The AI software market is sorting into five floors, from raw inference up to agent-native data. Naming them is easy. Holding one is not. Each floor has exactly one question that decides whether you can actually hold a position there — and most teams answer for the floor they wish they were on, not the one they're standing on.
The AI software market is sorting into five floors, from raw inference up to agent-native data. I mapped the whole thing — the named players on every floor and the forces that could stall the shift — in the market map.
Naming the floors is easy. Holding one is not.
Here is the test. Each floor has one question that decides whether you can actually hold a position there. Most teams answer for the floor they wish they were on, not the one they are standing on.
One decision question per floor
Inference & Infrastructure
Do you control compute or distribution? Control neither and you're renting your position from someone who does.
Agent Data Platforms
When you open your data to agents, do switching costs go up or down? If wiring agents into your schema doesn't make you harder to leave, you're a thin app on someone else's data.
Context Engines
For a real workflow, do you own all five context steps — curation, synthesis, consolidation, prioritization, storage — or do you own one and label the other four? This floor goes to whoever does all five, not whoever names it first.
Orchestration
Do you own the record of what every agent did — the one a compliance team can audit? Here, governance is the moat, not the feature set.
Agent-Native Data
Is your data proprietary and years deep, or are you orchestrating data you rent from the floor's incumbents? Workflow wins deals. Data wins the war.
One test underneath all five
The pattern underneath every floor is the same. You hold a floor only if you own something the floor above cannot absorb and the floor below cannot turn into a commodity. The question just looks different depending on where you stand.
You hold a floor only if you own something the floor above can't absorb and the floor below can't turn into a commodity.
Floor 3 is the one I spend my time on, so I'll be blunt about it. "We have a context layer" is the most over-claimed sentence in the category right now. The honest version of the Floor 3 question is whether you run all five steps of the context loop for a real workflow, or whether you do retrieval well and call the other four steps "context." The floor goes to whoever does all five. Naming it first buys you a press cycle, not a position.
Answer for the floor you're on
The framework only works if you answer for your actual floor, not your aspirational one. The thin app answers the Floor 2 question as though it owns the data. The retrieval vendor answers the Floor 3 question as though it does all five steps. The wrapper answers the Floor 1 question as though distribution is a roadmap item rather than a thing you either have or don't.
That self-deception is the expensive part. The market will answer the question for you eventually — in pricing power, in who gets bundled, in who gets repriced as commodity infrastructure. Better to answer it honestly now, while you can still act on the answer.
Of these five, which question is the hardest to answer honestly about your own floor?
That difficulty is the signal. The floor whose question makes you flinch is usually the one you're standing on without a defensible position.
Bring that one question to your next planning meeting. Not a feature list. The question for your floor, asked honestly. If you can answer it, you've picked your floor. If you can't, the market is about to pick it for you.
Frequently Asked Questions
What are the five floors of the AI software stack?
From the bottom up: Floor 1, inference and infrastructure; Floor 2, agent data platforms; Floor 3, context engines; Floor 4, orchestration; Floor 5, agent-native data. Each floor is a distinct position in the stack with a different basis for defensibility.
How do you know if you can defend a position on your floor?
Each floor has one decisive question. Floor 1: do you control compute or distribution? Floor 2: does opening your data to agents raise switching costs? Floor 3: do you own all five context steps or one labeled as five? Floor 4: do you own the audit log? Floor 5: is your data proprietary and years deep, or rented? The unifying test: you hold a floor only if you own something the floor above can't absorb and the floor below can't commoditize.
Why do most teams misjudge their floor?
Because they answer the defensibility question for the floor they wish they were on, not the one they're standing on. A thin app on someone else's data platform answers as if it owns the data; a context vendor that does one step well answers as if it does all five. The question only helps if you answer it honestly about your actual position.