In the first three weeks of November, three central banking institutions described the same worry in slightly different words.
On November 3, Alexandre Tombini of the Bank for International Settlements told a financial stability seminar that “if many institutions rely on similar models, data sources or third-party providers, their decisions may become more correlated,” which could “amplify herding behaviour.”¹ On November 16 in Milan, Federal Reserve Governor Michael Barr warned that traders relying on “a small set of similar AI models trained on similar data” face the risk that their strategies “synchronize reactions to news and other information.”² Two days later, the ECB’s Financial Stability Review reported that AI-driven equity funds show “a persistently higher downside tail dependence than comparable non-AI funds.” In plain terms, they are more likely to take large losses at the same time.³

All three were mostly talking about trading. Most private-markets firms don’t run AI trading strategies, so it is tempting to file this under someone else’s problem. We think that would be a mistake. In a firm that raises capital, writes IC memos and answers LPs, the herd forms long before any order reaches a market. It forms in what the firm’s AI reads, and in what that AI is allowed to learn.
The Quieter Kind of Herding
Consider how AI is actually entering small and mid-sized managers this year. An analyst has a model summarize the week’s podcasts and newsletters. A partner asks it to sharpen the macro section of the quarterly letter. Someone drafts a sector view for the IC deck. Each step is sensible on its own. Now multiply it across a few hundred firms that follow the same handful of shows, read the same newsletters and build on one of a few foundation models.
The result is a convergence of judgment, and it is much harder to see than a crowded trade. Theses start to rhyme. LP letters pick up the same phrases in the same quarter. Diligence questions get borrowed instead of earned. Barr made a point that applies to a GP as much as to a trading desk: many of the important questions “will be answered not in the abstract, but in the technical choices made about model design, training data, testing, and deployment.”² For a fund manager, the training data that matters most is the firm’s own information diet.
What Happened When We Fed Our Own System the Conversation
We run OakTech on our internal AI operating system. One part of it, which we call the information harness, exists to do exactly what the analyst above is doing: absorb the long-form markets and technology conversation so the founder doesn’t have to, and surface anything that should change how we operate or what we publish.
We pointed it at 55 episodes of widely followed markets and technology podcasts. It produced 134 candidate insights. Of those, 78 were proposed changes to how the system itself should work, and 56 were ideas for content.
Here is the number we found most useful. The system rated exactly one of the 134 as high confidence. Ninety-one were medium and 42 were low. Forty were marked skip, at least by default, usually because the material was political, would be stale within weeks, or leaned on figures nobody had checked against a primary source.
It would have been easy to treat dozens of hours of well-known investors talking as a stream of edge. The harness treated it as a pile of claims. That ratio says a lot about how little of the consensus conversation a disciplined firm should actually absorb.
Most of What Your AI Learns Should Stay in the Ledger
The rule that makes this safe is short. None of those 78 proposed improvements changes how the firm operates until the founder approves it. Each one is written to a ledger with its source, a confidence label, an owner and a suggested next step, and there it waits as a candidate. Nothing gets deleted, so we can always see what was considered and turned down.
What surprised us was the most common fate of the insights that did hold up: they confirmed something we already did. A conversation about acquisitions lined up with our existing rule that any deal recommendation must include downside debt scenarios, and the harness logged “no new rule” instead of inventing one. A segment on new computing hardware lined up with a rule we already apply to vendors: when we add a model or provider, our prompts, evaluations and data cannot be trapped inside one company’s API.
Confirmation is underrated. An AI system that turns every interesting episode into a new process will drift toward whatever the loudest voices said this month. That is herding too, carried out one firm at a time.
Herding Starts With Vocabulary
Some of the most useful catches were about words. Several insights flagged that a phrase was a show’s own branding or framing and told our content pipeline not to adopt it. Another became a standing writing rule: describe what a model actually did on a named task, rather than repeating a vendor’s label for whatever era of AI we are supposedly in.
That can sound like copy editing. It matters more than that. When thirty managers describe the same quarter with the same borrowed metaphor, an allocator who reads all thirty letters learns nothing from any of them and starts to wonder who is thinking for themselves. Language is the first place an LP can tell whether your view is actually yours.
