checklist. The errors you saw in testing are the ones to look for every time.
6. Hand it to someone else
Ask a colleague to run it from the written version alone. Every question they ask is a gap in the document.
7. Review after a month
Look at a few outputs, the reviewer's corrections and any near-misses. Keep, change or retire the workflow.
What does a real AI workflow look like?
Northgate Lettings is a fictional 30-person lettings agency. Property managers spend Friday afternoons writing update emails to landlords about repairs. The agency turns this into a workflow. Trigger: every Friday. Input: the week's repair log entries for one property, copied from the maintenance system, with tenant names replaced by [TENANT]. Prompt: 'You are writing a short, factual update to a landlord. Use only the log entries below. For each repair, state what was reported, what has been done and what happens next. If a cost, date or contractor name is not in the log, write [CHECK]. Do not apologise on the agency's behalf or promise completion dates.'
The first test run shows why the rules matter. Without them, the draft told one landlord the boiler would be 'fixed early next week', which nobody had confirmed, and estimated a repair 'at around £150'. With the rules, the draft lists the boiler as 'engineer visit booked; completion date [CHECK]' and leaves the cost out. The check is three lines: every date and cost appears in the log, nothing is promised that the contractor has not confirmed, and the landlord's name and property address are added in the agency's own email system, never in the AI tool.
Sign-off sits with the property manager for routine updates and with the branch manager for any repair over the agency's approval limit. The workflow is one page, kept in the shared drive, and a new property manager ran it correctly in her second week from the document alone. That is the test that matters.
Want to see how this connects to a role?
Explore the planned team trial ↗Why does every AI workflow need a human check?
Because the tool will occasionally produce confident errors, and people who use a workflow every week gradually stop looking. NIST's Generative AI Profile describes confabulation, the production of confidently stated but erroneous or false content, and automation bias, excessive deference to automated systems, which it says can make other risks worse. A written check counters both: it tells the reviewer exactly what to look at, and it keeps looking from becoming optional.
Keep the check short enough to do every time. Three or four specific questions beat a long list nobody finishes. If the check takes longer than doing the task by hand, the workflow is not saving anything and should be redesigned or retired.
Which AI workflow automation tool is best?
There is no best tool in general, and most teams should not choose one until a manual workflow has proved itself. Automation platforms connect apps and can call AI models on a schedule, which removes copying and pasting, but it also removes the natural pause where a person reads the output. If you automate, keep a review step before anything goes to a customer, and check where the platform stores the data it passes through.
Automation is a better fit when the inputs are structured, the output goes somewhere internal, and errors are cheap to reverse, such as tagging incoming emails for a person to handle. It is a poor fit for anything that sends messages, changes records or makes decisions without someone looking first.
When is an AI workflow the wrong answer?
Skip the AI step when the task is rare, when the inputs are too sensitive to put into your approved tool, or when a template would do the job. A standard letter that changes two fields each time needs a mail merge, not a model. A task that needs judgement about a specific person, such as a performance concern, should stay with a person.
And if a workflow keeps producing outputs the reviewer rewrites from scratch, retire it. That is useful information, not a failure.
How do people learn to run AI workflows well?
By running them on realistic examples with feedback before the stakes are real. Practise the check as much as the prompt: give people a draft with planted errors, such as an invented date or a cost that is not in the source, and see whether they catch them. Our prompt templates guide gives reusable structures for the prompt step.
Tosheo.ai is being built around this kind of practice. Its planned practice scenarios are role-plays in which the learner uses AI on a realistic work situation and explains what they changed and why, scored against a visible rubric. It is pre-build; the planned pricing is on the pricing page.