Workflows give users control over how multiple prompts should be combined together at once.

Internally, this leads to more consistent prompting. Instead of team A directing AI to synthesize content in one format, then for team B to write their own prompt to generate an open-text response, workflows allow a single owner to define prompts for every step in a process in a consistent and centralized fashion. They might also define workflows as templates for others to re-use from as needed.

Generative workflows also give us more control over how our internal data flows across systems. Prompts that allow us to access data in third party systems can instruct the AI to retrieve specific data into the centralized workflow for further prompts to reference, summarize, or combine into a RAG-based response as references.

For example, take a workflow that synthesizes customer insights and builds draft knowledge center articles on a weekly basis using a string of prompts. In this scenario:

  • A synthesis prompt analyzes customer insights and summarizes the key takeaways
  • That summary can be run through ChatGPT to process it against a framework for severity and need
  • The AI then generates a sample outline for the most severe knowledge base needs, and adds the remaining ideas to the backlog to keep pulse on
  • Knowledge base managers are alerted by the AI when a new outline is ready to be reviewed

Even with generative AI, this process could take multiple hours to run and manage every week. Instead, AI can orchestrate the process in a matter of minutes. This type of process will be especially common as agents become more common as an AI Service.

Workflows aren't always the right tool for the job, but they are a powerful tool when leveraged.

Details and variations

Generative workflows hone closely to existing analog workflow patterns. They can be built manually or by using an open ended prompt to generate a draft, which the user can modify (Zapier takes this approach).

  • The workflow can contain infinite steps
  • Each step is a self contained prompt or action. Generative and non-generative steps can be combined
  • Steps can integrate into other platforms, either referencing external data or retrieving it into the workflow's core system
  • If a workflow is generated using an open text prompt, consider having the AI state its plan before executing, since creating a workflow can be a data and time intensive process
  • Workflows can be tested by running each prompt once and generating a sample result to test its accuracy



Decrease waste, increase consistency

AI powered workflows allow you to build prompts into your processes in a way that scales. Rather than expecting people to write prompts the same way themselves, or manage access to a prompt library, workflows allow you to define the prompt once and use it over and over again. Users only need to enter information in places where you have indicated in the prompt. This saves you time and increases consistency in your AI work.

Logic-informed prompting

Multi-stepped prompts are difficult to build in a way that returns predictable and accurate results, and even then the AI can sometimes go off the rails. Workflows can break prompts down into their essential building blocks. Set up a step to get a summary of customer insights. Then have ChatGPT summarize it. Synthesize key insights and adds tags by product area. Then email a summary of the insights to each relevant PM. As a prompt, this instruction would be unweildy. As a workflow, each step is simple, guided by logical flows, and can easily scale.

Potential risks

Over complication

Like any workflow, AI-driven workflows can have the tendency to be over built. Ensure you have a way to trace the output through each generative step to review which prompts could be tuned or written incorrectly, leading to sub-optimal results.

Use when:
Multiple generative prompts that normally run together based on the same trigger can be combined into a single workflow to centralize the flow of data and the management of the tasks.

Examples supports mutli-step workflows that can be saved as templates for others to use
No items found.