There is a version of this article that opens with wonder. AI is incredible. The future is here. I will not be writing that version.
I have been running an AI-native research operation for several months now. I want to tell you what it actually is — how it works in practice, what it does well, where it fails, and what it has changed about the way investment research gets done.
The short version: it is not magic. It is management.
The Team
I work with two AI agents directly. Hermes — a research agent running on a Mac Mini in my home — manages the automated research infrastructure. ChatGPT serves as a thinking partner and drafting assistant. They are not interchangeable. They have different strengths, different tendencies, and different failure modes. Managing them well requires understanding the difference.
That word — managing — is deliberate. An AI agent does exactly what you tell it to do. The difficulty is telling it the right thing. Every morning brief, every dashboard, every article on this site has a human instruction behind it. The quality of the output reflects the quality of the direction.
Every weekday morning at 06:30 UTC, a full macro briefing assembles automatically: live market data, central bank signals, geopolitical developments, commodity complex, cryptocurrency, emerging markets. Data is fetched, structured, and formatted. The briefing lands in my inbox before I open my eyes.
Five specialist dashboards publish weekly on a fixed schedule — EM Liquidity, China Credit, Commodities, BTC Treasury, ETH Staking. Each one draws on multiple live data sources, applies the same analytical framework, and produces a formatted report for review.
None of this replaces analysis. All of it makes analysis faster, broader, and more consistent.
What AI Does Well
Coverage. The analytical frameworks I use — liquidity cycles, credit transmission, capital flows — require tracking a large number of variables simultaneously. A human analyst can follow five things carefully. An AI agent can monitor fifty things adequately. The difference matters at the edges, where signals appear before they become obvious.
Consistency. Human attention is irregular. On a busy day, the emerging market section gets less thought. The AI agent does not have busy days. Every dashboard runs to the same specification, every week, regardless of what else is happening. That consistency has real analytical value.
Speed. From raw data to formatted, readable report: minutes, not hours. This is not a marginal improvement. It changes what is possible within a research workflow.
Memory. Every decision, every framework, every instruction I have given is recorded and retrievable. The research infrastructure improves over time because it remembers what worked.
Where AI Fails
Judgement. This is the central limitation, and it is not a small one. An AI agent can identify that copper is down 3% and that this has historically preceded EM weakness. It cannot tell you whether this time is different — whether the data point is signal or noise, whether the regime has shifted, whether the macro narrative that explained last cycle still applies.
That distinction requires context, experience, and the kind of pattern recognition that comes from watching markets across multiple cycles. It cannot be automated. Every report I publish passes through that filter. The AI provides the data surface. The judgement is mine.
Creativity. AI agents are very good at executing within a defined framework. They are poor at questioning the framework itself. The best macro calls are often the ones that step outside the prevailing consensus view. That requires a different kind of thinking — one that I have not yet seen reliably replicated by a machine.
Errors. AI agents make mistakes. They misread data, apply the wrong context, occasionally invent facts that are plausible but wrong. The research infrastructure I have built includes parity checks — automated verification that catches the most obvious failures. But the final responsibility for what is published belongs to a human reader. Every report gets reviewed before distribution.
What Has Actually Changed
The honest answer is: access.
What used to require a research team, a Bloomberg terminal, a compliance department, and a six-figure infrastructure budget, I now run on a Mac Mini, with two AI agents, at a fraction of the cost. The analytical frameworks are the same. The data sources are broadly equivalent. The coverage is comparable.
This is not a marginal efficiency gain. It is a structural shift in who gets to participate in institutional-quality research.
Twelve years in asset management taught me how to think about markets. The AI infrastructure gives me the reach to act on that thinking at scale, consistently, without requiring a team.
The judgement is still mine. The leverage is new.