The Agent-Native Stack

In these seven pieces of the series on ‘Enterprise SaaS’, the picture of where enterprise software is heading has come into focus. Following is a TLDR of thesis presented:

• The unbundling cycle is real and is reorganising the industry around the agent as the new primitive.

• Margins are compressing for any business that does not rebuild its pricing around usage and outcomes.

• The data moat was always a myth and the real moats, workflow entrenchment, distribution, compliance, are uneven across incumbents and challengers.

• Enterprise software is separating into three distinct layers, record, action, and judgment, each with its own timeline and its own set of winners.

• The plumbing layer for agent identity is unsolved and worth a category.

• Vertical depth is more valuable than ever.

• And the twenty-year buy-not-build orthodoxy is inverting, selectively, in a way that will change how CIOs allocate capital for the next decade.

Each of these pieces describes a force. The question for this final piece is what all the forces produce when they compose. What does the resulting enterprise software stack actually look like. Who wins in which segments. Where should builders, buyers, and investors be paying attention.

The honest answer is that the post-SaaS stack is not a single thing. It is four distinct archetypes (of agent-native software), each with its own logic, its own go-to-market, its own unit economics, and its own set of probable winners.

The first archetype is the least glamorous and the most under-appreciated.

These are the systems that hold canonical data for an enterprise, with all the conservative properties that implies (the general ledger, the customer master, the electronic health record, the asset master, the core banking platform etc.) These systems have existed, in some form, for decades, and they will continue to exist, in updated forms, for decades more.

AI does not unseat this archetype. It does not even significantly disrupt it.

What AI does is allow a new generation of these systems to be built with modern architectures, modern user experiences, and native agent-action capability, for industries where the incumbents have aged poorly. The opportunity is for rebuild, not for displacement of the category.

Who wins. The incumbents in most established categories will keep winning, because the switching cost is too high for most buyers to justify. SAP in enterprise ERP. Oracle in databases. Workday in HR. Epic in health records. The interesting territory is in industries where the incumbents have under-invested or where the regulatory ground has shifted. These are the categories where new systems of record can be built and adopted in greenfield or replacement circumstances.

How to spot the winners. They speak the language of durability, auditability, regulatory alignment, and disaster recovery. They have weak marketing and strong reference customers. They are boring in a way that pattern-matches badly to the current hype cycle, which is why the market has tended to underestimate them in every cycle for twenty years and continues to.

The second archetype is the most hyped, because it is the easiest to describe.

These are the horizontal AI platforms that aspire to be present everywhere in the enterprise (Microsoft Copilot, Google’s equivalent, OpenAI’s enterprise offerings, Anthropic’s, Salesforce Einstein etc.)

All these hyperscalers’ AI platforms are betting that an enterprise-wide assistant, deeply integrated with a broad productivity surface, becomes the default interaction layer for knowledge work, and that everything else becomes a plugin into this layer.

The archetype will be very large in revenue terms. Horizontal assistants will be used by hundreds of millions of enterprise users, priced per seat, and will become a significant portion of the enterprise’s software spend. The question is whether they will be as dominant, across the full stack, as the marketing suggests.

Probably not. The horizontal platform will be excellent for general-purpose knowledge work, and less good for the specialised workflows that sit deeper in the enterprise. It will be a premium layer of productivity, not a replacement for the vertical and action-layer systems that do the actual work of the business. Buyers will pay for it in addition to, not instead of, their other systems. This is still a large outcome, but it is smaller than the total displacement narrative.

Who wins. Microsoft has a near-unassailable distribution advantage in enterprise productivity, and Copilot will almost certainly be the horizontal platform of choice for most large enterprises. Google will win in some segments, particularly education and certain verticals. The independent AI labs will win in specific categories where their model capability is materially ahead, but their distribution challenge is significant. Salesforce and ServiceNow will win horizontal-assistant positions within their existing customer bases, and will compete with the hyperscalers at the edges of those positions.

The third archetype is the one I think is most under-appreciated at this moment in the cycle.

These are AI-native applications that go deep into a specific industry or function, build or buy the domain vocabulary, accumulate labelled examples, design evaluations, integrate with systems of record, and align with regulatory regimes. They compete directly with traditional vertical SaaS, with consulting hours, and with offshore back offices. They produce measurable outcomes inside specific workflows. They are priced on usage or outcomes rather than seats.

This archetype is harder to see clearly, because each specific example is small, and the category sprawls across many industries. But in aggregate it is likely to be the largest source of new enterprise software value over the next decade, and the best outcomes per dollar invested. The winners in this archetype will look, in ten years, like the enterprise software companies of the 2030s in the way that Salesforce, Workday, and ServiceNow looked like the enterprise software companies of the 2010s.

Who wins. In each vertical, typically, one to three dominant players within five years of category formation. The dynamics favour focused founding teams with deep domain expertise, substantial initial capital to fund the domain engineering, and disciplined commercial models that price on outcomes from the start. Large horizontal players rarely win here, because they cannot match the depth without fragmenting their organisation into vertical business units, which they resist. Consulting firms can compete, up to a point, but usually lose to focused product companies over time.

