The workflow first, the model second
In Building effective agents, Anthropic distinguishes workflows with predefined steps from agents that decide dynamically how to proceed. It recommends increasing complexity only when necessary. This principle is useful for planning a project, even as tools and products evolve. Source: Anthropic, guide dated 19 December 2024.
An example proposed by TM Studio
Suppose we want to organise enquiries received through a website. An initial workflow could validate the fields, create a CRM record, assign an owner and set a reminder. These operations follow clear rules: a language model doesn't need to handle every step.
AI could help at one specific point, such as summarising a free-text message. The original would remain available, and the owner would check the summary before using it to reply. This is a proposed design, not a feature already active on this website.
Plan for when things go wrong
- If the CRM doesn't respond, the enquiry must remain recoverable.
- If the same submission arrives twice, a rule must prevent duplicate records.
- If the text is ambiguous, the record should be passed to a person.
- If the process stops, someone needs to know.
Measure the outcome
For an initial trial, we suggest recording time spent, errors corrected and enquiries left without follow-up. Then compare the same process before and after the change. We don't promise percentage savings without a baseline: the value lies in work that has actually become easier.