The Same Questions, A Different Vertical

On the architecture questions every AI-native vertical company must answer — and why care is the domain where the answers are paid in human presence.

I read an interview this week with John Glasgow, who is building an AI-native ERP for accounting. Almost every question the interviewer asked him is a question someone should be asking me. Our verticals are unrelated. The architecture questions are identical. So I answered them.

When you started, it was manual work before AI. What broke in that process?

The thing that broke was the assumption that the exception is rare.

In home care back-office work, the exception is the job. A claim denies because a caregiver clocked in four minutes late on a visit six weeks ago, and the authorization units for that service code ran out two days before anyone noticed, and the documentation that would have supported the appeal was never attached because the supervisor was covering a shift that day. That is not an edge case. Medicaid managed care plans deny claims at an average rate of 16.7% — the highest of any payer class. And by HHS Office of Inspector General accounting, 89% of denied prior authorizations are never appealed at all.

So the first thing we got wrong was building automation for the clean path and calling the rest "manual fallback." The clean path was never where the money was. The money was in the two claims out of a thousand that nobody had the hours to fight.

What we build now inverts it. The system handles the volume and then routes the exceptions somewhere specific — to a named human, with a named authority, in a named domain. Not to a queue that fills up. Not back to the owner's sister who does the billing twenty hours a week from her kitchen.

How do you know when the model fails? What has the error rate looked like over time?

Let me answer the more useful version of this question, which is: what would I have to be able to prove for that answer to mean anything?

Here is the number the system puts at risk — on CareBravo's Managed Billing plan, where the Careonomy service entity named in the customer agreement runs the billing and stakes its fee. A 99% clean-claim rate, in writing, measured from electronic visit verification and shift approval through to billing. Not "accuracy" in the abstract — a specific rate, on a specific measurement path, with a defined start and end.

And the form of the commitment matters more than the number: when the number is missed, the miss costs the fee — automatically, as credits written into the agreement, not as a negotiation after the fact. Not the customer's fines. Not indemnities. The fee, on the performance. I architected the system that makes a commitment like that survivable; the Careonomy service entity named in the agreement is the one that signs it. And the commitment attaches to a payer regime only after we have seasoned that regime — a rule we hold ourselves to rather than grant as a sales favor.

That last clause is the whole answer to "how do you know when the model fails." You know because you wrote down, in advance, what would count as failing, and you priced yourself so that failing costs you money.

Anyone can report an error rate. The interesting question is whether they were willing to attach a fee to it before they knew the answer.

Do you run your own model?

No. And I think the instinct to want one is the most expensive mistake available to a vertical AI company right now.

Our doctrine is: own the harness, rent the model.

The harness is the part that doesn't commoditise. It's the specification format every agent is rendered from. It's the confirm-back gate that fires before any write. It's the append-only ledger that makes every action reconstructable eighteen months later when an auditor asks. It's the authority matrix that says which role may do which thing at which risk level. It's the escalation routing that names, per category of exception, exactly which human sees it. None of that improves when a frontier lab ships a new model, and none of it degrades either. It is ours.

The model is a registered parameter. Per agent. Changeable only through a gate that replays a fixed corpus and compares results. If a better model appears on Tuesday, we change a value and re-run the corpus. If we had spent two years training our own, Tuesday would be a threat instead of an upgrade.

The companies claiming a proprietary foundation model as their differentiation are, in most cases, describing a fine-tune on someone else's base wrapped in a good harness — and the harness is doing the work. I would rather say the true thing and be right about which half matters.

How competitive is it to hire AI engineers?

I don't compete for the San Francisco engineer who has an offer from a frontier lab. I would lose, and the loss would be correct.

Our engineering team sits a long way from the Bay Area, and I won't pretend the economics of that are incidental — they aren't. But the economics are not the reason the arrangement works. The reason it works is that our hardest problems are not model problems. They are specification problems, verification problems and workflow problems. What separates a good engineer on our team from an average one is whether they can hold a payer's billing rules and a state's regulatory overlay in their head at the same time, and write a spec precise enough that an agent can't wander. That skill is not concentrated in one zip code.

The pitch I make is the same one every vertical AI company should be able to make and most can't: nobody's home care operational data is on the internet. Not the authorization patterns. Not the denial reasons by payer by state. Not what a good supervisory note looks like when a surveyor reads it. All the code is public, all the case law is public, and this is not. Which means the models arrive at our domain undertrained, and every operational outcome we verify is a signal that didn't exist before we captured it.

Engineers who want to be the ten thousandth person tuning a chat interface don't take that pitch. Engineers who want to build the corpus do.

How long is AI a competitive advantage before it's table stakes?

It's already table stakes and it was never the advantage.

Every home care software vendor now has an AI announcement. Some of them have shipped something real. Within eighteen months the answer to "do you have AI" will be yes across the entire category, and it will mean as little as "do you have a mobile app" means today.

