Every fund operator has had the same conversation in the last six months. A vendor walks in with an "AI agent platform" that promises to eliminate operational headcount. The CTO nods politely and asks for the sandbox. The pilot runs. The agent hallucinates a P&L number. The pilot dies.
We've walked away from that pitch dozens of times. We've also shipped agentic AI into production at three funds. Here's what we've learned about where it works.
Where It Works: Knowledge Retrieval Over Institutional Memory
The single highest-ROI agent we have deployed is the one that answers the question "have we seen this before?"
Every fund accumulates institutional knowledge in the form of past trades, past compliance memos, past board materials, past LP correspondence, past vendor evaluations. Most of this is unsearchable. People who have been at the firm for 8 years remember things; people who joined 18 months ago do not.
A retrieval agent over the firm's document corpus, governed by access tier, returns the right document and the right paragraph in 30 seconds instead of 30 minutes. The LP queries that used to escalate to the CIO now resolve at the analyst layer.
This is unglamorous. It also pays back inside the first quarter.
Where It Works: Reconciliation Triage
The second deployment is the boring one. A reconciliation agent reads the daily break report and classifies each break: "known pattern, auto-resolve via this rule," "escalate to ops because this counterparty does this every quarter-end," or "novel - escalate to senior."
The agent doesn't fix breaks. It triages them. The ops team's actual reconciliation time drops 40 to 60% in our deployments because they stop spending mornings re-deriving what kind of break they are looking at.
This works because the agent is working over a closed corpus of break patterns the firm already understands, not generating new analysis.
Where It Works: Pre-Read Generation for Investment Committee
The third deployment is the highest-judgment one. Before each investment committee meeting, an agent assembles the pre-read packet from the firm's own data - current portfolio exposures, the deal pipeline, the open compliance items, the LP commitments timeline. The CIO reviews it, edits it, and sends it.
The pre-read used to take a senior analyst eight hours. The agent does the assembly in twenty minutes. The senior analyst spends ninety minutes reviewing and adjusting.
The judgment stays human. The assembly is automated. The CIO gets a better packet because more iterations are possible inside the same time budget.
Where It Doesn't Work: Anything Producing a Defensible Number
We have been pitched, and we have refused, agents that produce P&L numbers. NAV numbers. Performance attribution. Anything an investor or a regulator might ask the firm to defend.
The reason is not that the math is hard. The reason is that the firm cannot point to a deterministic, auditable trail when the regulator asks how a number was produced. "An LLM generated it" is not an acceptable answer.
For these workflows we run agents next to the deterministic system, not inside it. The agent suggests anomalies, flags variances, and writes the explanatory narrative. The actual numbers come out of code that has been reviewed, tested, and signed off.
What This Means for Adoption
The pattern across our deployments is the same: agents that route, retrieve, summarize, and triage are ready. Agents that produce ground truth are not.
Funds that deploy agents in the first category are quietly compounding operational leverage. Funds that try to deploy agents in the second category produce headlines and then quietly walk them back.
If you're evaluating agentic AI for fund operations: pick the boring use cases first. The exciting use cases will be ready in 18 to 24 months and will require a deterministic foundation that is easier to build now, while the agent layer is still being figured out.