How would you measure the success of Ramp's Agents for Controllers, the AI agents that review expenses against company policy?

Introduction

In July 2025 Ramp introduced Agents for Controllers, AI agents that review expenses against company policy, flag violations, auto-approve compliant spend, and help code transactions so controllers spend less time on manual review. Since then Ramp has extended agents into accounts payable and procurement. Measuring the success of an AI agent is different from measuring a traditional feature: the agent does work on behalf of a user, so we must measure the quality of that work, how much of it humans trust, and the time it gives back, not just clicks.

Step 1

Product Context (3 minutes)

Agents for Controllers sit inside Ramp's expense management workflow:

1. An employee spends on a Ramp card or submits a reimbursement.
2. Ramp collects receipts, memos, and merchant data.
3. The agent reads the company's expense policy and the transaction context.
4. It decides: approve, flag with a reason, or request missing information from the employee.
5. It suggests or applies accounting coding (GL account, department, class) for sync to the ERP.
6. The controller reviews flagged items and audits a sample of agent decisions.

Key stakeholders:

• Controllers and accounting teams: want faster, more consistent review and a cleaner close
• Employees: want fast approvals and fewer follow-up questions
• Managers: want to approve less and only see what matters
• Auditors: need a defensible trail of who approved what and why
• Ramp: agents justify Ramp Plus and enterprise pricing and differentiate against Brex, Navan, and Expensify

Lifecycle stage: early adoption. Customers are testing how much autonomy to grant, so trust-building metrics matter as much as efficiency metrics.

Why this is different from measuring a normal feature: a traditional expense review screen succeeds when people use it. An agent succeeds when people use the review screen less, because the agent has already done the work correctly. Engagement metrics like daily active controllers can actually fall as the agent improves. That inversion is the most important point to make early in the interview, because it changes which metrics are healthy. Interviewers at Ramp often probe whether candidates notice that "less time in the product" can be the goal for a finance automation tool.