It’s about a 34-year-old guy from New Jersey, his dying mother, and a workflow built with tools you and I already have access to.
Let me tell you what happened.
On October 20, 2025, Pratik Desai’s mother threw a 70-person Diwali party. She was completely healthy. No warning signs. Nothing.
Exactly one month later, she was diagnosed with Stage 4 duodenal adenocarcinoma, a rare and aggressive cancer of the small intestine.
The hospital’s response? They planned to discharge her without even scheduling an oncology follow-up.
Let me say that again. Stage 4 cancer. And the plan was: go home, good luck.
Pratik is not a doctor. He’s a technologist who worked at Accenture on systems integration, then moved to Salesforce where he led their global personalization practice. He later started his own AI company called 1to1, which exited to ListEngage.
But none of that prepared him for what came next.
“I’m a type A person, so I wasn’t going to take that as an answer,” he said in his Business Insider account. “I’d never been a caregiver before. I don’t know what it looks like without AI. I took the tools in front of me and asked, what can I do now?”
Source: Business Insider via Yahoo (April 5, 2026)
Here’s what Pratik did. And it’s shockingly simple.
Step 1: Daily export from Epic. Every day, he pulled his mother’s medical records from the hospital’s Epic system.
Step 2: Synthesis in NotebookLM. He fed the records into Google’s NotebookLM along with symptoms he observed and things she told him. He’d tell it: “Synthesize the data.” NotebookLM only answers from the documents you upload. It doesn’t hallucinate from external sources. For 1,600 pages of medical records, this was something no human could do manually.
Step 3: Interpretation with Claude. He took the synthesized output to Claude and asked the questions that mattered: “What should I know for tomorrow’s appointment?” “What are the scenarios in front of us?” “What doesn’t look right here that I should push back on?” “What second opinions should I be asking for?”
That’s it. Three tools. NotebookLM for grounding. Claude for reasoning. Epic exports as the raw data.
No custom code. No $50K consultant. No medical degree.
He started with Google AI Studio but switched to Claude because, in his words, “Claude’s models were just getting better. I wanted the best possible model, knowing I had such a short amount of time left with my mom.”
Source: Business Insider via Yahoo (April 5, 2026)
Pratik was at his mother’s bedside from 5am to 10pm every single day for 76 days. NotebookLM was his second opinion. Claude was his interpreter.
Here’s what the workflow caught:
Emergency 1: Christmas Day pulmonary embolism. On Christmas morning, Pratik noticed something off in the way his mother was walking, breathing, and talking. She was pushing back on spending time with the family. The hospital’s emergency line wasn’t returning calls on Christmas.
He typed what he was observing into AI. It determined she was dealing with complications of a pulmonary embolism. He sent an SOS to his cousin who is a doctor, got confirmation, and got her to the hospital. Without AI, he estimates they would have waited 4 to 5 hours for a callback. That delay could have killed her.
Emergency 2 and 3: The blood transfusion bleeding pattern. AI detected a pattern that nobody on the medical team had caught. Seven days after every blood transfusion, his mother would start bleeding dangerously. The cause? 48 hours after each transfusion, the hospital switched her from a liquid diet to solid food. The solid food irritated an ulcer she had, which accelerated bleeding to a point where she was essentially dying from blood loss.
This happened twice. Both times, the AI-powered workflow flagged it before the medical team did.
The CT scan that could have changed everything. The workflow also caught two diagnostic errors in a CT scan report and three instances where the wrong cancer type was stated. They were making critical treatment decisions based on that report. Neither Pratik nor his mother could have read that report themselves, but Claude could.
Source: Business Insider via Yahoo (April 5, 2026)
Some doctors pushed back on Pratik’s approach. They said AI hallucinates. They said it’s only right 70% of the time.
His response is the most quotable line I’ve read all year:
“What if we took the medical system and graded it the same way? We expect AI to be perfect, but the medical world isn’t.”
That reframe matters. Because Johns Hopkins research has found that more than 330,000 patients die annually in the US from diagnostic errors alone. The medical system is not infallible. It never was. The question isn’t whether AI is perfect. The question is whether having a second set of eyes, even imperfect ones, is better than having none.
His mother lived 76 days after diagnosis, 67 as an inpatient. She got to say her goodbyes. She got to kiss his two-year-old daughter.
