November 2026 · For small practices

State of AI in Healthcare

Since 2016, Y Combinator has backed 690 healthcare companies. In 2016, one in ten mentioned AI in its one-line pitch. In 2025, almost nine in ten did. That is the market that is about to call your front desk. Here is the trend, what it means, and how to start getting ahead of it.

Part 1. The trend

AI went from a feature to the default

Share of healthcare companies whose one-liner mentions AI, by the year they joined Y Combinator.

690 companies. 2026 is a partial year (75 companies so far), shown as an open dot.

When nearly every new company says "AI," the word tells you nothing. What matters is the job it does and how you will know it did it right.

It landed first on the work that runs your office

Share mentioning AI by who the company serves, 2025 and 2026 batches. Number of companies at right.

Small groups swing on one or two companies. Read the order, not the decimals.

Front desk, operations, billing and insurance went from 14% of new companies in 2016 to 2019 to 36% in 2024 to 2026. These are the jobs a small practice runs on one or two people, so those people are who the vendors are calling.

Part 2. What to expect

The next 12 months, in plain terms

More pitches, all alike

Every front desk and billing vendor now says AI. Expect the same promise from many names: answer every call, book every slot, code every claim.

AI wins where the result is easy to check

A call answered, a claim coded, an appointment booked. It is weaker where the answer takes judgment and slow feedback, like whether a patient is getting worse. That follows the argument from Princeton's Arvind Narayanan, who calls it a hypothesis.

The checking stays with you

A tool can do the task. Someone in your practice still reviews it, catches the misses and answers for the result. A tool is a win only if the total work goes down.

The rules keep moving

Payers change requirements, visit types get renamed, templates get edited. A tool trained on last quarter's rules keeps running on them until someone tells it.

The insurance side is automating too

Companies built for insurance and prior auth are adopting AI at 64% in the latest batches. Expect more automated back and forth with payers.

Part 3. What to do

Five steps before you buy anything

  1. Pick one job you can check.Missed calls, claim denials or scheduling gaps. Choose a task where you can count the result.
  2. Get a baseline first.For one week, write down how long that job takes and how often it goes wrong. You cannot tell if a tool helped without it.
  3. Name one owner.One person in the practice reviews the tool's output and answers for it. No owner means no one notices when it slips.
  4. Write down your definitions.Visit types, codes, payer rules and who decides. One shared list means staff and tool work from the same meaning, and you can see what changed.
  5. Start short.Run a 30 day trial or a month to month term. Prices are falling fast, so a long contract at today's price is a bet.

Seven questions for any vendor

Part 4. How to start getting ahead

The practices that come out ahead

They are not the ones with the most tools.

One tool at a time

Add the next one only after the first has a measured result.

A monthly 15 minute check

Look at the numbers, ask what changed in the payer or schedule rules, and update your list of definitions.

People stay on the hard calls

Use AI for the routine and the checkable. Keep a person on anything a patient finds sensitive or unclear.

Know your own workflow

The vendors are building for the average office. The practice that can say how its day really runs gets the most out of any tool.

Want a second set of eyes before you decide?

My Communication Audit looks at how your practice talks about what it does, with patients and with vendors. It is free for practices that reach out.

Book a 15 minute intro call

Sources and method. Company data comes from the public dataset by Garrison Research of healthcare companies in the Y Combinator directory, 2016 to 2026. "Mentions AI" is a yes or no tag on the one-line description, so it undercounts companies that use AI without saying so. Shares and segment groupings are calculated from that data. Y Combinator is one accelerator. It shows what early-stage founders think will sell, not what practices have bought.

The idea that AI improves fastest on tasks that are easy to verify comes from Arvind Narayanan. It is a conditional argument, not an established result. The steps and questions above are my own practical judgment, not legal or financial advice.