Loading
Loading
For twenty years, every application was built the same way: a database, some business rules, a screen on top. AI is forcing us to re-open that settlement. This is the decoder for what the three layers are, what agents actually change, and the question most people skip: which parts of the “stable bedrock” were only ever shaped that way because a human was driving.
Strip any piece of software you use down to its skeleton and you find the same three layers, stacked. They have not changed in two decades.
The data model is the nouns: the structured tables that define what a “customer,” an “invoice,” a “trip” even is, and how they link. The business logic is the verbs and the math: the rules that process the data (a discount calculation, a matching algorithm, a payout split). The UI is the skin: the buttons, forms, and dashboards a human taps to trigger the logic and change the data.
The old SaaS business was the act of bundling all three. If you wanted to manage customers you bought Salesforce, and you got Salesforce's database, Salesforce's rules, and Salesforce's screens as one inseparable package.
The layers are abstract until you map a real product onto them. Take a ride.
Notice the asymmetry already. The data model and the logic must be rigid: a trip links exactly one rider to one driver, the payout is exactly fare minus the platform fee, every time. You do not want a creative system inventing a new way to store who owes whom money. The UI, by contrast, is the soft part, the part that exists purely to translate a human's intent into a command the rigid parts can execute.
Satya Nadella's framing, from the Microsoft Build 2026 podcast crossover, is that AI does not rewrite the stack so much as pull it apart. Each layer gets a different verdict.
The bedrock holds. In his words, “my general ledger better be a general ledger. I don't need new schema creation… that entity relationship is actually a pretty good, robust thing.” The data models and the core financial rules of the world are stable and should not be reinvented. You do not let a model hallucinate a new way to store debits and credits.
The UI dissolves. The rigid screen was always a bottleneck: a human had to log in, navigate menus, and click to translate intent into data. With agents, that step disappears. You say “get me to JFK by 3pm” and an agent calls the booking logic directly, returning a minimal confirmation instead of an app to operate. Nadella calls the shift “chat to canvas.”
The apps go headless. Once the UI is no longer the only door, software stops being siloed. An agent reads your calendar, notices a flight delay in the airline's logic, pushes back the Uber, and texts your spouse, treating four apps as four rule-engines it orchestrates. You stop being the human bridge clicking between tabs.
Here is where the tidy version of the framework gets interesting, and where it is worth pushing past Nadella's reassurance.
“Don't touch the data model and the logic” is right for the genuine bedrock, the general ledger really should stay a general ledger. But a lot of what looks like bedrock was actually shaped by a constraint that is about to lift: a human was the operator. Pagination exists because a person scrolls a page at a time. Multi-step wizards and confirmation dialogs exist because a person needs to be walked through a task and double-check before committing. Data was flattened into the exact shape a form could display. None of those are laws of the business. They are accommodations for human hands and human attention.
When the operator becomes an agent with tool use, those accommodations can become costs. A pagination scheme that felt natural to a scrolling human forces an agent into fifty round-trips to read what it could have pulled in one. A confirmation gate built for a cautious human becomes a wall an automated workflow has to route around.
The hard, valuable judgement is not “preserve everything” or “rewrite everything.” It is telling the two apart: which structure is a genuine business invariant that must be preserved (the payout must equal fare minus fee), and which is a human-era artifact that an agent operator should collapse, while keeping the invariant intact. Preserve the rule; redesign the mechanism the rule was wrapped in for a human.
If code is cheap to generate and the UI is dynamic, the old moat, “we own the software,” evaporates. When anyone can stand up an interface in minutes, value moves to two places that do not commoditize.
Context (Nadella's “Work IQ”). Not the software, but the proprietary record of how a specific business actually operates: its history, its private data, its institutional knowledge. That is what cannot be regenerated by a model that just met you.
Private evals and traces. The differentiator is no longer which frontier model you use, since everyone rents the same ones. It is your private evaluation suites and the history of agent runs (the traces) that encode what a “correct” outcome looks like for your business. The model is rented; the definition of done is owned. This is the same conclusion Pere keeps reaching from the other direction: durable value sits with the scarce complement and the customer relationship, not the undifferentiated middle. See where value accrues in a stack and the substitute price floor.
The one-screen summary, by layer:
The era of logging into a dozen apps to click buttons is giving way to agents that navigate stable data models on your behalf. The data and the genuine rules are the bedrock. The interface is gone. And the most underrated move in between is auditing which “rules” were only ever rules because a person was holding the mouse.