“There was not a single instance where we felt like the medical world was actually looking out for her goals as a patient,” Pratik said. “Mainly, an optimized length of life to say the goodbyes she wanted to say.”
Source: Business Insider via Yahoo (April 5, 2026) | Humai.blog (April 14, 2026)
Pratik’s story is powerful on its own. But it’s part of something much bigger.
Matt Rosenberg saved $163,000 on a hospital bill using Claude. After his sister-in-law’s husband died of a heart attack, the hospital sent a bill for $195,628. Rosenberg opened Claude and did a line-by-line audit. Within an hour, it identified duplicate billing codes, a charge for a coronary bypass surgery that never happened, and a ventilation management fee that Medicare explicitly prohibits when a critical care code is also present.
He didn’t just accept Claude’s output. He ran the same analysis through ChatGPT to cross-check, then spent 20 minutes verifying the Medicare billing rules against primary regulatory documents. He wrote a six-page letter citing specific violations. The hospital reduced the bill to $33,000.
A medical billing attorney would have charged $500 to $2,000 for that work. Claude costs $20 a month and is available at 2am.
A cancer patient used AI minutes after her biopsy results. A woman diagnosed with cancer in both breasts simultaneously uploaded her biopsy report to an AI chatbot within minutes of receiving it. It translated the staging into plain language and generated questions for her oncologist appointment three hours later.
“I had a much stronger baseline understanding of what was happening,” she told Cancer Today magazine. “It doesn’t replace medical advice, but it is a fantastic bridge to help you engage better with your medical team.”
One of Pratik’s friends used the same workflow for his own mother. He studied up using the workflow, then called a meeting with his mother’s medical team. The doctors were so impressed by his understanding of the case that they kept asking if he was a doctor or biology major. He told them he’s a marketer.
The numbers are staggering. A January 2026 OpenAI survey found more than half of respondents had used AI tools for healthcare advice over a three-month period. An estimated 40 million people worldwide are using AI for healthcare daily.
Source: Humai.blog (April 14, 2026) | Business Insider via Yahoo (April 5, 2026)
If you’re building in AI, here’s what this story is telling you:
The killer app isn’t a model. It’s a workflow. Pratik didn’t build a custom model. He chained together existing tools (Epic exports + NotebookLM + Claude) into a repeatable workflow. The value wasn’t in the AI. It was in the structured process around it.
The moat is trust and accessibility, not technology. Anyone can access Claude and NotebookLM. What’s rare is the trust layer: the workflow design, the verification process, the human judgment about when to act on AI output and when to double-check. If you’re building healthtech, build there.
Vertical AI wins when it solves life-or-death problems. Not every vertical has these stakes. But the ones that do (healthcare, legal, financial compliance) are where AI creates the most undeniable value. You don’t need to convince anyone that catching a CT scan error matters.
The subscription math works in the patient’s favour. A $20/month Claude subscription versus a $2,000 medical billing consultant. A free NotebookLM tool versus hiring a patient advocate at $500/day. This economics gap is where new businesses get built.
You don’t need to be a technologist to do what Pratik did.
If someone you love gets a diagnosis, here’s the workflow:
Get daily records from their patient portal (most hospitals use Epic or MyChart)
Upload them to NotebookLM and ask it to synthesize
Take the synthesis to Claude and ask: what should I know, what looks wrong, what questions should I ask the doctor
Verify anything critical against primary sources before acting
Show up to every appointment with better questions than the doctor expects
That’s it. That’s the whole thing.
You’re not replacing doctors. You’re becoming a better-prepared participant in your own care. Or your parent’s care. Or your partner’s care.
And sometimes, that preparation is the difference between catching a misdiagnosed CT scan and not.
Pratik’s mother lived 76 days after diagnosis. Three of those were directly extended because of AI-assisted interventions.
But here’s the part that stays with me.
She got to kiss her granddaughter. She got to say goodbye.
The AI didn’t cure her cancer. Nobody is claiming that. But it gave her family something the medical system wasn’t prioritising: time. And the dignity of being properly cared for.
If a guy with no medical background, a Claude subscription, and a free Google tool can catch errors that a medical team missed, what does that say about the information asymmetry we’ve accepted as normal in healthcare?
And what are you going to do about it in your own life?