Expense Policy as Culture

Expenses as a shared belief system

Companies sit on a spectrum from the 37signals approach of handing everyone a card and trusting them, to the large-company approach of tier-two-city hotel limits booked fourteen days in advance. Eric Glyman argues Ramp can back-test policy strictness against real outcomes, employee time spent on expenses, growth rate, end margins, and finds that high-growth companies tend to resemble Netflix's "no rules rules": trust but verify, keep it visible, stay loose.¹

The deeper point is that a moral code cannot be encoded as a simple if-this-then-that rule table. The two obvious extremes are both wrong: forcing an executive to stay at a Motel 6 and take a two-hour taxi ride before signing a two-million-dollar contract technically saves money but is penny-wise and pound-foolish; the right call might be to stay at the Four Seasons. What a company actually wants is the owner's instinct, the sense that spending unnecessarily on the company's money is the same as stealing from yourself, which is nearly impossible to impose consistently across tens of thousands of employees and decays as a company scales past its founders. Glyman frames the judgment call with a 12 Angry Men logic: with more context and evidence, the moral conclusion about a given expense can change entirely, so the design principle that follows is to build tools that hold more context and do more jobs, rather than narrower rules that do one job worse.

The breakthrough

The mechanism is to take an expense policy written in plain English, for instance a rule that a flight longer than five hours can be booked in business class, and run it through a large language model with full context: transaction metadata, receipt data, timing, and the policy itself. Ramp now reviews more than one hundred thousand expenses a day this way, agentically, at better than ninety-nine percent accuracy, with a full audit trail of the reasoning behind each decision, a level of consistency higher than most human reviewers achieve, since most people do not actually know the policy by heart. The system still satisfies the separation-of-duties requirement under Sarbanes-Oxley, since an independent reviewer applies the rule rather than the spender self-certifying, while removing the roughly one hour a month a manager used to lose to a job that was never really part of anyone's actual role.

The engineering view: a self-improving program

Karim Atiyeh describes the same policy agent from the engineering side, and it is a clean instance of treating code as a combination of a large language model, instructions, and a running loop.² The plain-English policy functions as the instructions; the tools give the agent more context than any single human reviewer typically has, including integration with calendar and email; and the loop runs 24/7 as transactions arrive. As it runs, the system gets better at suggesting how to make the underlying policy clearer, so the plain-English document becomes a living text that evolves over time, with a human still holding sign-off. The same shape generalizes past expenses into other undocumented finance work, invoice fraud checks, whether an order was actually placed, whether it was actually received, whether the price matches, because Ramp's exposure to how many customers actually run their finances lets it infer unwritten policies from behavior and ship them as the next generation of agents.

Why it matters

The policy agent is a concrete instance of selling an outcome rather than a tool: real work a manager would otherwise have had to do gets done, at higher fidelity, and is simply gone. It also reframes compliance from friction into culture transmission, since an agent capable of reading a values-laden policy can keep a founder's moral code legible even at a scale no printed rule table could reach. The obvious limit is that ninety-nine percent accuracy across one hundred thousand daily expenses still means roughly a thousand errors a day, which is exactly where the human sign-off and audit trail carry the real weight, and the published accuracy figure does not disclose whether those errors skew toward false approvals or false declines.