Integration · AI models
OpenAI and WhatsApp: generate replies, images and transcripts from a workflow
How the native OpenAI and ChatGPT apps work in Broadchat automations: generating content and images, asking a model a question, and turning WhatsApp voice notes into text, with setup and limits.
By Broadchat teamPublished 6 min read
You will need
- An OpenAI account with API access and billing set up, since OpenAI charges for API usage separately from your Broadchat plan.
- An OpenAI API key, created in the OpenAI platform dashboard.
- The Automations module on your Broadchat plan, with paid tasks available: each OpenAI or ChatGPT step uses one paid task in addition to OpenAI's own usage cost.
- For voice-note transcription, a publicly reachable URL for the audio file, which Broadchat's media steps can provide from an incoming WhatsApp voice note.
Two apps, one connection
Broadchat's workflow builder has two separate apps that both call OpenAI's models: OpenAI, with Generate Content, Generate Content (Newer Model) and Generate Image, and ChatGPT, with Ask AI and two speech-to-text modules. Both use the same OpenAI connection in Automations → App Authentications, authenticated with an API key from your own OpenAI account rather than a Google-style sign-in. This is a raw, prompt-it-yourself connection to OpenAI's models inside a workflow or chatbot step, distinct from Broadchat's own AI agents, which are built for answering as a business from its own instructions and knowledge base.
Reach for this integration when a workflow needs a one-off generation task that does not belong inside a full AI agent: turning a short product description into ad copy, summarising a long enquiry for a sales agent, generating a quick illustrative image, or converting a voice note into text so the rest of the workflow can read it.
Example: summarising a long enquiry for an agent
A B2B distributor gets detailed WhatsApp enquiries that run several messages long before a customer finishes explaining what they need. A workflow triggered by the chat's Assign to Agent event:
| Step | Module | Key inputs | Output used later |
|---|---|---|---|
| 1 | Assign to Agent (trigger) | Full conversation text | conversation_text |
| 2 | ChatGPT → Ask AI | Instructions: "Summarise this enquiry in two sentences, extracting the product, quantity and delivery city if mentioned." User Message: conversation_text | answer |
| 3 | CRM → Update a Custom Field | Save answer to a "Summary" field on the contact |
The agent picking up the chat sees a two-sentence summary on the contact record instead of scrolling back through the full history.
Example: reading a voice note
A customer sends a voice note describing a complaint in Hindi. A workflow triggered by an incoming audio message runs ChatGPT → Speech to Text (Translation API) with the voice note's public URL, which returns an English transcript. That transcript can then be logged to the CRM, routed with a keyword check, or handed to a support agent who does not speak Hindi, without anyone having to listen to the audio first.
Setting up the connection
Create an API key in your OpenAI platform account and connect it once in Automations → App Authentications. Because this uses your own OpenAI billing, keep an eye on usage from the OpenAI dashboard directly; Broadchat's paid-task count tells you how many times a step ran, not what OpenAI charged for it.
Troubleshooting
- The step fails with an authentication error: the API key may have been revoked or the OpenAI account's billing may be inactive; check both in the OpenAI dashboard.
- Generated replies are too long or inconsistent: lower Max Token and Temperature for shorter, steadier output.
- Speech to Text returns nothing: confirm the audio URL is publicly reachable and under 20 MB, and that it is a direct link to the file rather than a page that contains it.
- Insufficient task to perform the action: your Broadchat plan's paid tasks for the period are used up, separate from any OpenAI billing issue.
For a fuller picture of how AI fits into a WhatsApp business number, see how a WhatsApp AI chatbot works. For connecting other automation tools, see the full list of integrations. Agencies building AI-assisted workflows for clients can offer them under their own brand through the partner program.
Set it up
Create an OpenAI API key
In your OpenAI platform account, create a new secret API key and copy it immediately, since OpenAI does not show the full key again afterwards.
Connect it in Broadchat
Open Automations → App Authentications, click Add Connection, and choose OpenAI (the connection type used by both the OpenAI and ChatGPT apps). Give the connection a name and paste the API key.
