Using AI to automate an internal FDA/HHS newsletter

API usage cut by three-quarters inside a fixed, restricted hour budget.

The client, a US Federal Government Agency, needed an efficient and reliable way to summarize the week’s top AI news into a digestible email format. They approached a design agency, which subcontracted the task to Common People.

Government agency responsible for healthcare and related technology

The project presented two significant challenges:

1. Limited Project Hours: The client had restricted hours they could dedicate to the project.
2. Overflow of Content: The vast and fluctuating volume of AI-related information made it difficult to filter out valuable content from the sensationalist and clickbait material.

Given the overarching impact AI discussions have on various industries, the client sought to sift through immense amounts of data to provide only relevant information to their audience. Traditional PR monitoring tools were either prohibitively expensive or lacked the necessary specificity.

Common People utilized their expertise in building AI workflows to create a streamlined, efficient process. The basic workflow consisted of the following steps:

Identifying Trending Articles

Using Feedly, Common People pinpointed trending articles based on specific keywords.

Checking Relevance and Applying Categories

Articles are then run through an AI-powered workflow that validates AI and healthcare relevance while filtering marketing or PR-focused articles.

Scoring and Sorting

Relevant articles are scraped and scored using a custom prompt that outputs scores in JSON key: value format, enabling articles to be sorted by score in the database. The top 10-14 articles per category are then manually screened and compiled into the newsletter every week.

Refining the Process

This seemingly straightforward workflow was refined iteratively:

• Developing Scoring Criteria: Initially, there was no clear definition of category types or audience value. An iterative approach refined these definitions.
• Detailing the Rubric: A detailed rubric was created to resolve the AI’s tendency to overly score articles.
• Press Release and Relevance Filtering: Additional steps were prepended to this workflow to filter out marketing-heavy press releases and ensure articles genuinely centered on AI and healthcare.

1. Efficiency in Filtering: By refining the process, Common People reduced the number of articles needing intense review, thereby cutting API usage for processing by three-fourths. This reduction was crucial due to the high cost and inconsistent results from the GPT-3.5 model compared to the more reliable GPT-4.
2. Sustainable, Low-Code Solution: The final workflow utilized no-code tools like Zapier and Make.com to allow non-technical team members to maintain and adjust the system easily.
3. High User Satisfaction: The end product not only saved significant hours but also delivered high-quality, relevant content. The client’s feedback highlighted the utility and simplicity of the tool in keeping up with valuable AI trends.