Teams that already pay OpenAI or Anthropic and want records summarised, sorted and drafted inside their automations, on their own account, with their own choice of model and their own agreement with the provider.
Run AI steps on your own OpenAI account — drafting, summarising and classifying inside automations.
Run Claude steps on your own Anthropic account — drafting, summarising and classifying inside automations.
A one-line summary on each record lets somebody scan a list of enquiries without opening any of them. The step runs as each record arrives, so the summary is there before anybody looks.
You'll know it worked: Add a test enquiry, or press Run now, which uses the table's newest record. Summary should fill in within a few seconds, and the run history should show the step with no error.
Sorting is where a model's near misses hurt, because a category that almost matches is invisible to every view grouped by it. Ask for one word from a fixed list, and let RowFold match the answer to the field's options.
You'll know it worked: Send three test requests written to belong clearly to one category each. Filter the grid by Category and check each one sits under the option you expected, spelled exactly like the option.
A draft saves the blank-page minutes on routine replies while a person still decides what goes out. Keep it in a field, and nothing reaches the customer until somebody sends it.
You'll know it worked: Create a test request and open it. Draft reply should hold a reply that uses the customer's name and plan, ready to edit.
Pasting a record into a chat window, asking for a summary and copying the answer back works once. Nobody does it for the four hundred enquiries that arrived this month.
A Generate text step runs by itself when a record is created or changes, with the record's values in the prompt through tokens. Save the result to puts the answer in a Text or Long text field, where the grid, filters and exports can use it.
Asking a model to pick a category produces near misses such as “sales”, “Sales.” or “The category is Sales”. A near miss isn't one of the field's options, so the record looks sorted while every board and filter grouped by that field passes over it.
Write the answer into the choice field with an Update the record step using {{connector.result}}. An answer that matches an option apart from case, quotes or a closing full stop is stored in the option's own spelling.
RowFold has AI of its own, which makes it unclear when connecting a provider is worth doing at all.
The built-in Generate with AI step needs nothing connected and uses the workspace's monthly allowance of 500 AI actions per editor. Connect your own account when you expect more runs than that, want to name the model, or need your company's agreement with the provider to cover the data.
Sending customer records to a model is a decision about data, and it is hard to see exactly what an integration sends.
The connected step sends the prompt you wrote, with its tokens filled in, as one request to your own account. Nothing else from the record goes with it, so the prompt is a complete list of what leaves.
Only what the prompt contains. The step sends the prompt you wrote, with its tokens filled in from the record, as a single request to your own account, and nothing else from the record goes with it. What the provider does with that request is governed by your account's terms, so check them if the records are sensitive.
Any text model your account is allowed to call. Leave Model blank and the step uses a sensible default, or type a model's name from the provider's list. A name your account can't use fails the step and the run history says why, so try a new model with Run now before relying on it.
No. Steps that run on your connected account are billed by your provider and leave the workspace's allowance alone. The allowance, 500 AI actions per editor each month, covers RowFold's own AI features, including the Generate with AI step. A workspace can use both: its own key for the high-volume job and the built-in step elsewhere.
Not through Save the result to, which accepts Text and Long text fields only. Use an Update the record step that sets the choice field to {{connector.result}}. An answer that is one of the options apart from case, quotes or a closing full stop takes the option's own spelling; anything else is stored exactly as it came back.
Anthropic replies stop at 4,096 tokens, roughly 3,000 words, and OpenAI replies at the model's own limit. Saved into a field, an answer keeps its first 4,000 characters. Neither limit matters for a summary or a category, and for anything longer the reliable fix is in the prompt: ask for the length you want.
There is no template to wait for. Type a sentence about what you run and RowFold assembles the tables, links, views and dashboard around it — then you change whatever does not fit.