How to spot the serious players. They have a clear, defensible view of their vertical. They have specific named customers in that vertical. They have evaluations that measure domain-specific performance, not general benchmarks. They have named regulatory experts on their team. They have honest pricing tied to outcomes. They do not over-index on model capability in their marketing, because they know the model is a component, not the product.

The fourth archetype is the most invisible and, for certain investors and certain builders, the most interesting.

These are companies whose product is not an application at all. They provide the infrastructure that other agents consume. Identity and access for agents. Orchestration across multiple agents. Evaluation and monitoring of agent behaviour. Memory management. Tool registries. Retrieval and context management. Compliance and audit scaffolding. Policy enforcement. Simulation environments. None of this is glamorous. All of it is load-bearing for the agent economy.

This archetype is the direct equivalent of cloud infrastructure in the 2010s. It is invisible to end users but is critical to everything that end users interact with. It has strong network effects as adoption compounds. It is mis-priced early, because the buyer is other software companies rather than end customers, and the market for it is not yet legible. The winners ten years out will include a few public companies with extremely durable positions.

Who wins. Unclear at this point, which is part of what makes the archetype interesting. Some of the winners will be hyperscalers extending into this layer (e.g., Microsoft recently launched Agent 365). Some will be specialist companies founded in the last two to three years. Some will be acquired into larger platforms before they become independently visible. The opportunity for careful investment here is significant, because the category has not yet attracted the attention or valuation premium it eventually will.

The four archetypes are not competitors. They are layers that will work together, and the enterprise of the 2030s will typically use all four.

The system of record keeps its seat. The horizontal agent platform becomes the general-purpose productivity surface for knowledge work. The vertical agent operators run the deep, high-value workflows in the specific industries where the enterprise operates. The agent-native infrastructure is consumed by the other three archetypes, invisibly but essentially.

The mistake is to treat the four archetypes as alternatives, or as phases. They are complements. A vertical AI operator in insurance will read data from the insurer’s system of record, be supervised by users working through their horizontal assistant, and run on top of agent-native infrastructure for identity, orchestration, and evaluation. Each layer does what it is good at. None of them replaces the others.

This is why the “SaaS is dying” and “horizontal AI will flatten everything” narratives are both wrong in the same way. They mistake one archetype for the whole stack. The stack has always had multiple layers. The agent-native stack will continue to have multiple layers. The layers are shifting, rearranging, and in some cases being added. None is being eliminated.

For builders, the map is a forcing function. Pick an archetype. Build for that archetype. Do not try to span two. The teams, commercial models, capital requirements, and strategic positions for each archetype are different enough that trying to be both a horizontal platform and a vertical operator, or both a system of record and an infrastructure layer, leads to confused execution and unattractive outcomes. The best companies are clear about which archetype they belong to and optimise ruthlessly for that archetype’s dynamics.

For buyers, the map is a procurement framework. Each archetype needs a different set of procurement questions, a different commercial structure, a different integration approach, and a different governance model. Buying a vertical operator as if it were a horizontal assistant, or an infrastructure layer as if it were an application, leads to disappointment on both sides. Enterprises that build procurement functions that understand the archetypes will get significantly more value out of their AI spend than those that do not.

For investors, the map is a portfolio lens. Each archetype will produce winners and losers. The risk profile, the capital efficiency, and the expected holding period for a winner in each archetype are different. An investor whose portfolio is entirely in one archetype is making a concentrated bet that may or may not be intentional. A disciplined portfolio across the archetypes produces a more durable set of outcomes.

For operators thinking about careers, the map is a guide to where the interesting work will be. The system of record archetype rewards patience, durability, and depth. The horizontal platform rewards scale and platform thinking. The vertical operator rewards domain expertise and commercial discipline. The infrastructure layer rewards technical depth and operational maturity. People gravitate to the archetype that matches their disposition and their strengths. Getting the fit right matters for a decade of career work.

There is no single post-SaaS paradigm. Anyone who tells you there is has not spent enough time looking at the actual enterprise stack. The agent-native stack is multiple things at once, and it will be so for the foreseeable future.

What is changing is not the number of layers, but the logic of each layer, the primitives around which each is organised, and the commercial dynamics that determine winners in each. The term SaaS, as a label, will mean different things in different layers, and the vendors most successful in each layer will increasingly not share a common business model with the vendors in the other layers. The industry will gradually stop using the single label SaaS altogether, because the label no longer describes anything coherent.

That is what the end of an era actually looks like. Not a cliff-edge disruption but a gradual fragmentation of a category that used to be unified, into several distinct categories with their own rules, their own winners, and their own stories. Enterprise software is not ending. It is getting more interesting, more specialised, and more valuable in aggregate than ever before. The next decade will produce more enterprise software category winners than the previous two combined, and the winners will be unusually spread across the archetypes described here. The best way to watch this will be to recognise the archetypes, to read each company against its archetype rather than against a generic playbook, and to be patient about the outcomes. The interesting time is just beginning.