What doesn't become table stakes is the shape of the commitment. Software gives an agency tools its team operates. We deliver the completed result. Nine outcomes, owned end to end — scheduling and payroll, compliance, billing, real-time communication with families and caregivers, documentation, marketing, hiring, retention, training. Not nine features. Nine results, with the work finished rather than assisted. CareBravo carries the AI-native work; where a function requires sustained human execution, Careonomy carries that layer — named, transparent, and measured on its own line.

That's not a model capability. It's an accountability structure, and you can't ship it in a release note. A competitor who wants to match it has to change what they sell, how they price, who they hire, and what they're willing to lose money on. Most won't, because their existing business is the reason they can't.

What's your right to win against incumbents adding AI on top of what they have?

Two answers, one technical and one about the buyer.

The technical one is the same one you hear from every AI-native challenger, and it happens to be true: intelligence layered over a data model that can't carry the detail produces confident summaries of things nobody checked. If the underlying system only stores the aggregate, the agent reasons over the aggregate. Ours stores the event — append-only, corrections as new events that supersede, never overwrites — because the audit is the product and the audit needs the sequence, not the total.

The answer about the buyer matters more. Our customer is the single-location independent owner with no franchisor behind her. She has no royalty-funded marketing, no corporate training program, no back office waiting in another state. The implicit promise is everything a franchise fee buys, without the franchise.

She also cannot evaluate our software. She has no test environment, no analyst, and no spare Thursday. The standard enterprise answer — extended sandbox access so the buyer can validate on their own data — assumes a buyer with the capacity to run a sandbox. She doesn't have it.

So the price does the evaluating for her. The floor is published and stated up front. There are no per-seat fees and no modules that nickel her. The percentage plans stay in the low single digits of what she bills, and the clean-claim guarantee credits our fee back automatically when we miss. A competitor whose data is thinner than ours cannot offer that structure without losing money on it — which means the price is not a marketing decision, it's a claim about the corpus, made in a form she can verify without technical skill.

What does this look like in three to five years?

I think the interface question is settled and everyone is just catching up to it. Dropdown menus, thirty screens of configuration, a human clicking through a workflow someone designed in 2011 — that is going away, in our category and in every category. An owner will talk to one agent, that agent will route to specialists, and the specialists will do the work.

The part I think people are getting wrong is what happens to the humans.

The popular version is that the exception queue shrinks toward zero and the specialists disappear. I don't believe it, and our architecture doesn't assume it. Every regulatory change, every new managed care contract, every state that revises its waiver creates a fresh population of exceptions nobody has ever resolved. Judgment work doesn't get automated away in an environment that keeps changing the rules. It gets relocated — from the owner's kitchen table at eleven at night, to a specialist who has seen the same kind of exception again and again across many agencies.

That's the version I'd bet on. Not fewer humans. Humans placed where judgment actually is, with the volume taken off them.

And underneath it, the thing that compounds: every one of those resolutions gets captured with its outcome — not just what the specialist decided, but whether the payer paid. Frequency is not evidence. A pattern that recurs often and resolves badly is worse than no pattern at all, and a knowledge system that weights by how often it saw something rather than whether the answer worked will confidently teach itself to be wrong. We weight by the verified downstream result. It is slower and it is the only version that doesn't compound its own errors.

Tell me about internal AI tooling.

Everyone on the team is expected to attempt a task with AI before doing it by hand. That part is unremarkable now and anyone not doing it is behind.

The part I'd push back on, in the version of this I read elsewhere, is the free-for-all. Build whatever you want, post it in a channel, celebrate the velocity. I understand the appeal and we get most of the benefit a different way.

Internal tools at our company are chartered. There's a workbook that defines the goal, the tasks under it, the person who makes each one and the different person who checks it. Unchartered work isn't buildable — not because it's forbidden, but because nobody has said what it's for or who verifies it. Maker and checker are never the same person. Nobody merges to production alone.

It sounds slower than a Show & Tell channel. It is slower, for the first month. Then the fiftieth internal tool still has an owner, and the person who built the eleventh one has left, and the thing still works.

The claim I'd rather be judged on

As of August 2026, CareBravo serves 130 care organizations. As those agencies grew, their back offices did not grow with them — zero added back-office hires across the program cohort. That clause is the whole thesis: not that the agency grew, but that it grew without the administration growing with it.

Every payer in the world is compressing the administrative envelope — by mandate, by rate cut, by demographics. That pressure is not a policy episode and it won't be reversed by an election. Which means the only durable answer is a substrate that costs less than the envelope it replaces.

AI carries the volume; humans remain where judgment, authority, relationships and care require them. That is the operating truth. The manifesto version, the one I'd put on the wall: AI does the admin; humans do the caring. Everything else we've built is downstream of knowing the difference.

Anand Chaturvedi® is Founder & CEO of Caryfy AI, building economic infrastructure for care, and the author of The Care Manifesto™.