The same discipline applies to market narratives. When one segment covered a debt scare at a single AI infrastructure company, the lesson the harness logged was to re-check leverage and offtake name by name, and not to conclude that AI demand was over. This month’s data shows why. Nvidia reported record data center revenue of $107.6 billion for the quarter ended October 25, up 110% from a year earlier.⁴ The same week, the ECB said euro area banks are showing “signs of deteriorating credit quality in exposures to software firms and semiconductor producers.”⁵ Both things are true at once. A firm whose AI soaks up only the narrative of the week will get at least one of them wrong.
Your Edge Is First-Party Data, Graded Honestly
If the public conversation is what the herd feeds on, what is left that belongs only to you? For most GPs, it is first-party operating data: the pipeline, the LP conversations, the meeting outcomes, the record of what you said you would do and whether you did it. No other firm’s AI has read it.
That record is only an edge if it is complete and you know where it isn’t. Our system’s latest production report, run on the evening of November 21, fully covered 21 of the 24 sources it is required to check and flagged the other three as stale. That is an 87.5% green rate, and the run was still graded yellow, not green.
Conflicting data gets the same treatment. When a task or meeting appears in two systems with different details, our rule is that the newest update wins. On one development run, the system found 49 such conflicts. Thirty-eight already matched our database and 11 did not. All 49 were reported, so a person could see what would change before trusting it. In the version we are rolling out now, we also tightened what counts as “read.” Seeing a message in a preview or a chat list no longer counts as processing it. Every source goes through the same capture-and-check path before the system treats anything as known.
None of this is exciting. It is what allows a founder to trust a summary without rereading the inbox underneath it, which is the whole point.
Questions We Would Put to Any Manager Using AI
Tombini’s list of safeguards opens with “transparency, so that significant AI-assisted decisions can be understood, challenged and audited.”¹ He was speaking about financial markets broadly, but the idea carries straight over to a GP. Four questions we ask of ourselves, and would ask of any firm:
- When your AI suggests a view or a process change, where is it written down, with its source and a confidence label, before anyone acts on it?
- Who has to approve before a learned “insight” changes a memo template, a diligence checklist or an LP letter?
- How much of what your AI reads is the same public material your competitors’ AI reads, and how much is your own operating record?
- Could you switch models next quarter without losing your prompts, tests and data?
Our view is that AI herding in private markets won’t look like a flash crash. The firms it hurts will fade slowly, sounding a little more like everyone else each quarter, while the LPs who read every letter quietly stop telling them apart. The defense is unfashionable. Read widely, believe little, write everything down, and require a person to sign off before the firm changes its mind.
That is the operating discipline we build with clients in Business Process AI (BPAI), and it shapes the investor operations we set up for emerging managers through CRS Launch. OakTech has helped clients raise $235M+ in capital, and we run our own firm by the same rules.
Find Out What Your AI Is Actually Learning
If your team uses AI to read the market or draft what goes to LPs, we can map what it absorbs, where a person should sign off, and which of your own data is the real edge.
References
- Tombini, Alexandre. “Navigating financial stability in the age of AI.” Speech at the Mexican Banking Association Seminar on Financial Stability and Markets, Bank for International Settlements, November 3, 2026. https://www.bis.org/speeches/20261103-navigating-financial-stability-age-ai.pdf
- Barr, Michael S. “Some Potential Financial Stability Implications of Artificial Intelligence.” Speech at the Seventh Annual ACM International Conference on AI in Finance, Milan, Board of Governors of the Federal Reserve System, November 16, 2026. https://www.federalreserve.gov/newsevents/speech/barr20261116a.htm
- European Central Bank. “Financial stability implications of AI adoption in the financial industry and the real economy.” Special feature, Financial Stability Review, November 2026 (published November 18, 2026). https://www.ecb.europa.eu/press/financial-stability-publications/fsr/special/html/ecb.fsrart202611_01~dd6e15a84a.en.html
- NVIDIA. “NVIDIA Announces Financial Results for Third Quarter Fiscal 2027.” Press release, November 17, 2026. https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Third-Quarter-Fiscal-2027/default.aspx
- European Central Bank. Financial Stability Review, November 2026 (published November 18, 2026). https://www.ecb.europa.eu/press/financial-stability-publications/fsr/html/ecb.fsr202611~e03b919d36.en.html
For a practical way to turn this into operating controls, see our capture, verify, act guardrail pattern. To map the processes that need those controls, start with an AI business process assessment.