AI operations in finance start with someone who knows the work

Every month, finance teams at growing companies and investment firms repeat the same work: closing the books, preparing board and investor reports, answering diligence questions, screening opportunities and tracking portfolio positions. Much of it is now within reach of AI agents. In our view, the projects that succeed start from someone who knows the work stating, in plain terms, what each process must achieve, and only then turn to the tools.

We believe the most common failure in AI operations is a build that begins with a capable developer and a vague goal. A developer can promise to build almost anything. Without the expertise to know what the right output looks like, the result is difficult to trust and harder to maintain.

Why expertise comes first

Financial work is full of rules that are obvious to practitioners and invisible to everyone else. A term sheet is not an executed agreement, gross burn is not net burn, and a cash balance that has not been reconciled is an estimate. An agent that confuses any of these will produce output that looks complete and is wrong.

Yet the problem runs deeper than definitions. A specification is only as good as the judgment of the person who writes it, and a test is only as good as the expected output it checks against. Thus, the practitioner who will rely on the process should write both, with modest technical help, before the build starts.

Keep each process as simple as possible

We find that the best specifications are short. Each one states what the process should achieve, from what inputs, for whom, and how a person will know the output is correct. The specification for a monthly investor report, for example, can fit on a page: the figures it must contain, where each comes from, who reviews it and when it is sent.

Hence, the specification should be as simple as possible: this model should achieve this, and this agent should produce that. Simplicity keeps the outcome in view for the people who own it, and it gives the developer a target that can be tested. Detail belongs in the build, where it can be checked against the statement of what the process is for.

Sequence the work

The order in which processes are automated matters as much as the processes themselves. A company that automates customer service before it has fixed the marketing that brings customers in has built a reply system for an empty queue. The same logic applies to finance: automating investor updates before the monthly figures are reconciled only sends unreconciled figures faster.

Thus, we start with the process whose output the others depend on. For most companies that is the monthly close and the core figures. For most funds it is the record of positions and transactions. Once those are reliable, reporting, updates and screening can be built on top of them.

Governance from day one

AI operations in finance need governance as much as code. The National Institute of Standards and Technology published its AI Risk Management Framework 1.0 in January 2023, organized around four functions: Govern, Map, Measure and Manage. It is voluntary, and it is written for organizations of every size. Its core point applies directly to a finance team: decide who is accountable, understand the context in which the system runs, measure whether it performs as intended, and manage the risks that remain.

In practice, we translate that into a few working rules:

• Every agent has a named human owner who reviews its output before anything leaves the firm.
• Instructions, status and decisions live in a repository that the team can read, with a log of changes.
• Credentials are granted through the tools' own access controls, never pasted into chats or documents, and a scan checks every change for secrets.
• Each process has a test: a known input and the output it should produce.

What's more, firms registered as investment advisers in the United States should note that the SEC's marketing rule, Rule 206(4)-1 under the Investment Advisers Act, applies to advertisements however they are drafted. An agent that drafts marketing or investor material for an adviser is drafting within that rule, and its output needs the same compliance review as any other.

What a first engagement looks like

For a company or fund new to this work, we would choose one recurring process with a clear owner and a clear output, such as call notes, inbox triage or the monthly investor update. We would write its specification in a page, build the agent and its test, run it alongside the existing process for a few cycles, and hand it to the team once the outputs match.

Hence, the measure of success is simple: the team runs at least one recurring process with agents, without outside help. A second and third process follow faster, because the repository, the guard and the habits are already in place.

Where AI operations fit in a raise

For a company preparing a round, the same approach turns diligence from a scramble into a routine. An agent that answers investor questions from an approved data room, with every answer reviewed by the company before it is sent, saves time for the team and gives investors faster, consistent responses. For a wealth manager or VC firm, agents that screen opportunities against a written mandate and assemble portfolio reports let a small team cover more ground without lowering its standards.

Our view

We believe AI operations in finance should start with a person who knows the work writing down, simply, what each process must achieve, followed by a sequence that automates the foundations first and governance that keeps a human accountable for every output. Firms that work this way get systems they can trust and maintain. Our view would change only for processes with no judgment in them at all, such as moving files between folders, where off-the-shelf automation is enough and expertise adds little.

AI Operations is the offer we run for this work, on its own, inside Fund Ops, or as the optional phase of Raise Ready.

Sources

1. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)", NIST AI 100-1, released January 26, 2023. nvlpubs.nist.gov
2. U.S. Securities and Exchange Commission, Division of Investment Management, "Marketing Compliance: Frequently Asked Questions". sec.gov

This article is general information, not legal or investment advice.