Using ChatGPT in the chat interface by now probably is an everyday experience for most. But AI also can be extremely valuable when it comes to automation: GenAI can solve tasks in the background in a digital workflow, e.g. extract certain data from a document, classify emails, detect risks in any kind of text, or suggest decisions based on internal company data.
Still, using AI in digital workflows is a little different. In the chat interface, at least I often aim to extract some knowledge from the LLM: How would you solve this specific problem in Python? Which book would you recommend if I like books A, B, and C? Please suggest 10 diverse examples of goods that pharmaceutical companies typically need suppliers for. At the very core, I rely on the LLM to have that knowledge stored internally and to output it in the right format for me.
But workflows are different: Here we usually need the AI’s capability to think, that is, to derive an answer only by using information from the input text itself. For example, we may give it a contract, ask it to extract a certain clause, and then compare it against a set of criteria that this clause must fulfil. We explicitly do not want to get the LLMs general opinion on that clause, we just check criteria.
As a consequence, in digital workflows, AI tasks usually require less creativity. Plus, they ideally have one clear solution (yes/no, criteria fulfilled/not fulfilled), and they are simple enough so they can be solved by AI reliably. In doubt, we should split tasks into smaller steps and solve them with different prompts one after the other.
By the way, also when designing AI prompts for workflows, we usually start in the chat interface. Simply because this is the perfect environment to test whether AI is capable to solve a task at all, and to improve the prompt iteratively. Only when we have reached a certain level, we can transfer the prompt into the workflow and test at scale.
