At Operator Collective's Spring Gathering, Erica Dorfman, CFO of Brex, led a roundtable on how AI is reshaping finance, legal, and risk management. Erica brought a particular vantage point to the conversation: the Capital One acquisition of Brex, completed in early April 2026, was weeks old when the gathering took place, giving her a live view of what AI integration looks like inside a major bank-fintech deal. Here are the session notes:
The spending gates are open, for now
The consensus across the table was that most companies are running with minimal restrictions on AI spend for the next twelve months. Teams have open access; tokens and projects are flowing freely. The underlying logic is that constraining experimentation costs more than the spend itself, at least while organizations are still learning what AI can actually do inside their specific businesses.
At Brex, Erica's approach has been to leave the gates open until AI costs become a meaningful constraint relative to overall engineering team budgets. But she's equally clear about what comes next. In one to two years, AI spend will need to be managed the way headcount is managed, and CFOs who haven't started building that framework will find themselves behind. The conversation she's already having with her peers: AI investment is becoming a P times Q equation, where organizations need to know not just how many people they need but how much AI capability, and how those two numbers trade against each other. Most companies can't answer that question yet.
The workforce math is changing structurally
The efficiency gains showing up in this generation of startups are not incremental improvements on existing operations. They're structural shifts. Teams of ten are doing work that previously required thirty or thirty-five people. Several companies represented at the table have effectively stopped entry-level hiring: the most junior person is in their late twenties with six or seven years of experience and already promoted twice.
The downstream consequence of that shift tends not to get enough attention. In law, AI is handling first drafts and junior associate work, with senior lawyers doing expert-level review. That arrangement is more efficient today. But the senior lawyers of 2035 were supposed to be the junior associates of today. If that pipeline closes off, the profession develops a structural problem that takes a decade to surface. Finance is heading toward a similar dynamic, and the time to think about it is before it arrives.
Accuracy requirements in finance and legal are non-negotiable
In most business contexts, a tool that's right 99% of the time is genuinely valuable. In finance and legal, that same tool can be catastrophic. The tolerance for error in a contract, a financial filing, or a compliance determination is effectively 100% or nothing. This isn't a caveat about AI's potential; it's a constraint that should shape how organizations design their workflows.
The table's view is that the answer isn't simply putting a human in the loop at the final step. That's necessary but not sufficient. The more important work happens at the front end: context setting, scaffolding, and workflow design that makes the AI's task tractable before it begins. Multi-agent review systems, where one agent does the work and a second reviews specifically for anomalies, are showing genuine promise. Deterministic workflows with intelligent steps embedded at key decision points are gaining traction over fully open-ended generation.
And the accountability question doesn't transfer to the tool. AI can assist, flag, and review. The decision, and the responsibility for it, still sits with a person. Building AI systems that obscure that chain of accountability is a governance failure, not just a design choice.
What finance leaders should be doing right now
Erica's practical guidance for CFOs at this stage: keep the spend gates open, but build the measurement framework before the costs force the conversation. Token costs alone tell you almost nothing. The metrics that matter are tied to actual business outcomes -- faster close cycles, reduced error rates, lower cost per transaction, faster customer response. Building those baselines now means having something to measure against when the reckoning arrives.
Getting specific about where AI changes the risk profile of the finance function is also worth doing now rather than later. Which processes are being automated, where are the error modes, and who owns the accountability chain? Finance leaders who can answer those questions clearly will be ahead when regulators start asking the same ones.
And the talent pipeline question deserves more attention than it's currently getting. Efficiency gains today are real. But building a function that depends entirely on AI to operate, without a bench of people who understand the underlying work, creates a fragility that may not be visible until it's expensive to fix. The CFO's job has always been to hold both the short-term and the long-term in view simultaneously. AI makes that harder, and more consequential, than it's ever been.
This article was adapted from a roundtable discussion at Operator Collective's Spring Gathering, held under Chatham House rules. Insights are shared without individual attribution.
