ExpertsFMAI in the Investment Office – Abby Barlow, Laura Hill, Brian Sugrue, Jenny Heller, John Lawrence, Matt Bank, Kristin Kallergis Rowland, Jon Webster (EP.515)
Capital Allocators – Inside the Institutional Investment Industry · September 14, 2026
5 Selected Lessons:
framework · 3 heuristic · 1 priority · 1
Preserve human judgment in the core decision, automate everything else
Source
Capital Allocators – Inside the Institutional Investment Industry — AI in the Investment Office – Abby Barlow, Laura Hill, Brian Sugrue, Jenny Heller, John Lawrence, Matt Bank, Kristin Kallergis Rowland, Jon Webster (EP.515) (September 14, 2026) · at 8:28 ▸
When this applies
when introducing automation into any workflow that contains a core judgment task
The takeaway
Identify the single step in a process that constitutes your real edge, and explicitly wall it off from automation even while automating its surrounding scaffolding.
Why it works
The value of a decision memo lies in forcing a human to reason through what could go wrong; outsourcing that reasoning to AI removes the very friction that produces insight and accountability.
One level up
Decide what NOT to automate before deciding what to automate — the exclusion list protects your differentiator.
Where it stops holding
This stops working when the 'core judgment' step can itself be decomposed into verifiable sub-decisions that AI can reliably replicate — at that point the boundary may need to move.
Detail
hierarchy:
- Human judgment, confidentiality, and accountability on core investment questions (preserved, never automated)
- Formatting, polish, graphics of memos (can be automated)
- Routine work and time savings (accelerated by AI)
rationale: The purpose of a memo is to force the team to ask what would make the bet blow up — AI must not answer that question for them.
In their words
We will use AI to augment human expertise, accelerate routine work, and improve decision quality while preserving human judgment, confidentiality, and accountability.
Master context document for consistent AI outputs across a team
Source
Capital Allocators – Inside the Institutional Investment Industry — AI in the Investment Office – Abby Barlow, Laura Hill, Brian Sugrue, Jenny Heller, John Lawrence, Matt Bank, Kristin Kallergis Rowland, Jon Webster (EP.515) (September 14, 2026) · at 6:11 ▸
When this applies
when multiple people or multiple projects need an AI tool to behave with consistent institutional knowledge
The takeaway
Invest once in capturing organizational and personal context in a reusable document, then feed it to AI at the start of any new task to raise output quality and consistency.
Why it works
An AI tool without persistent context treats each session as a blank slate; a structured interview-generated context document gives it durable background so outputs are tailored rather than generic.
One level up
Shared infrastructure (a context layer) compounds in value as more people and projects draw on it.
Where it stops holding
This stops working when the organization or roles change frequently enough that the context document becomes stale faster than it's maintained.
Detail
procedure:
- Have AI interview you for an hour about the organization, its history, your working style
- Have AI synthesize this into a shared context document
- Require every new project or new hire to load this document into AI before starting work
rationale: A new hire's AI-assisted work improves immediately if the AI already 'knows' where he sits and what he's working on.
In their words
I want the output at the end to be a document that is shared for any new project that you start working on. We want to create an enterprise-level context doc that everyone on the team is using consistently.
Trust layers for safe AI adoption (perimeter, output, provenance)
Source
Capital Allocators – Inside the Institutional Investment Industry — AI in the Investment Office – Abby Barlow, Laura Hill, Brian Sugrue, Jenny Heller, John Lawrence, Matt Bank, Kristin Kallergis Rowland, Jon Webster (EP.515) (September 14, 2026) · at 23:14 ▸
When this applies
when scaling any new tool across a team handling sensitive data and consequential decisions
The takeaway
Before scaling a powerful but imperfect tool across a team, build explicit layered trust: trust the boundaries (where data goes), trust the output (verification discipline), and trust each other (transparency on what's human vs. machine).
Why it works
Trust is multi-dimensional — confidentiality boundaries, output accuracy, and human provenance each address a different failure mode, so earning permission to expand usage requires satisfying all three, not just one.
