AI Agents Don't Multiply Analysts. The Right Data Does.

On June 26, Anthropic published the latest edition of its Economic Index, and one line in it should matter to anyone who runs an analysis team. "One year ago, most Claude usage took the form of a conversation between a user and an assistant. With the rapid growth of Claude Code and Cowork, Claude sessions now increasingly consist of long-running agentic tasks." The shape of knowledge work changed in twelve months. We stopped asking models questions and started handing them jobs.

The numbers around that shift are not subtle. Across roughly 9,700 surveyed users, 86% reported that AI made them faster. Sixty-eight percent said they learned more when they used it. A majority said it made their skills more valuable, not less, and more than a third expected AI to handle most of their work within a year. The people delegating the most were the most optimistic about their own careers, not the least. This is the productivity story everyone now expects: point an agent at the work, get back in an hour what used to take a week.

But there is a second reading of the same data, and it is the one that decides whether your firm sees any of those gains.

Intelligence is now abundant. Information is not.

The report is consistent on one point: outcomes are uneven. Experienced users succeed far more often than newcomers, even on the same task. Anthropic's March edition put a number on it: long-tenured users had roughly a 10% higher success rate, and the gap held even after controlling for what they were working on. The model is the same for everyone. The results are not.

That gap is easy to misread as skill. Some of it is. But most of what separates a good agent run from a useless one is not the prompt and not the model. It is what the agent can reach. An analyst who has spent a year learning where the numbers live, which figure is authoritative, and how one fund's disclosure maps to another's is feeding the model better information. The newcomer is feeding it a guess.

This is the part of the AI story that gets undersold. We talk about it as a revolution in intelligence. It is just as much a revolution in information. A frontier model pointed at a folder of inconsistent PDFs returns confident, well-written, wrong answers. The same model pointed at structured, reconciled, source-linked data returns work you can put in front of an investment committee. The intelligence is identical. The information is not. Until your data is structured, you will be disappointed by the results, whichever model you run.

In real estate, the data was never the easy part

Our industry is the hard version of this problem. The information that matters lives in rent rolls, valuation reports, fund factsheets, data rooms, and twenty years of PDFs that never agreed on a format. The same building appears under three names. The same fund reports NAV two different ways. An agent dropped into that returns the mess it was handed.

So before we built anything that looks like an analyst, we built the layer underneath it. The Quanthome Data Engine takes the information scattered across your documents and structures it into one secure, audited base, entity-matched against the market and ready for agents to use. From scattered documents to one structured, audited view. Every figure stays traceable to the record and the document it came from. That last part is not a detail. A number an agent cannot trace is a number a reviewer cannot trust.

This is the unglamorous work that makes the rest possible. Structure the data first, and the productivity figures from the Economic Index stop being someone else's story.

What changes once the data is ready

Inside Quanthome, our own analysts already work this way. Quanthome AI runs managed agents in parallel on different projects, in one interface connected to our data, the client's, and any source they approve. You ask a question in real estate terms and it generates the page for it, grounded in CHF 5.8T+ of indexed asset value and 140+ vehicles, every figure linked back to its source. It changes how we reach information. The dashboard is no longer something we wait for a team to build. It is something the question produces.

Next week we put the next layer in front of a first group of clients. Quanthome Workflows enters beta preview: managed agents that turn a structured base into the deliverables the business actually runs on, the annual report, the factsheet, the due-diligence pack, generated on templates you control. The line we keep coming back to is a simple one. A high-quality, sourced report analysis in two prompts: one to frame the question, one to refine. Clients measuring it see roughly 90% less time to a fund's annual report and 95% less on due diligence, with every figure still tracing back to its source. A late correction reflows everywhere it touches instead of restarting the cycle.

Two prompts is not a trick of a faster model. It is the payoff of having structured the data first.

The model is not your advantage

Here is the uncomfortable part of the Anthropic data for anyone betting on intelligence alone: everyone has the same models. Your competitor can buy the exact capability you can, on the same day, at the same price. What they cannot buy is your data, structured and ready for an agent to work. As my colleague Nathan wrote about what the models find when they arrive, the constraint was never the model. It is the environment you point it at.

The shift the Economic Index describes is real, and it is already changing how analysts and researchers spend their days. But the firms that get the tenfold, and not the disappointment, will be the ones who did the quiet work first. They structured their information. Then they handed it to the agents.

If your data is not ready, that is where to start. We built the Data Engine for exactly that.