AI Did Not Kill Enterprise Software. It Split It Into Structural Winners and Losers
There is a narrative that has dominated boardroom conversations and venture capital discussions for the past two years: artificial intelligence will devour enterprise software in the same way that software devoured analog business models. It is a powerful image. And like every powerful image that circulates without friction, it deserves to have pressure applied to it before it dictates investment decisions with real consequences.
Charlie Gottdiener, CEO of Anaplan, recently published an essay in Fortune that proposes a different reading. His thesis is not that AI will not change software. It is that the change will not be horizontal or democratic: it will be a selection, a sorting process that will amplify certain vendors while rendering others redundant. For Gottdiener, the decisive variable is not the technology itself, but the nature of what each layer of software calculates or represents.
What follows is neither a defense of that thesis nor a refutation. It is an audit of its commercial logic.
---
The Three-Layer Model and What Is Actually at Stake
Gottdiener proposes that the enterprise software architecture is fracturing into three levels with distinct roles. At the top, large-scale language models act as a universal conversational interface. At the bottom, what he calls the Model Context Protocol executes commands toward existing systems. In the middle, the layer he calls Deterministic Domain Authority — his term for governed, auditable, and reproducible computation engines — is where, according to his argument, defensible value will reside.
The technical distinction that sustains that argument is precise: a language model is probabilistic. It generates responses that vary according to statistical pattern, not according to a fixed computation logic. When a company needs to know the exact impact of a modification to its financial plan, or to calculate the effect of a change in compensation structure on total labor costs, probability is not enough. What is needed is an engine that produces the same result given the same data, every time, under any condition of regulatory or fiscal audit.
That limitation of language models is real, documented, and not in serious technical dispute. What is in dispute is what happens to the layers of software that do not live in that deterministic space.
Gottdiener's diagnosis of business intelligence and data visualization tools is specific and carries weight: if the primary value of a product was to allow a user to ask questions about their data in natural language and receive a visual response, that product now competes against a conversational interface integrated into the working operating system. Not in three years. Today. The barrier to entry for replicating that basic functionality has collapsed.
The same applies, with nuance, to workflow automation tools that do not possess their own computation: they move data between systems, but they are not the source of truth for any of them. When a language model can orchestrate those integrations directly through natural language instructions, the intermediate layer loses its argument for existence.
Where the analysis becomes more interesting, and more susceptible to scrutiny, is in the other half of the classification.
---
The Moat That Gottdiener Proposes and What He Does Not Say About It
Gottdiener argues that enterprise planning engines, human resources record systems, customer management systems, and specialized regulatory databases are the structural survivors. The reason: they possess governed computational truth. The hire date of an employee, the amount of a closed deal, the maximum permitted dose of a pharmaceutical compound. Those are facts, not suggestions. And a language model cannot fabricate or validate them with the precision that an audit demands.
That argument has technical solidity. But it introduces a trap that Gottdiener partially acknowledges and then does not fully develop.
If the value of a deterministic engine is computational precision within a domain, and if that engine can be replicated by another vendor offering the same precision at lower cost, then the language model — acting as a universal interface — will be indifferent between providers. The moat is not in the engine. It is in the specific model that the customer has built within that engine.
Gottdiener says as much: the real advantage is the planning or operations model that a particular company has codified into the platform over years. Migrating that model to a competing system is not a data export. It is rebuilding institutional logic from scratch. That is painful, and that pain is what retains the customer.
This is where it is worth separating the narrative from the commercial finding. Because what Gottdiener is describing, without naming it as such, is a retention mechanism based on accumulated switching costs, not on continuous technical superiority. It is an argument about stickiness, not about permanent innovation. That does not invalidate it. But it radically changes how the value proposition must be read for a customer who has not yet begun implementation.
A buyer who is not yet tied to any platform must ask — and here Gottdiener's analysis does not help them very much — how much of the value they will receive comes from the engine itself, how much comes from the depth of their own model built over time, and how much comes from integration with the language model ecosystem that will follow. Those are three value propositions with radically different cost and retention structures.
---
The Variable That Does Not Appear in the Visible Argument
Gottdiener writes from a position of conflict of interest that he declares with honesty: Anaplan is, according to his own description, exactly the type of platform that his theoretical framework declares a winner. That does not disqualify the analysis, but it compels the reader to pay closer attention to what is not said.
What does not appear in the text is the pricing dynamic in a market where deterministic engines multiply. If the language model acts as a neutral interface and selects the most precise engine at the lowest cost, the price of deterministic engines will tend to fall as more vendors offer them. Competition will not disappear because the computation is precise. It will shift to another level: who offers the same precision with better performance and lower initial implementation cost.
In that scenario, the only durable defense is not the engine, but the depth of the customer's institutional model. Which means that the vendor's value becomes progressively concentrated in the implementation and model-building phase, not in the software license itself. That has direct consequences for margins and for revenue structure: if the value lies in professional services and in the complexity of the model that has been built, the business increasingly resembles a consultancy with its own platform rather than a software product with high margins and growth through license expansion.
That shift is not necessarily bad. But it is also not the narrative that circulates when people speak of SaaS platforms with pure-software valuations.
The other notable absence is the speed of improvement of language models themselves in mathematical and logical reasoning tasks. Gottdiener assumes that the probabilistic limitation of language models is structural and permanent for complex enterprise computation. That assumption may be valid today. It may not be in four years. The improvements in formal reasoning of new-generation models are consistent and well documented. If that gap closes partially, the dependency on external deterministic engines is reduced, and with it, the central premise of the argument.
---
What Separates a Structural Moat From a Self-Confirming Narrative
Gottdiener's argument is not smoke. It identifies a real technical distinction — deterministic computation versus probabilistic inference — and connects it correctly with a concrete business need: auditability, regulatory precision, and reproducibility in high-impact financial and operational decisions.
Where the argument requires more work is not in its technical premise, but in its projected commercial architecture. Retention through accumulated switching costs is powerful, but it operates on companies that are already implemented. The value proposition for a new buyer is more fragile than the framework suggests, because that buyer can choose to implement with greater caution, with less initial model depth, and with more exit options. Pricing pressure on deterministic engines will increase as the category matures. And the continuous improvement of language models in complex reasoning will keep narrowing the space where delegation to an external engine is mandatory.
What does remain clear, and deserves to be taken without reservation, is the diagnosis about the layers of software based primarily on user experience and visualization. Those layers do not have a retention argument based on their own computation. Their moat was the interface, and the interface already has a more convenient substitute. It is not that they will disappear tomorrow. It is that their capacity to sustain pricing and retention without transforming their value proposition toward governed computation or data deteriorates with every passing quarter.
AI is not consuming enterprise software in a uniform way. It is applying asymmetric pressure that favors those who possess governed computation and penalizes those who were primarily selling convenient access to data that others calculated. That separation was not invented by Gottdiener. It was accelerated by AI. And the companies that have not yet audited which side of that line their current vendor lives on will need to do so before renewing their next contract.
