Fred Turner, CEO of Curative, just fired a warning shot across the bow of every major SaaS provider. His company, once a $5 billion COVID testing juggernaut now valued at $1.3 billion as a health insurer, recently ditched its $600,000 annual Salesforce contract. Why? Because an in-house, AI-driven CRM built in two months is simply doing a better job.

Turner isn't stopping there. He expects Curative to slash 80% of its total SaaS spend this year, betting big on AI to automate back-office operations that traditionally eat up enormous budgets and countless hours. The message is clear: the era of expensive, off-the-shelf software might be nearing its end, replaced by nimble, custom AI agents.

Key Takeaways

  • Curative canceled its $600,000 annual Salesforce contract, replacing it with an AI-driven internal CRM built in just two months that is working better.
  • The company projects an 80% reduction in SaaS spend by leveraging AI for core back-office functions, moving away from traditional software vendors.
  • AI agents have shrunk Curative's credentialing process from an average of two to three months and $50 per credential to just 12 hours and 20 cents, leading to the first one that went to zero people in that department.
  • AI can now ingest scribbles on a napkin or any other unstructured format, converting it into a standardized format in about 15 minutes, a task previously prone to delays and errors.
  • Curative uses an AI-Generated Single-Use Python Script for Data Ingestion to tackle its most stubborn data challenges.

The AI-Generated Single-Use Python Script for Data Ingestion

Fred Turner's team isn't just talking about AI; they're deploying specific methods that deliver results. One core strategy they use for dealing with messy, inconsistent data is a novel approach to generating code on the fly:

  • Input any file format: Send us whatever you've got, whatever format. It can be scribbles on a napkin, it doesn't matter.
  • AI agent writes Python script: Literally to give the files to an agent, tell it to write Python to get these files into a standardized format because they're not very good at parsing files. But they're incredibly good at codegen.
  • Test, loop, and iterate script: And so you can tell it to write a Python script to convert any random file into this known format and then test it and loop and iterate on your script until it's working.
  • Convert and discard script: And then you throw away that script. And so it's single use code that never gets used again. and you just generate that code one time and then throw it away.

When This Works (and When It Doesn't)

This method excels in environments with highly varied, non-standardized data inputs. Turner emphasizes that Now send us whatever you've got, whatever format. It can be scribbles on a napkin, it doesn't matter. The model will figure it out. The model will convert it into our standard format and it will do it in about 15 minutes. It shines where traditional parsers fail, or when building custom integrations for every edge case becomes too expensive.

However, this approach isn't a silver bullet. It requires a developer-savvy team to oversee and validate the AI-generated code, ensuring accuracy and security. It's less useful for data that is already perfectly structured or for systems where absolute auditability of every code change is paramount. You're generating and discarding code, which is efficient but might not fit every compliance need.

What to Do With This

Think about the worst, most inconsistent data you regularly deal with. Maybe it's customer feedback forms, partnership agreements, or supplier invoices that arrive in a dozen different layouts. Here's how to apply Curative's framework this week:

1. Identify a Pain Point: Pick one recurring data ingestion bottleneck. Let's say it's new vendor onboarding forms, which arrive as PDFs, Word docs, and even scanned images, all with slightly different fields.

2. Gather Inputs: Collect a small batch of these varied vendor forms. Identify the key pieces of information you need to extract (e.g., vendor name, address, tax ID, payment terms).

3. Prompt an AI Agent: Using an LLM like Claude or GPT-4, describe your problem. Give it a few examples of your messy input files and explicitly state your desired output format (e.g., a JSON object with specific, standardized keys). Tell the agent to write a Python script that takes these inputs and produces that output.

4. Test and Refine: Run the AI-generated script on your sample data. If it fails or misinterprets something, feed the error back to the AI. Point out the specific areas that need correction and ask it to iterate on the script. This loop is crucial for success.

5. Process and Discard: Once you have a working script for a particular vendor's format or a batch of similar forms, run it to extract the data, and then—critically—discard the script. You'll generate a new one for the next unique data challenge, embracing the "single-use" philosophy.