‹ Retour

Best Practices for Automating Data Workflows with AI

 

By Vladimir Williams, PhD

02/07/2026

 

Workflow automation is one of AI's most promising use cases, however, deploying AI in production workflows requires much discipline. AI models are not a replacement for human judgement, but rather powerful tools which can take on much of the weight of data workflows. The question is not whether to use AI or not, but how to use it responsibly. Based on current methods and my own experience, here are some best practices for integrating AI into automate data workflows while minimizing hallucinations and maintaining quality.

Design Solutions Around Human Review

Use AI for What It's Good At (and Nothing Else)

Minimizing Hallucinations Through Structure and Context

For example, I recently built an AI-powered workflow automation for a client. The problem was straightforward: automate the generation of a Gantt chart using a text document as an input. This automation would knock hours off of their workflows while eliminating human error introduced by manual chart generation. One could have proposed to input the text document to an LLM and ask it directly to output a Gantt chart, however, this route would have limited the capacity for human review and modification.

 

Instead, I designed a pipeline which would input the text document into the LLM (using an API) to extract the relevant data and export it as a JSON file, which would then be fed into a deterministic Python code. The Python code used the JSON file to generate an Excel spreadsheet formatted as a Gantt chart. Because the data structure and its rules were known, several validations could also be made by the Python code, for example, checking to make sure the input's event start dates came before their end dates. The result was a hybrid solution which joined a probabilistic AI step with a deterministic step to finish the workflow while checking the AI output.

Path to Production: Iterate and Test