The $600K Headline Everyone Missed the Fine Print On

The views expressed here are solely my own. They do not represent the opinions, positions, or policies of any current or former employer, client, or affiliated organization.

It's a great story. It's also an incomplete one, and the incomplete part is the part that actually matters.

Fred Turner, Curative's founder, told the 20VC podcast that his health insurance company killed its Salesforce contract because an internally "vibecoded" CRM now does the job, and does it better. Curative plans to cut roughly 80% of its total SaaS spend this year. The company's AI compute bill has grown sixfold, month over month, for the past six or seven months. Turner also pointed to Gwen, an internal agent that now negotiates provider contracts for about $70 each, down from $1,500-2,000 with a human team, letting Curative do 10-20x the volume.

Those are real, verifiable numbers, and they matter. But there's a sentence buried halfway through the coverage that deserves far more attention than the $600K figure: Turner himself said maintenance is "definitely one of the most challenging pieces."

That's the whole story, actually. Everything before that sentence is the pitch. That sentence is the P&L.

Three things don't add up cleanly when you follow the math instead of the narrative:

First, the savings are gross, not net. Canceling $600K in licensing is real. But nobody has published what two months of engineering time cost in opportunity terms, what the AI compute specific to building and running the CRM costs (the sixfold increase Turner cites is company-wide, not CRM-specific), or what ongoing maintenance is costing now that the tool is live. A number that only counts what you stopped paying, not what you started paying, isn't a savings figure. It's half of one.

Second, two months is a fast timeline for software with almost no room for error. Curative is a health insurer. A CRM touching provider and patient-adjacent data in a regulated environment normally needs data integrity validation, security review, audit trail design, and some form of compliance sign-off, none of which reliably compress into eight weeks no matter how good the code generation is. Salesforce's own response to this story leaned on exactly that point: 150,000 companies still use its platform partly because it's built for exactly this kind of regulatory weight.

Third, "maintenance is challenging" is not a minor caveat. It's the tell. Software that was hard to build fast usually gets harder to run reliably at scale, not easier, and Turner is describing that difficulty in month one, not month twenty-four.

Run the actual math on what "maintenance" costs once it stops being a side project. A single senior engineer dedicated to keeping enterprise software running, fully loaded with benefits, typically runs $180,000-220,000 a year in the US. Add even part-time QA or support coverage and a sliver of product ownership, and a small dedicated team lands somewhere between $400,000 and $750,000 a year, before accounting for the AI compute the tool itself consumes.

That range sits right on top of the $600,000 Salesforce contract Curative canceled. If maintaining this CRM eventually requires anything close to a dedicated team, the "savings" aren't a reduction in cost. They're a transfer of the same number from a licensing line to a payroll line, with Curative now also carrying the uptime, security, and compliance risk it used to pay Salesforce to own. The unanswered question the story doesn't address is whether that maintenance is marginal work absorbed by engineers already on staff, or headcount that gets added later. That distinction is the entire difference between a real $600K win and an accounting illusion.

Curative's internal engineering culture (the same one that scaled from 7 to 7,000 employees in nine months during the pandemic) absorbs the maintenance load without adding headcount, the model replicates across the other tools Turner wants to cut, and this becomes a legitimate build-vs-buy case study other companies can actually learn from.

and the most likely one: the savings are real but smaller than advertised once someone properly accounts for the engineering time already being spent on fixes and support. As Curative scales the volume it's explicitly targeting, the system needs a real owner, real QA, and real governance, at which point it stops being "free" and starts looking like any other piece of enterprise software with its own budget line.

edge cases that two months of testing never surfaced show up in production. Data migration gaps, missing audit trails, permission structures Salesforce had solved a decade ago. Add key-person risk if the engineers who built it move on, and Curative ends up re-buying commercial software after an expensive incident, the same cycle the software industry has run since long before AI made it faster to build the first version.

This isn't a story about whether AI can replace SaaS. It clearly can, for the right kind of workflow, at the right kind of company. Gwen's numbers on contract negotiation are genuinely impressive and reflect real judgment work AI is suited for.

The story is about how easy it's become to announce a win before the bill for maintaining it has arrived. "We built it in two months and it's free now" is a headline. "We built it in two months and here's what it costs us to keep running, eighteen months in" is the actual answer, and almost nobody is publishing that number yet, including Curative.

Build-vs-buy was never a question AI changed the fundamentals of. It just made the build side faster to demo and slower to fully cost out.

Juan Carlos Vásquez has spent ten years inside enterprise content operations, repairing content supply chains before scaling them. Fix to Flow is the discipline that work produced. The views here are his own.