The Smartest AI Model Yet Still Needs Something Real to Work In

Natascha
CMO, Yolm
Nov 17, 2026

Claude Fable 5.2 can build software with a thousand agents. Here is why it still cannot fix a business that runs on several partial truths.

Eight days, a thousand agents, one empty folder

On November 9, 2026, Anthropic released Claude Fable 5.2. In one internal experiment, an Anthropic engineer gave it an empty folder and a single goal: build an operating system. The model coordinated more than 1,000 sub-agents over eight days and produced software that booted on a laptop.

The release came with a 20% price cut and output up to 30% faster than the previous version, as Computerworld reported. Read that as a business leader and the question is obvious. If AI can build an operating system from nothing, why is it still so hard to get AI to answer a simple question about my own business?

The answer has very little to do with the model.

What is your AI standing on?

A customer calls to ask why an order is late. The scheduling system says it is complete. The spreadsheet called FINAL_v7.xlsx, the one that tracks rush-order exceptions, says it is waiting for inspection. The dashboard still shows yesterday. Nobody made a mistake. Each person is looking at a different copy of the business.

People hold those copies together with experience, phone calls, and asking Dave. AI cannot. Point the most capable model in the world at that business, and it has to guess which version is real. A smarter model guesses more fluently. It still guesses.

The analysts reacting to the Fable 5.2 launch saw the same limit. Omdia chief analyst Lian Jye Su told Computerworld that "reliability remains an issue without human supervision." Pareekh Jain of Pareekh Consulting noted that complex legacy systems and real business requirements could limit how much work companies can hand to AI agents.

The research points the same way. MIT-affiliated Project NANDA, in its July 2025 preliminary report, found a sharp divide: about 95% of organizations in its studied integrated GenAI efforts showed no measurable P&L impact, while about 5% extracted significant value.

So here is our view. AI doesn't fail at intelligence. It fails at friction, and the biggest friction is the business underneath.

Your AI pilot did not fail. Its foundation did.

One more copy is not a foundation

The usual fix is to gather everything into one more place: a warehouse, a lakehouse, a "context layer" for the AI. Those tools have a real job. They are good for analyzing history and for combining genuinely separate companies.

But a copy is refreshed after the business has already moved. It rarely sees FINAL_v7.xlsx. And it can describe the business, but it cannot run it. Ask it to reschedule an order, and it has nothing to reschedule.

Left: Assembled Software, where sales, planning, production, shipping, finance, reporting, a spreadsheet, a reporting copy, and AI each hold their own copy of the data. Right: your business system, where the same applications and AI all work on one live database.

Every extra copy is one more version to reconcile, and the AI still has to pick one. AI does not need another copy of the business. It needs the live business.

When AI writes more than anyone can review

Go back to the thousand agents. Writing software is no longer the hard part. AI will soon produce more of it than any team can read line by line. The constraint moves from generation to verification: how do you know a change fits everything else the business depends on?

In a business spread across separate systems, nobody can answer that quickly. Every change touches a seam, and the seams are where things break silently. A better model writes the change faster. It does not make the seams safer.

The answer is a system that can check itself. When the whole business is described in one system definition, a compiler can check a change against all of it in seconds, before it goes live. That does not prove the change is what the business intended, and people stay accountable for that. But it does catch the inconsistencies that used to surface as a late order and an angry phone call.

Something real to work in

On Yolm, the whole business runs as one system on one live database. That does not mean one giant application, or everyone seeing everything. Quoting, scheduling, shipping, and finance each keep their own role-specific applications, standard and unique alike. Each person, application, and AI agent sees only what its permissions allow. They all work from the same live records.

That changes what AI can do in three ways.

  • It can read the business. Not yesterday's extract, but the current operation, including the parts that used to live in spreadsheets. The data is consistent by construction: two parts of the system cannot disagree about which quantity is real, because there is only one.
  • It can act safely. The AI logs in like an employee, not a superuser. It has the same permissions, approvals, and audit trail as the people it works beside. That is the human oversight the analysts are asking for, built into the system instead of bolted on.
  • It can help build. Yolm Forge is where AI agents and people build and change the system, and every change is checked against the whole system definition before release. Yolm Spark is AI that answers questions and does work inside the live operation.

Yolm Core holds the business. Forge builds and changes it. Spark works in it. None of them keeps a separate copy.

You control your system: what it does, what changes, and when changes go live. Yolm runs the foundation underneath it.

The test to run before the next model arrives

We call this the AI-native business system: your own business system, standard where the business is standard and unique where it is unique, on one live truth and one managed foundation. Yolm is the AI business platform where a business builds and runs it.

There will be a Fable 5.3, and a model after that. Each one will be smarter than the last. None of them will fix a business that exists as several partial truths.

So before you evaluate the next model, ask one question about your own business:

If two parts of your system can disagree about a number, which one does the AI believe?

If you do not like the answer, the next step does not have to be big. Start with one capability where the pain is sharpest, live in weeks, and let every next step join the same live system. The first capability can be small. The foundation must be complete.

Talk to us about your first capability · See a real answer from Yolm Spark