How to auto-reply to Outlook emails with a trained OpenAI assistant and Power Automate

Go to platform.openai.com and sign in or create an account. Add a payment method and buy some credit. OpenAI charges by usage, so set a usage limit on the account in case something goes wrong in the flow and it keeps calling the API.

Open the Dashboard menu and select Assistants, then create a new assistant. In the video I name it Automated customer support replies.

Under the name you see the assistant's unique ID (it starts with asst_). Copy it. The flow needs it later.

Paste your instructions into the Instructions box. The instructions in the video tell the Assistant to:

• Expect an email subject and an email body in every message
• Reply only with information from the uploaded documents
• Write professional, precise replies and begin by greeting the sender
• Reply that it cannot help and an agent will get back to them when the answer is not in the documents
• Use <br> tags where a new line starts, so the email reply keeps its line breaks

For the model I select gpt-4o.

Switch on File search, then use Files and Click to upload to add your documents. In the video I upload two files: company information and a list of technical issues and solutions. These files are the only knowledge the bot will use.

Select Playground and send a test prompt that looks like a real email, with an Email subject: and an Email body:. Select Run. The answer should match your document and include <br> tags for line breaks.

Then send an unrelated question, such as who the president of the US is. The Assistant should reply that it cannot help with that. If it answers anyway, tighten the instructions.

In Power Automate, select Create and then Automated cloud flow. Name it something like Outlook email replies with OpenAI Assistant. Search for email and pick the Office 365 Outlook trigger When a new email arrives (V3), then select Create.

If you already have an older version of this flow running, turn it off first so two flows do not reply to the same email.

Add an Initialize variable action named OpenAI API, type String. For the value, go back to the OpenAI dashboard, open API keys, select Create new secret key, and copy the key. You cannot view the key again after you close the dialog, so save it somewhere safe.

Add a second Initialize variable action named Assistant ID, type String, and paste the assistant ID you copied earlier. Renaming each action to match its variable makes the flow easier to read.

The Assistants API normally takes five calls: create an assistant, create a thread, add a message, create a run and read the output. The assistant already exists, and you can create the thread and add the message in one call, so the flow only needs three HTTP requests.

Add an HTTP action and set:

• URI: https://api.openai.com/v1/threads
• Method: POST
• Headers: Content-Type = application/json, Authorization = Bearer followed by the OpenAI API variable, OpenAI-Beta = assistants=v2
• Body: a JSON object with one user message, for example {"messages":[{"role":"user","content":"Email subject: [Subject], Email body: [Body]"}]}

Replace [Subject] and [Body] with the Subject and Body dynamic content from When a new email arrives (V3).

Save the flow, select Test, choose Manually, and send any email to the mailbox to trigger it. Open the HTTP action's output and copy the whole Body. It contains the new thread id.

Back in edit mode, add a Parse JSON action. Set Content to the Body of the HTTP action, select Use sample payload to generate schema, paste the copied body and select Done. Rename the action to Thread ID so you can find it later.

Add another HTTP action and rename it Run. Set:

• URI: https://api.openai.com/v1/threads/ followed by the id from the Thread ID Parse JSON action, then /runs
• Method: POST
• The same three headers as before
• Body: {"assistant_id":"[Assistant ID]","instructions":"[your instructions]"}

Use the Assistant ID variable for [Assistant ID]. In the video I paste the same instructions from the Assistant into this body so the run follows them.

Save and test with Automatically and a recently used trigger. Copy the run's output body and add a second Parse JSON action (rename it Run ID) using that sample, with Content set to the Body of Run.

Add a Delay action set to 1 Minute. In my tests the Assistant needs around 30 to 40 seconds to generate a reply, so one minute gives it room. If you read the messages too early, the reply will not be there yet.

Add a final HTTP action and rename it Output. Set:

• URI: https://api.openai.com/v1/threads/ followed by the thread id, then /messages
• Method: GET, because you are reading data, not sending it
• The same three headers
• No body

Save and test with the last successful trigger. In the output body, the Assistant's answer sits inside data, then content, then text, then value.

Add the Office 365 Outlook action Reply to email (V3). For Message Id, pick Message Id from When a new email arrives (V3) (select See more if you do not see it).

Under Advanced parameters, add Body and enter this expression:

body('Output')?['data'][0]?['content'][0]?['text']?['value']

It takes the first message in the data array (the newest one, which is the Assistant's reply), then the first text item in its content, then its value. Save the flow.

Select Test, choose Manually, and send an email about an issue covered in your documents. The run waits on the Delay action for a minute, reads the output and sends the reply. In the video the bot replies with the exact fix steps from the troubleshooting guide.

Sometimes the bot mentions the name of the source document in the reply. To stop that, edit the first HTTP action and add an instruction to the message content, next to the email subject and body. In the video I add: if the information is not available in the document, reply that you cannot help and an agent will get back to you soon, and do not mention the document reference.

Save the flow and send another test email. Keep adjusting the instructions until the replies read the way you want.