Varun Agarwal, CFA ← Home
A brief

The investment-ops bench of 2031 looks nothing like the one of 2021.

For most of the last twenty years, investment operations at a foundation or endowment has looked the same: a tight team of two or three on the ops bench, with a separate accounting/tax/audit track running alongside. Roughly half the ops week goes to reconciliations and vendor relations — custodians, analytics platforms, data feeds. Most of that work isn't specialized. It's just context-switching across five different systems. That cost just collapsed.

Varun Agarwal · ~4 minute read · Investment operations · AI · institutional

The shape the function actually takes

An investment operations team at a foundation or endowment investment office below $5B in assets is typically tight — two to three people on the ops bench specifically, plus a director or COO on top. Accounting, tax, and audit run on a separate track with their own staff and an external auditor each year. Reporting sometimes sits with ops; sometimes it sits with whoever in the risk or investment team owns the analytics platform. That part varies by org.

For the two-to-three on the ops bench, the work splits into a small number of buckets that consume real chunks of the week:

None of this is deeply specialized in the sense of requiring years of unique training. The reason it splits across two or three people is throughput (there's too much of it on any given Tuesday) and context-switching (every vendor portal has its own schema, vocabulary, login, and quirks). That second one is what AI changes.

What AI actually changes

The change isn't that AI replaces reconciliation or vendor management as functions. It's that AI collapses the context-switching cost between them. A modern ops operator with Claude open can move from a custodian portal to a Caissa refresh to a parser for the latest manager statement without re-loading the mental model each time. The mechanical parts of recs — matching, flagging breaks, drafting break notes — compress dramatically. Vendor relations compress too, because AI can hold the schemas, conventions, and quirks of all five systems at once.

The director is one person. The orchestra is software. The 2—3-person ops team becomes 1—2, and the reporting + analytics work that used to sit outside ops can come back inside it.

That's not a forecast. It's already what works. The cost of building internal tooling — a reconciliation harness, a CRM, a custom dashboard, a vendor-data refresh script — has dropped by an order of magnitude in eighteen months. What used to require a vendor RFP, a six-month integration build, and a consultant now requires a weekend and a clear thesis about what the team actually needs. The accounting/tax/audit track stays roughly where it was — that work has its own external dependencies and doesn't compress the same way — but the ops bench, and the reporting and analytics layer adjacent to it, looks fundamentally different.

What this looks like in practice

This site is itself the proof. One person. One week. End-to-end — design, build, hosting, copy, the whole stack. The exact same posture I'd bring to running an investment operations function:

Practical examples on this site
  • /work/ Six production-style dashboards covering the operational backbone of a foundation office — deal flow, performance attribution, public & private reconciliation, alt-asset document ingestion, sign-off controls. The kind of internal reporting infrastructure an office runs monthly.
  • /crm/ An LLM-driven investment-research CRM that consolidates manager meeting notes, with sentiment and productivity scoring on each conversation and source-cited Q&A across thousands of documents. Built locally; PDFs never leave the user's machine.
  • /about/ The full track record — fifteen years of risk and operations work across Harvard Management Company, Johns Hopkins, UBC Investment Management, and The Nature Conservancy. $21B AUM modeled in MSCI Total Plan at Harvard.
  • /conductor/ The long-form narrative — a 30-panel graphic novella tracing the arc from a 2006 neural-network thesis to a fevered week in March 2025 that re-opened a door I'd shelved for almost two decades.

Where this leaves the function

The shape of an investment operations team five years from now isn't a smaller version of today's team. It's a different team altogether — fewer seats, more leverage per seat, and a working assumption that internal tooling gets shipped in days, not quarters. The directors and operators who'll do this well are the ones who already think this way today: who treat reconciliation harnesses, custodian integrations, and analytics platforms not as fixed infrastructure they inherit, but as design problems they own.

This essay is one operator's working notes on what that looks like in practice. The examples linked above are the most concrete version of it I can show in the open — built end-to-end under the Blue Maple Foundation banner.

Notes from the field

If you're thinking about this too, I'd be interested to compare notes.

Reach me at va2238@caa.columbia.edu or through LinkedIn — especially if you're a senior operator or director building or rebuilding an investment operations function in the AI age. Always happy to trade ideas.