Everyone in AI asks some version of the same question: what's your moat?
It's a fair question. It's also, these days, usually asked by people who assume the answer is the model.
It isn't. It was never going to be.
Satya Nadella said it plainly: foundation models are becoming increasingly similar and widely available, and "models by themselves are not sufficient" for a lasting competitive edge (Unite.AI).
Sit with that for a second. The most capable software ever built is available wholesale, to everyone, at roughly the same price. Your competitor buys the same intelligence you buy. Asking what's your moat in that world is like asking what's your moat when everyone gets electricity from the same grid.
So the question gets sharper: if intelligence is a commodity, what isn't?
Hamilton Helmer spent a career on this. In 7 Powers, his framework for durable competitive advantage, he argues that a real strategic power has exactly two parts: a benefit (you earn more than competitors) and a barrier (something stops them from copying you). No barrier, no power — just a temporary lead (summary).
I like that test because it's brutal. It kills most AI pitch decks on contact.
Warren Buffett's version is older and blunter. In his 2007 shareholder letter he wrote that "business history is filled with 'Roman Candles,' companies whose moats proved illusory and were soon crossed" (as quoted here).
A demo is a Roman candle. A fine-tuned model is a Roman candle. A fundraise announcement is a Roman candle.
Let me tell you what isn't. I've spent 25 years in healthcare technology — EMRs, revenue cycle, utilization review, workers' comp — and the last eight running real clinical operations. Here's my honest inventory of what actually defends a business when the model belongs to everyone.
Denial data
We have a record of more than a hundred thousand clinical review decisions. Not just the outcomes — the reasoning. What was requested, what evidence was considered, what was approved, what was denied, and why, case by case, for the better part of a decade.
No model vendor has this. No competitor can scrape it. You cannot download it, license it, or prompt-engineer your way to it. It only exists because someone did the work, one file at a time, for years.
Helmer would call this a cornered resource: preferential access to an asset competitors literally cannot get. He's right. Every denial letter we ever wrote is a small piece of a dataset that cannot be recreated without reliving the decade.
This is the part of AI strategy nobody wants to hear, because it's slow. Data moats aren't built. They're accumulated.
And the data teaches things no model knows. Which treatments get denied and then quietly approved on appeal. Where the guidelines say one thing and the actual decisions say another. The patterns that only appear when you've watched the same decision get made ten thousand times. A foundation model trained on the public internet has never seen any of this. It lives behind our firewall, in files no crawler will ever touch.
War stories
Our system runs on 12,366 deterministic rules. Each one of those rules is a scar.
Rule 47 exists because of a specific denial that went wrong in a specific way in 2019. Rule 83 exists because a payer tried something clever and we had to answer it. I remember the cases behind most of them. That's not a figure of speech — when you've run the operation, the edge cases have faces.
Helmer calls this process power: operational excellence embedded so deep that competitors can observe it but can't replicate it. The barrier isn't secrecy. It's that you can't shortcut experience. A competitor can read our marketing page. They can't read the eight years.
Every AI demo works. That's what demos are for. Production is where the edge cases live, and edge cases are the entire business in healthcare. Anyone can generate a prior authorization letter. Almost no one can tell you what happens when the adjuster calls back angry, the physician disagrees, and the deadline is Friday.
We've handled that call. Thousands of times.
The workflow, not the portal
Here's a mistake I watch smart teams make: they build a beautiful product and ask the customer to come to it. Another login. Another portal. Another tab.
We did the opposite. PhyPal is an orchestration layer inside the EHR and practice systems teams already use — not another portal. The work happens where the work already happens.
This is the least glamorous moat on the list and maybe the strongest. Once you're inside someone's workflow, displacing you means asking a busy clinic to change how they work. That's Helmer's switching cost, except the cost isn't contractual — it's human. Nobody has time to relearn the thing that already works.
Distribution beats product. Embedded distribution beats everything.
Governance as the product
In workers' comp, every clinical decision has to be defensible. Not just correct — provable. Who decided, on what evidence, under whose authority, with what audit trail.
So we built the audit trail into the foundation: every input, every reviewer, every outcome, captured at the decision level. AI does not make clinical decisions. Governed humans do. The system exists to make that sentence true and provable.
Here's what took me years to understand: in a regulated market, the ability to prove how a decision was made is the product. The model is interchangeable. The proof isn't.
When a regulator, a payer, or a plaintiff's attorney asks "show me exactly how this decision was reached," there are two kinds of companies. The ones that can produce the record in ten seconds, and the ones that schedule a meeting to figure out what happened. That gap is a moat you can drive a truck through.
Nobody buys governance because it's exciting. They buy it because the alternative is terrifying.
The people
I've worked with hundreds of physicians over 25 years. That sentence doesn't sound like a moat until you've watched a medical director take your call at 9 PM because you've been right for a decade.
Trust compounds. It compounds slowly, then all at once — the same way it does in every other part of life. Helmer files this under branding, and he's describing the mechanism correctly even if the word feels too soft: a willingness to work with you, specifically, that a competitor can't buy with a better feature list.
You earn it by being right, repeatedly, in public, for years. There is no growth hack for it. That's the point.
What the moat is not
It's not the model. Everyone has the model.
It's not the demo. Demos don't survive contact with a real operation.
It's not being first. Being first to a commodity is just being early to a price war.
And it's not a secret. I'm writing the whole thing down. Knowing the recipe doesn't give you the kitchen, the staff, or the decade.
The short version
Denial data. War stories. Governance. The workflow you're already inside. The people who pick up the phone.
None of it is exciting. All of it took years. That's exactly why it works.
The moat was never the intelligence. Intelligence is wholesale now. The moat is everything around it that can't be wholesaled: the decisions, the scars, the proof, the trust.
You can't download any of that. You have to live it.
I write about the operating details of this at claimatix.ai/blog. The frameworks referenced above: Hamilton Helmer's 7 Powers, and Buffett on moats via Berkshire Hathaway's shareholder letters.
References
- Unite.AI. "Are AI Models Becoming Commodities?" (quoting Satya Nadella). unite.ai
- Helmer, Hamilton. 7 Powers: The Foundations of Business Strategy. Deep Strategy, 2016. Summary: befreed.ai
- Buffett, Warren. Berkshire Hathaway 2007 Shareholder Letter. berkshirehathaway.com
- Tools for Investors. Coverage quoting Buffett's "Roman Candles" passage. toolsforinvestors.com