TL;DR: You might have expected AI to cut healthcare costs, whether it’s by reducing paperwork, automating the doctor’s notes, or thinning out hospital staff. But a new 60-page PwC report suggests the reverse: So far, one of its most widespread uses is making medical bills bigger. It’s an example of how AI isn’t only good at making tasks more efficient—it’s also very good at finding more granular ways to boost a sector’s bottom line.

What happened: AI is one of five potential drivers of health costs climbing up to 9% in 2027—matching this year's rate, the highest since 2010–11—per PwC. The key reason: AI note-taking tools are documenting more specifics about diagnoses and medical complications that a rushed human clinician might have lumped into one broad “code”—a standardized billing label that tells insurers what to pay. Those extra details can justify a higher severity (read: higher paying) code, even if the actual care a patient receives is the same as before.

The devil is in the billing details: One Blue Cross Blue Shield analysis found that some hospitals saw the billing code for acute posthemorrhagic anemia in new mothers jump from 4% to 12.3% of maternity admissions between 2022 and 2025. The number of blood transfusions (a common treatment for this condition), meanwhile, barely budged. An audit of the hospital system with the steepest rise in this code found that fewer than 20% of the cases actually met the clinical criteria for a diagnosis. The rise in higher-intensity coding coincides with hospitals’ growing use of AI for billing. According to BCBS, “coding intensity” added $22 million to maternity spending at the hospitals studied in three years.

The big but: AI is the report's top-ranked new pressure, but it's not the biggest driver of costs overall—old standbys like labor and supply costs still account for more of the increase, one of the report's authors told Healthcare Dive. And AI tools could eventually push the other way, driving down costs by automating hospital administrative work or catching diagnoses earlier.

Bottom line: AI is often pitched as a way to optimize whatever industry it touches—trimming waste and making systems faster and cheaper. But in healthcare, one of the first things it has optimized is how to charge you more. As one health insurance exec put it: Companies “will take AI and say, ‘How can I use this to further my self-interest?’” —WK

A few prompts to help you actually understand your work data

I've spent a lot of time over the past five months feeding data into AI—reader survey results, click data, open rates—and the difference between a useless response and a genuinely insightful one usually comes down to how you frame the ask. Here are the prompts that have worked for me, made generic enough to use on whatever data you're staring at.

1. Pressure-test your own read on the data.

The biggest unlock for me wasn't asking AI to summarize the data—it was writing down my own takeaways first, then asking AI to check them against the numbers by uploading the raw data files to a project:

"Here are my takeaways from the data for [X project]: [paste your takeaways]. Analyze the data itself and tell me which of my conclusions it supports, which it contradicts, and which it complicates. Be specific about why, and flag anything important I haven't noticed. Before you get started, is there anything you need to know about the project itself?"

That ordering matters. If you read AI's analysis first, it'll anchor your thinking and you'll lose the chance to compare your instincts against the data cleanly.

2. Find the patterns you're not seeing.

Then, in a fresh chat (so it isn't anchored by what you've already told it), go fishing:

"Analyze this data and identify the patterns that show up consistently. Then tell me: What's surprising here? What would you expect to see that's missing? If you could only tell me three things from this data, what would they be and why?"

The "what's missing" part is the underrated bit—absence can often be more interesting than presence (or at least spark new questions).

3. Make it repeatable.

If you look at the same kind of data regularly, ask AI to help you build a reusable structure:

"Based on this analysis, create a template of questions we should analyze this data against every [week/month]. Structure it so results are comparable across time periods and we can see patterns emerging."

I did this for our click data, which I review in depth every month, and it means each analysis builds on the last instead of starting from scratch. And, again, AI is only as good as the data you give it. Garbage in, confident-sounding garbage out. —SM

If you’re using AI to dig into data in an interesting way, or have a topic you want us to dig into, let us know.

Here’s Claude’s response when asked what Anthropic should do if AI really causes mass unemployment: Fork over taxes high enough to make Sen. Bernie Sanders blush, and get transparent about which jobs its products are replacing. It even compares AI firms to tobacco and oil companies—publicly debating the harms they cause while privately lobbying against fixes.

Claude is a lot blunter than the economic policy framework Anthropic actually released Wednesday, which pledges about $350 million toward researching solutions for AI-driven job loss—roughly $1 per American, a rounding error next to the walloping that Great Depression-level unemployment could dole out. The internet was unimpressed by the paper, in which the company admits that it’s “less certain about the right answers” if the worst unemployment scenario comes to pass. (“Maybe don't leave the plan in the hands of the people who made the problem,” one Redditor wrote.)

The AI industry’s “concept of a plan,” as Gizmodo put it, lands amid a flurry of attempted image rehab by AI execs. (See: Sam Altman suddenly “delighted to be wrong” about a jobpocalypse, Dario Amodei backpedaling the white-collar wipeout prophecy, and Jensen Huang saying AI is a "lazy" excuse for layoffs.) But it’s hard to run a charm offensive when even your own chatbot is skeptical of your solutions. —WK

Chaos Brewing Meter: /5