Harshil Shah is VP of Product and Delivery at AltaDX.gettyA year ago, getting your hands on a frontier model felt like an edge. Today, you and your competitor call the same API, and a better one ships every few months. I've sat in plenty of build-versus-buy debates where the team wanted to own the whole orchestration stack, and most of the time, my answer is the same: not yet. Time to market matters more than owning a layer that's commoditizing by the quarter. I don't lose sleep over which model is under the hood. I lose sleep over whether the thing really works in production.So, what's left? After shipping a fair number of agents into the real world, I keep landing on the same three things: data, domain and distribution. You need all three. Miss one and you don't have a business—you have a demo.Data: The Loop, Not The LakeEvery founder claims a data moat. Most have a data lake, which isn't the same thing. The cleanest cautionary tale is Bloomberg, which had decades of proprietary financial data and reportedly spent around $10 million training a finance-specific model. Then, a general frontier model with no special training matched or beat it on finance tasks. A pile of historical data a competitor can approximate isn't a fortress.What's defensible is the loop: data your product generates through use, that feeds back into how the system behaves and that a competitor starting today doesn't have yet.Most enterprise teams get this part backwards. The real ask I hear is rarely "Give me a chatbot." It's "Extract this, decide that and keep a human in the loop until I trust it." Every one of those review steps is a labeled data point your competitor never sees. That feedback is the moat, not the storage bill. If your data doesn't make the next decision better, it's a warehouse.Domain: It Has To Be Built In, Not Talked AboutThis is the one that's changed the most, and not in founders' favor. Building vertical software used to require a rare mix of deep industry knowledge and serious engineering. LLMs collapsed the engineering half. The expert with 20 years in a field can now encode their methodology directly, so a category can go from three incumbents to 300 startups in a year. "We understand this industry" is a starting condition now, not an advantage.Domain only counts when it's wired into the boring parts: the edge cases, the compliance and the definition of what a good outcome is. In regulated consumer messaging, a 17-year-old can't legally opt in to marketing texts, and a general model has no idea. An agent that charges for a missed appointment has to wait before it bills because a human will correct the record after the fact. These things decide whether the agent can be trusted to act on its own.The specialists that endure live in exactly these high-stakes, rule-bound corners. Harvey in legal and Hippocratic AI in healthcare didn't win on a secret model. Harvey runs multimodel, which tells you the model was never the point. They lead because the standards of the work are encoded into how the system behaves and because trust in those fields takes years to earn and one bad answer to lose.Distribution: The KingmakerThis is the one technical founders most want to ignore. The uncomfortable fact is that distribution beats algorithms, because algorithms are replicable and reach isn't. The minute a clever feature works, an incumbent folds it into a product already in front of millions and ships it in an update.Distribution today means being embedded where the work already happens, with no rip-and-replace. Harvey is the textbook case of getting the order right: Secure the domain, the data and the law-firm relationships first, then extend reach by becoming available inside Claude through a connector, so the product meets lawyers where they already work. The twist makes it sharper. Anthropic also launched its own legal tooling, moving up into Harvey's lane. Rather than retreat, Harvey integrated deeper, betting that life inside the ecosystem beats life outside it.Outcomes and distribution are colliding in how agents get sold, too. Sierra charges only when its agent resolves the customer's issue, and it's reportedly inside more than 40% of the Fortune 50. Enterprises have stopped paying for pilots and started paying for results.Why You Need All ThreeEach of these is weak alone. Data you can't get in front of anyone is a better mousetrap nobody installs. Reach without real domain depth gets you a shallow feature that a foundation model copies next quarter. And all the industry expertise in the world never compounds if it isn't learning from usage. They only work as a loop. Reach earns usage, usage generates data, data and domain produce outcomes a generic tool can't match and those outcomes buy trust and more reach.It's why selling an agent has gotten unforgiving. The moment you charge for outcomes, the agent has to deliver one, and it can't do that without data to improve on, domain to be trusted with and a place in the workflow to act. It's also why I'm allergic to demos. An agent that looks brilliant in a scripted run and falls apart on a live customer has none of the three working in production. Polish isn't a moat.The Gut CheckBefore you build, buy or invest, run the test honestly. Does the data compound with use or just pile up? Is the domain knowledge built into the product or sitting on a slide? And are you embedded where the work happens or waiting for people to find you? One weak answer is a real problem.The model will keep getting cheaper and better for everyone. Plan as if it’s free, then ask what you have that isn't.​​​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?