Add a generation step to a workflow or chatbot
In a chatbot, add an Add Automation App step and search for Open AI or ChatGPT. In a workflow, add the app after your trigger. Choose your OpenAI Connection and the module you need: Generate Content, Generate Content (Newer Model), Generate Image, or, under ChatGPT, Ask AI.
Write the prompt and pick a model
For Generate Content or Ask AI, select a model (for example gpt-4o-mini for routine replies, gpt-4.1 or gpt-4o for harder questions) and write the prompt, mapping in variables from earlier steps such as the customer's question or saved contact fields.
Use the result in the conversation
The step returns the generated text (or an image URL, for Generate Image) that a later Send Message step can send back to the customer, or that a CRM step can save to a Super Field.
Add transcription for voice notes, if needed
For a workflow that reacts to incoming voice notes, add ChatGPT's Speech to Text (Transcription API) with the audio URL from the message, or Speech to Text (Translation API) if you want the result translated into English.
What you can do
Generate text for a reply or a document
Generate Content, Generate Content (Newer Model) and Ask AI take a prompt and return generated text, for drafting a reply, summarising a conversation, or writing a description to drop into a template variable.
Generate an image from a description
Generate Image takes a text description and returns an image at a chosen size, useful for a workflow that needs a one-off visual, such as a personalised graphic, without a design tool in the loop.
Turn a WhatsApp voice note into text
Speech to Text (Transcription API) converts an audio file to text in its original language; Speech to Text (Translation API) converts it to English text, so a voice note a customer sends can be read, searched or routed like a normal message.
Tune the output
Generation steps expose Max Token, Temperature, Presence Penalty and Frequency Penalty, so you can keep replies short and consistent for a support flow, or more varied for creative copy.
Limitations to know
- This is your own OpenAI account and your own API cost; Broadchat does not supply OpenAI capacity for this integration, and each call is billed by OpenAI in addition to the paid task it uses in your Broadchat plan.
- There is no trigger from OpenAI; every step runs only when a chatbot or workflow reaches it, never on its own.
- The model list is fixed to the options Broadchat has added (for example gpt-4o-mini, gpt-4o, gpt-4.1-mini, gpt-4.1, gpt-3.5-turbo, and o4-mini for Ask AI); a newer model OpenAI releases is not available until Broadchat adds it.
- This is a different thing from Broadchat's own AI agents, which answer from a business's own instructions and knowledge base inside the AI module; the OpenAI and ChatGPT workflow apps are raw model calls you prompt and wire up yourself.
- Speech to Text needs a public URL for the audio file; it cannot read a file that only exists on a customer's phone.
Sources
- OpenAI API reference — OpenAI, checked 12 October 2026
- Speech to text guide — OpenAI, checked 12 October 2026
Broadchat team · Product and partner success team
The people who build and support Broadchat, the white-label WhatsApp Business API platform. We write from what we see while onboarding partners and their clients on the official WhatsApp Cloud API, and we check every article against Meta's current documentation.
Last updated:
Frequently asked questions
Does Broadchat pay for the OpenAI usage in this integration?
No. You connect your own OpenAI API key and OpenAI bills your own account for every call. Broadchat's plan separately counts each step as one paid task, which is a different charge from OpenAI's own usage cost.
Is this the same as Broadchat's AI agents?
No. Broadchat's AI agents answer as your business, using instructions and a knowledge base you set up in the AI module. The OpenAI and ChatGPT workflow apps are direct calls to OpenAI's models that you prompt yourself inside a workflow or chatbot step.
Can I transcribe a WhatsApp voice note automatically?
Yes, with ChatGPT's Speech to Text modules. Transcription keeps the original language; the translation module returns an English transcript. The audio needs to be at a public URL, which Broadchat's media handling can provide for an incoming voice note.
Which OpenAI models can I use?
The models Broadchat has added to the module's list, such as gpt-4o-mini, gpt-4o, gpt-4.1-mini and gpt-4.1 for most generation steps, and o4-mini as an additional option for Ask AI. A model OpenAI releases later is not available until Broadchat adds it to the list.