Ask an AI model in a flow

Coming soon. AI model steps aren't switched on for workspaces yet. Until they are for yours, there's no AI model providers section in Settings → Integrations and no Ask an AI model step in the flow builder, and nothing on this page applies.

An Ask an AI model step sends an instruction you write — plus only the inputs you choose — to ChatGPT (OpenAI), Claude (Anthropic) or DeepSeek, and uses the answer in the rest of the flow. For example: read what someone said they're looking for, and sort them into buy, book or question.

• Your own account. You connect your own API key. The provider bills your account directly, at its own prices, for what your steps use. HyperDM charges nothing for these steps, and HyperDM's own AI agent never uses your key.
• Only what you choose is sent. The instruction you write and the inputs you tick — nothing else from your workspace.
• Two branches. Answered — only when the answer matches the format you set — and Didn't work for everything else.

Who can do what: owners and admins connect a provider; only the owner sets its monthly token limit; owners, admins and agents set up, test and read the results of the step. See Roles and permissions.

Connect your AI model provider

1. Go to Settings → Integrations and find the provider's card under AI model providers — ChatGPT (OpenAI), Claude (Anthropic) or DeepSeek.
2. Click Where to create a key and create an API key on the provider's page. For OpenAI, a key with restricted permissions needs the Models permission.
3. Paste the key into the card's API key box (for example OpenAI API key).
4. Click Connect ChatGPT (OpenAI) (or Connect Claude (Anthropic), Connect DeepSeek). The button reads Checking with the provider… while HyperDM checks the key with the provider.

If the provider refuses the key, nothing is saved and what you typed stays in the box, with the reason under it. If it accepts it, HyperDM stores the key encrypted and never shows it again, and the card reads Connected with the number of models your key can use.

Set a monthly token limit

Owner only — and required. With no limit, no step runs and no test is sent.

1. On the same card, find Monthly token limit.
2. Type a whole number into Tokens per calendar month on the provider. The box shows a suggested starting number.
3. Click Save limit. It says Saved once HyperDM has stored it.

The limit counts the tokens every step sends and receives on this key per calendar month (UTC), tests included. A step that would go over it isn't sent and takes Didn't work. The card shows how much of the limit this month has used.

Add the step to a flow

1. Open a flow and Save it — a step can only be set up on a saved flow.
2. From Add step → Advanced steps, add Ask an AI model. On the canvas it reads Not set up, with two branches: Answered and Didn't work.
3. Click the step. Its panel shows two separate lines: the provider (for example "ChatGPT (OpenAI) — connected") and This step (Not set up). A connected provider doesn't mean the step is ready.
4. Click Set up. The Set up “Ask an AI model” window opens. Closing it keeps what you've typed.

Choose the provider, the model and the instruction

1. Under Provider and model, choose the Provider and the Model. The models are the ones your key can use, with the recommended ones first. Refresh models reads the list again (owners and admins).
2. Under Instruction, write What should the model do? — up to 6,000 characters. It's sent exactly as you write it: nothing in it is filled in. If you type {{, the window reminds you that inputs are sent separately — choose them as inputs instead.

Choose what the model may see

Under Inputs — what the model may see, tick each value the model may read. Nothing else is sent.

1. Tick from the groups: Contact field, Contact detail (for example their first name or @username), Sheet lookup, Earlier AI answer (from AI model steps whose Answered branch leads to this one), Run fact (flow name, step name, run time) and Workspace variable.
2. Under Chosen, set each input's longest value in characters, or Remove it. Up to 24 inputs.

Each input travels as its own labelled block after your instruction, never mixed into it. A workspace variable whose name looks like a secret — a key, a password, a token — is marked "(looks like a secret — can't be sent)" and can't be chosen. The check leans careful, so a harmless name such as day_pass or map_pin can be caught too: rename it under Settings → Workspace → Workspace variables and tick it again. The message that started the conversation isn't an input: to use what someone said, ask them with a Data collection block that saves to a contact field, and tick that field.

Choose the answer's format

Under Answer, choose one:

• Text — one answer, kept as {{ai.<name>}}. Give it a Name (lowercase letters, numbers and _; it starts as text, or text_2 when another step in the flow already keeps text — and a step brought in from another flow starts on the name its later messages already use), set Longest answer (characters), optionally Save to a contact field, and tick May contain links if links should be kept.
• Structured fields — named values, each kept as {{ai.<name>}}. For each field: Name (lowercase letters, numbers and _), Type (Text, Number, Yes / no or Choice), Choices (separate with commas) for a choice, Longest value (characters) for text, an optional Description for the model, Required, May contain links for text, and Save … to a contact field. Add a field adds another, up to 12.

HyperDM checks every answer against this format itself. An answer that doesn't match takes Didn't work. Without May contain links, any link in the answer becomes "[link removed]". With it on, a link the answer keeps is sent exactly as a link you type into a message would be.

Answer names must be different from every other AI model step's in the same flow, because {{ai.<name>}} has to mean one thing. That includes a Text answer's Name: a second AI model step with a text answer in the same flow needs a name of its own, for example reply.

