How Ramp Built LLM-Backed Expense Policy Agents Users Can Trust
The problem
Expense approval was a tedious, time-consuming task traditionally handled by managers, and applying LLMs in a finance product required careful design to avoid losing user trust through low-quality or unexpected outputs.
Workflow diagram · grounded in source
1
Expense submitted for approval
Trigger
The policy agent is applied to the traditionally manager-handled task of approving expenses.
▾ source quote
“approving expenses. This has traditionally been a manager's responsibility but we're now applying a "policy agent" that finance teams can trust to match or exceed human judgment”
2
Policy agent evaluates expense
Ai action
The policy agent evaluates the expense against the relevant sections of the user's expense policy.
▾ source quote
“the policy agent links directly to sections of the user's expense policy that its reasoning references”
3
Reasoning and citations output
Output
The agent outputs an explanation of why an expense was approved, with a link to the relevant policy section.
▾ source quote
“This first bullet point is output from an LLM explaining why the expense was approved. The info icon links directly to the section in the user's expense policy describing the requirements.”
4
Uncertainty escalation
Routing
When the agent is not sure, it falls back to the pre-agent escalation process.
▾ source quote
“When the agent isn't sure, we fall back to the pre-agent escalation process users are accustomed to”
5
Deterministic rule guardrails
Validation
Deterministic rules — dollar limits, vendor blocklists, and category restrictions — are layered on top of LLM decisions.
▾ source quote
“We layer deterministic rules on top: dollar limits, vendor blocklists, category restrictions”
6
Feedback loop improves policy
Feedback loop
User interactions and agent uncertainty explanations feed back to improve context and policy accuracy over time.
▾ source quote
“This feedback loop not only reduces the amount of work humans need to do over time, but also improves policy accuracy”
Reported outcome
Since enabling the policy agent at Ramp, more than 65% of expense approvals are fully handled by the agent, with a feedback loop that reduces human workload and improves policy accuracy over time.
Reported metrics
Expense approvals handled by agentmore than 65%
Human workload over timereduces the amount of work humans need to do over time
Policy accuracyimproves policy accuracy
Reported stack
LLMworkflow builder
◆ Does this fit your context?
Compare to your context
Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.
Since enabling the policy agent at Ramp, more than 65% of expense approvals are fully handled by the agent, with a feedback loop that reduces human workload and improves policy accuracy over time.
What tools did this team use?
LLM, workflow builder.
What results were reported?
Expense approvals handled by agent: more than 65%; Human workload over time: reduces the amount of work humans need to do over time; Policy accuracy: improves policy accuracy (source-reported, not independently verified).
How is this expense management AI workflow structured?
This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.