One level up
Permission to expand use of a new capability is earned incrementally through demonstrated trustworthiness, not granted all at once.
Where it stops holding
This stops working when speed pressure forces skipping the verification or sandboxing steps, eroding the trust the framework depends on.
Detail
procedure:
- Trust in the boundaries: use enterprise-grade tools that don't train on your data; sandbox agents before deployment
- Trust in the output: maintain a verification discipline and keep humans in consequential decisions
- Trust each other: track provenance of what's human vs. AI-generated work
rationale: This is what earns us permission to use it at all.
In their words
A principle that underlies how we're using AI today, and we'll continue to use it, is building layers of trust.
Encoding AI across the full alts investment lifecycle (sourcing, diligence, red-teaming, monitoring, servicing)
Source
Capital Allocators – Inside the Institutional Investment Industry — AI in the Investment Office – Abby Barlow, Laura Hill, Brian Sugrue, Jenny Heller, John Lawrence, Matt Bank, Kristin Kallergis Rowland, Jon Webster (EP.515) (September 14, 2026) · at 43:47 ▸ · at 45:53 ▸
When this applies
when deciding how deeply to integrate a new capability into an end-to-end operational process rather than bolting it onto one step
The takeaway
When a process's core constraint is unstructured, inconsistent document handling, embed the automating capability at every stage of the lifecycle rather than as a single bolt-on feature.
Why it works
Alternative investments run on unstructured, inconsistent documents at every stage (PPMs, side letters, GP reports), so a tool that structures and compares documents removes a scalability bottleneck that previously required slow manual retyping at each step.
One level up
Full-lifecycle integration, not point solutions, is what converts a tool into genuine workflow transformation.
Where it stops holding
This stops working when the organization lacks the scale or document volume to justify lifecycle-wide integration investment.
Detail
procedure:
- Fund research and data collection from managers
- Pre-populate first draft of investment review committee memo
- Run a red team agent as devil's advocate before IC
- Ongoing diligence checking against original thesis
- Structuring, onboarding, negotiating terms, compare-and-contrast
- Portfolio construction simulations
- Servicing and operations
rationale: It's taken off probably a third of the workload of people just in the last six to nine months; AI is embedded into every aspect, not bolted on the side.
In their words
It's taken off probably a third of the workload of people just in the last six to nine months. AI is embedded into every aspect of our work stream. It's not bolted on the side. It is a part of everything.
AI as always-on junior analyst requiring mandatory re-check pass
Source
Capital Allocators – Inside the Institutional Investment Industry — AI in the Investment Office – Abby Barlow, Laura Hill, Brian Sugrue, Jenny Heller, John Lawrence, Matt Bank, Kristin Kallergis Rowland, Jon Webster (EP.515) (September 14, 2026) · at 0:00 ▸ · at 3:52 ▸ · at 6:11 ▸
When this applies
when delegating any analytical first draft to an AI tool, especially one involving numbers
The takeaway
Treat AI output like a junior analyst's first draft: never ship it unreviewed, and build in an explicit self-verification step before trusting the numbers.
Why it works
LLMs can silently make computational errors (e.g., double-counting) that are easy to miss amid otherwise polished output, so a dedicated self-audit step surfaces errors a first read would not catch.
One level up
Speed gains from automation must be paired with a proportional increase in verification discipline, not a decrease.
Where it stops holding
This stops working when the reviewer lacks the expertise to catch subtle errors themselves, since the AI's self-check is not guaranteed to find everything.
Detail
rule: After any AI deliverable, explicitly ask it to revisit the original goal and triple-check its own numbers for errors.
threshold: Applied to every deliverable involving numbers or complexity, not just occasionally.
rationale: AI caught its own double-counting error in a pacing model only when explicitly asked to re-verify against the original goal.
In their words
I have gotten in a habit where I will ask it at the end of any deliverable. Go back to the beginning of this conversation and what we were trying to accomplish, review the deliverable, triple check the numbers again, find your own errors.