If you rename an answer, anything that still uses the old name — a message with {{ai.<old name>}}, a Condition, another AI model step's input — reads empty. Design checks lists each one, so you can update it to the new name.

Set the limits

Under Limits:

• Longest answer (tokens) — from 1 to 4,096; 300 to start.
• Inputs, at most (characters) — from 1,000 to 60,000; 12,000 to start. A run whose inputs are longer sends nothing and takes Didn't work.
• Time limit (seconds) — from 3 to 30; 20 to start.
• Try again after a timeout or a busy provider — Don't try again, Once (after a minute) or Twice (after 1, then 5 minutes). A timed-out request may still be billed by the provider.
• Creativity (0–1) — only for models that take it. 0 answers the same way most often.

Check what gets sent, then test it

1. Under What gets sent, read Instructions (system) and Message (user): exactly what the step sends, with your sample values in place.
2. Under Test, type a Sample value for each input.
3. Click Test with these values…. The window asks first: "This sends a real request to OpenAI and is billed to your account." (or Anthropic, or DeepSeek). If the provider needs reconnecting, isn't connected, or has no monthly token limit yet, it says so instead and sends nothing. If you fixed that in another tab since this page opened, reload the page.
4. Click Send the test, or Don't send. While it waits you can click Cancel — a request that already reached the provider may still be billed.

The result reads Answered, with each {{ai.<name>}} and its value, or Didn't work, with the reason. Both show the tokens used, the cost estimate (or "price not known") and how long it took. If the provider didn't report the tokens, the counts are marked "estimated from the limit". The same test can answer differently each time. A test never changes your setup, and changing the provider, the model or a setting clears the result, so it always matches what you see.

Save the step and the flow

1. Click Save step. HyperDM checks the setup once more — the provider still connected, the model still offered, the answer names free in this flow, the contact fields still there — and saves it. If anything's wrong, the reason shows next to the part to fix, and everything you chose stays.
2. Save the flow. The panel reminds you: "Setup saved. Save the flow to keep this step's link to it." If you changed the provider or renamed an answer, it says "Setup saved. Save the flow now — if it's live, this step takes “Didn't work” until you do, because its link, provider or answers changed." On a live flow, save it straight away.

The step now reads Ready — or No token limit until its provider has a monthly token limit (see above). If someone saved this step's setup in another tab while you were editing, the window shows their version above yours, and keeps it there if you close and reopen it: Save step again to keep yours, or Use their version.

Use the answer and connect both branches

• In a later message, click Personalize and pick from AI answers to insert {{ai.<name>}}.
• In a Condition, branch on ai.<name> — for example ai.intent is buy. An optional field the model left out is empty: is empty is true for it.
• Answered runs only when the answer matched your format.
• Didn't work runs for everything else — a timeout, an answer in the wrong shape, a key the provider stopped accepting, your monthly token limit. On it, the step's {{ai.…}} values are empty. Give it a path of its own, such as a message that hands the conversation to your team.
• A step answers once per run. Don't loop Didn't work back into the same step to try again — it returns the same result without asking the provider. Use Try again after a timeout or a busy provider under Limits, and to ask again later in the conversation, add a second AI model step.

The flow can't go live until the provider is connected, a monthly token limit is set and every AI model step is set up and ready; Set live says what's missing first.

Check how the step is doing

Open the step and expand Run results: the last 30 days of runs and tests, each Answered, Didn't work, Retrying or Running, with the reason in words, how many attempts it took, the tokens, the cost estimate, and when a waiting run tries next. The instruction, the inputs and the answer are never shown there. The provider's card in Settings → Integrations shows this month's totals and Recent runs across all your flows.

When a step needs review or reconnecting

• Needs reconnecting (on the card and the step): the provider stopped accepting the key. Every step on it takes Didn't work until you reconnect on the card.
• Needs review: the model isn't offered for your key any more (click Choose another model), the provider was reconnected with a different account, HyperDM's connection to the provider changed, or an input looks like a secret. Click Review setup, fix it, and Save step.
• Copied from another step: you copied or duplicated a step. Click Copy the original's setup — inside the same flow, the window opens with the original's setup filled in so you can rename its answers and save.
• Switched off: AI model steps or this provider aren't switched on for your workspace. The step takes Didn't work, its setup is kept, and you can view it but not change it.
• Design checks says a step "keeps answers its saved setup no longer gives": its setup was saved with new answer names, but the flow wasn't saved after it (for example the page was reloaded in between). Open the step, click Edit setup and Save step, then Save the flow. Until then the flow can't go live, and a step already live takes Didn't work without sending anything.

What AI model steps don't do

• They only run on your own key — there's no option to use HyperDM's AI.
• They never run, open or fetch anything in an answer; it's text a later message can show.
• They don't read the message that started the conversation, or anything you didn't tick.
• They don't stream, use tools, read images or hold a conversation — one request, one answer.
• A step that comes from another flow — a copied, shared or imported flow included — keeps its place on the canvas but not its setup: it reads Not set up, so set it up again there. Another workspace connects its own key.