Vesta raised $30 million to replace lenders' loan origination systems with AI agents. TD expanded agentic AI across its mortgage operations, built on a 200-person internal AI lab. And Valon, which raised $150 million earlier this year, now runs all of Newrez's loan servicing, a dependency large enough that Newrez's parent company flagged it as a risk in an SEC filing and went ahead anyway. Three companies, three answers to one question: who holds the operational memory of a mortgage business, and what does the institution lose if it walks away?

Vesta raised $30 million to replace lenders' loan origination systems with AI agents. TD expanded agentic AI across its mortgage operations, built on a 200-person internal AI lab. And Valon, which raised $150 million earlier this year, now runs all of Newrez's loan servicing, a dependency large enough that Newrez's parent company flagged it as a risk in an SEC filing and went ahead anyway. Three companies, three answers to one question: who holds the operational memory of a mortgage business, and what does the institution lose if it walks away?

Medical AI tools fall into five steps, ordered by how close each one gets to making the medical decision itself. From least to most: writing notes (typing up what was said at a visit), paperwork (insurance pre-approvals, summarizing a patient's records), spotting risk (flagging which patients a doctor should check), suggesting a diagnosis (proposing what's wrong) and acting alone (caring for the patient with no doctor involved). The FDA, the U.S. agency that approves medical products, is moving toward judging each finished tool by what it does, rather than by the general-purpose AI underneath it. The higher the step, the more checking a tool faces. A note-writing tool needs only the doctor's signature on the note. A tool that reads medical scans needs FDA sign-off.

At the bottom step, note quality no longer separates the tools. Most now write notes that are about equally good. Abridge, Ambience and Nabla are moving into billing codes and insurance pre-approvals. Epic, whose patient-record software many hospitals run on, is building its own note-writer (high confidence). Near the top, Aidoc is betting that one general system can spot many conditions. Viz.ai stays with a few. It also holds its own Medicare payment codes, so hospitals get paid extra when they use it. (Medicare is U.S. government health insurance for older people.) That makes Viz.ai the clearest case on this list of a tool paying for itself.

Medical AI tools fall into five steps, ordered by how close each one gets to making the medical decision itself. From least to most: writing notes (typing up what was said at a visit), paperwork (insurance pre-approvals, summarizing a patient's records), spotting risk (flagging which patients a doctor should check), suggesting a diagnosis (proposing what's wrong) and acting alone (caring for the patient with no doctor involved). The FDA, the U.S. agency that approves medical products, is moving toward judging each finished tool by what it does, rather than by the general-purpose AI underneath it. The higher the step, the more checking a tool faces. A note-writing tool needs only the doctor's signature on the note. A tool that reads medical scans needs FDA sign-off.

At the bottom step, note quality no longer separates the tools. Most now write notes that are about equally good. Abridge, Ambience and Nabla are moving into billing codes and insurance pre-approvals. Epic, whose patient-record software many hospitals run on, is building its own note-writer (high confidence). Near the top, Aidoc is betting that one general system can spot many conditions. Viz.ai stays with a few. It also holds its own Medicare payment codes, so hospitals get paid extra when they use it. (Medicare is U.S. government health insurance for older people.) That makes Viz.ai the clearest case on this list of a tool paying for itself.

• Writing notes: Abridge (built into Epic, sold to hospital technology chiefs); Ambience (100+ medical specialties, pitched on helping hospitals bill correctly and earn more); Nabla (works with any patient-record system, 35+ languages, free basic version); Suki (doctors give it voice commands, and it drafts orders for the doctor to approve)
• Paperwork: Cohere Health (handles pre-approvals for insurance companies); Navina (summarizes records for doctor groups paid on how healthy their patients stay); Commure (all of a hospital's office paperwork)
• Spotting risk: Healthleap (risk scores from health records; 50+ hospitals, by the company's own count). Watch this: the FDA says that if the "information" a tool gives gets specific enough, it counts as telling the doctor what to do, and that brings more FDA checking
• Suggesting a diagnosis: Aidoc (31+ FDA sign-offs, about 2,000 hospitals); Viz.ai (alerts the whole stroke team at once; Medicare pays hospitals up to $1,040 extra per patient)
• Acting alone: ARPA-H, a U.S. government health research agency, offers up to $62.7M through its ADVOCATE program for AI that checks in on heart-failure patients after treatment. Research only. No tool at this step is allowed in use yet

Time-sensitive: FDA's comment period on generative-AI medical device evaluation closes Monday. The proposed framework would judge the application's behavior, not the foundation model underneath it. If that holds, it resets the evidence bar for clinical AI. Read the Healthcare card before the long weekend.

Across all eight verticals this week, the structural moves outweigh the product moves. Schneider is spending $23.7B on data substrate. Ledgebrook earned rated carrier status. Shopify opened checkout infrastructure to every agent. Thrive raised $2B to own the firms it deploys AI into. Each of these creates a competitive position that takes years to replicate. Harvey and Google made large moves; neither has yet proved the depth that would change their vertical's structure. Watch the gap between scale of investment and evidence of operational impact — it's the pattern running through half these cards.

Time-sensitive: FDA's comment period on generative-AI medical device evaluation closes Monday. The proposed framework would judge the application's behavior, not the foundation model underneath it. If that holds, it resets the evidence bar for clinical AI. Read the Healthcare card before the long weekend.

Across all eight verticals this week, the structural moves outweigh the product moves. Schneider is spending $23.7B on data substrate. Ledgebrook earned rated carrier status. Shopify opened checkout infrastructure to every agent. Thrive raised $2B to own the firms it deploys AI into. Each of these creates a competitive position that takes years to replicate. Harvey and Google made large moves; neither has yet proved the depth that would change their vertical's structure. Watch the gap between scale of investment and evidence of operational impact — it's the pattern running through half these cards.