The world is changing and everyone in software is trying to read the same thing: where this goes. It splits into two questions, how software gets built and what the software does once it is built. This is an attempt to reason through both, written by two founders who each work one half. We are honest that the answer might not be us. What we think is that it will look like this, because the problem leaves few other shapes.

Software is now expected to be intelligent, and most applications are missing the part that would make them so: a predictive core that knows their own data. Aito v2, generally available today, is the predictive database built to be that core. It speaks SQL to the systems you already run, holds vectors and relationships for your LLMs and agents, and our own applications run on it.

The first wave of agentic AI tried to build autonomous software out of language models alone. That was the right first move for language and the wrong design for action. Software that acts needs reasoning for the exceptions and intuition for the rule, working as peers, and it needs both to sit on one foundation. Here is the paradigm, the architecture it implies, and how Aito runs it without a Frankenstein stack.

Written in August 2026, when v2 was in public beta. v2 became generally available on 21 September 2026, so this post is the record of the beta rather than the current state. It is concrete on purpose: here is what the engine does, here are queries you can run, and here are the numbers, including the ones that do not flatter us.

The efficient market hypothesis says you cannot beat the market. Graham, Buffett, and a lineage of growth investors spent decades saying otherwise. So I ran the experiment: grade 250 S&P 500 companies on value, quality, and growth signals, point-in-time, and let a predictive database score all three philosophies against twelve years of real outcomes. The answer surprised the value investor who built it.

A new architecture for B2B SaaS

Buyers ask whether a predictive database can hold up at SaaS scale. Here are the numbers from a 10-million-row invoice routing benchmark, run end-to-end through Aito's HTTP API the way production traffic actually hits it: low-hundreds-of-milliseconds predict latency, sub-linear scaling from 1k to 10M, and what the cold-start looks like before the cache warms.

A predictive database lets you query predictions (classifications, recommendations, missing values) from structured data, using the same query interface you use for data retrieval. No training step.

Databases are absorbing AI capabilities. A taxonomy of the four categories emerging in 2026: vector databases, ML-in-database platforms, LLM-augmented databases, and predictive databases.