Data shapes for AI steps in workflows explained simply

Why data shape matters in AI workflows

Every workflow step expects data in a certain shape. That shape can be a single text string, a table, a JSON object, an email, or a file. If the output shape of one step does not match the input shape of the next step, the workflow breaks or behaves in strange ways. As an AI workflow designer you do not need to code, but you must think clearly about what goes in and what comes out of each step. Once you see data as shapes on a conveyor belt, you can plan AI prompts and tool settings so they match and stay stable.

Common input shapes for AI steps

Most AI steps in tools like Zapier, Make or n8n accept a few common input shapes. Single text: for tasks like drafting emails or rewriting content. Key value fields: for prompts that refer to named pieces of data, such as customer_name or invoice_total. Lists of items: for summarising multiple emails or comparing several options. Files or URLs: for reading documents or web pages. When configuring an AI step, read the input description carefully, then map fields so the step gets exactly the shape it expects, instead of raw mixed data.

Typical AI output shapes you will design

AI steps often output plain text by default, like a drafted reply or a summary. For workflows this is only the start. More powerful shapes are bullet lists, tables, or structured JSON. A table output lets you loop through rows, such as tasks or product lines. JSON outputs let you route logic, for example send to Sales if lead_score is high. As a designer you decide which shape matches the next tool. If a CRM expects separate fields, ask the AI to respond in a structured way instead of a long paragraph.

Using JSON and key value pairs safely

JSON is a structured way to send named values, such as {"priority": "high", "owner": "Alex"}. Many platforms let you map JSON fields directly into form fields, spreadsheets or CRMs. The risk is that small AI changes break the shape, for example an extra comment or a missing field. To keep JSON stable, tell the AI to respond with JSON only and list every field you expect. Then validate or parse the output in a separate step if possible. Start simple, for example three fields, and add more once the structure proves stable.

Typical data shape mistakes and how to avoid them

Beginners often send too much mixed data into a step, like whole emails, signatures and system text. The AI then produces outputs that are hard to reuse. Another common issue is changing a prompt so the output shape shifts, for example extra commentary around JSON. To avoid this, strip inputs to what the AI actually needs, and keep business rules in plain language outside the model where you can. When you update a prompt, always re test any steps that rely on that output shape, using both clean and messy real data.

Useful links

• OpenAI — For general information about AI models and capabilities.

Keep learning