How Perplexity's finance team of 10 scales one of the fastest-growing AI startups
“Each member of our team has an outsized impact due to our focus on using high-leverage tools like Ramp.”
Controller, Perplexity
See insights on how 50,000+ customers spent on Ramp in 2025
Lauren Feeney is the financial controller at one of the world’s most talked-about AI companies. She leads the full scope of Perplexity’s finance function—global operations, hundreds of vendors, thousands of monthly transactions—with a general ledger team of one. That person is Patricia, Perplexity’s senior accountant—and the kind of operator who makes “exceptions only” systems actually work. Together, they process more in a month than most finance teams over three times their size.
Perplexity’s story isn’t just a case study in efficiency. It’s an argument made in real time that the operating model you choose early is the one you’ll live with at scale.
The tools most finance teams use weren't built for this velocity
When CEO Aravind Srinivas co-founded Perplexity in 2022, he made an early, deliberate decision about how the company would manage spend as it scaled. He chose Ramp early as the foundation for how money would move across the company—putting policy, visibility, and clean data in place from day one so the team could stay focused on building. The instinct was the same one that shaped the product: build it fast, make it transparent, and make it scale. "Velocity is core to our brand," Lauren says. "The market in which we operate moves extremely quickly and we cannot afford to let things drop."
The challenge wasn't just speed. It was control at scale. As Perplexity grew, so did its vendor footprint: hundreds of subscriptions, contractors, and services flowing in and out every month. The traditional response would have been to add headcount: more people to code transactions, chase approvals, reconcile accounts, and manually review line items. Lauren had a different instinct.
The card compromise that changed her approach
Lauren had seen what happened at a previous company when credit cards were distributed without structure. A co-founder’s card got compromised. Because it was a general-purpose card tied to dozens of vendors, the fallout hit 57 relationships at once. Fraud on one card. 57 downstream problems. The brittleness of an uncontrolled card program was impossible to ignore.
AP software alone didn't fix it either. Tools like Bill.com could process bills, but they couldn't surface why a specific invoice looked unusual. They couldn’t flag that a vendor’s charge was triple last month’s without someone doing the digging manually. At Perplexity's transaction volume, that gap was untenable. "You have hundreds of vendors," Lauren says. "It's not possible for me to remember what the monthly spend is for each one of those every time."
Move fast and trust the numbers: the AI finance stack that shows its work
Lauren’s answer was to build a finance stack optimized for leverage, not headcount: a punctilious system that could move fast without inviting sloppiness. The first initiative she stood up: a vendor-specific card program. Instead of issuing general-purpose credit cards across the team, each approved vendor relationship gets its own dedicated card with preset spend controls.
Each of those choices removes a small point of friction—and over time, that adds up. If a card is ever compromised, the damage is limited to one vendor relationship, not dozens. Coding context is captured at the moment of issuance rather than reconstructed weeks later. When an employee leaves, card management is frictionless: one card, one vendor, one clean cancellation. "My preference is to restrict the use of general-use credit cards," Lauren says. "It allows for greater efficiency at month-end and gives us generally better insight into our spend."
That architecture doesn’t just feel cleaner—it changes the math. Each vendor card is locked to its intended vendor and budget. If a charge doesn't match the preset parameters, it doesn't clear. There's no cleanup work after the fact, no disputed charges to resolve, no end-of-quarter surprises from vendor billing. Combined with Ramp’s cashback on every card transaction, Ramp has generated over $5M in cumulative savings for Perplexity. The savings come not from spending less, but from spending more intentionally, with every dollar on card earning a return.
The Accounting Agent that earns its keep
The card program set the guardrails. The real acceleration came from what happened next: the Accounting Agent. Perplexity was among Ramp’s earliest access customers for the feature, and they’ve spent months building a model that reflects how they actually code, not how any other average company does.
Nearly 100% of the team's card transactions are automatically coded through a combination of Patricia's rules and AI that handles the exceptions. When the Agent gets something wrong, the team corrects it in plain language, explains the context, flags edge cases, and gives the system enough information to update its own understanding. The model learns from each correction. The exceptions list shrinks. And the coding rate improves with every cycle—the system building an increasingly precise picture of how Perplexity codes.
"Finance professionals are by nature more skeptical of AI," Lauren says. "It's part of our training—always apply professional skepticism." What made her comfortable trusting the Agent with that kind of responsibility was the way it explains itself. She doesn’t get an answer. She gets the reasoning—and, crucially, the rationale is legible. In finance, that legibility is what turns automation from novelty into infrastructure.
"When I log in, it's clear to me what the Agent believes is ready to be approved. It draws my attention to anything worth an extra pair of eyes and tells me why—triple the spend month-over-month, a new line item description, a coding question. That's the magic."
— Lauren Feeney, Controller, Perplexity
This mirrors, not by accident, what Perplexity's own product does for users. Perplexity doesn't just return answers—it cites sources so users can verify. Ramp's Agent doesn't just code transactions—it explains decisions so the controller can judge. The same principle, transposed into a different domain.
Bills clear themselves, and exceptions find the right person
On the bill pay side, Lauren’s inbox surfaces only what needs human review. Each item comes pre-flagged with a reason. For large invoices, dynamic approval routing sends them directly to the right executive without finance playing scheduler.
"Being able to set thresholds gives me peace of mind that the required eyes are getting to the spend they need to see," Lauren says. "People get visibility automatically. Finance doesn't have to chase anyone down."
Month-end is a wrap-up, not a scramble
The clearest measure of what Lauren and Patricia have built is the ratio: roughly 7,000–9,000 card transactions per month, coded automatically at 97%—without manual touch.
That number has continued to climb. As the Accounting Agent builds an increasingly precise picture of how Perplexity codes—updated with every correction the team makes in plain language—the auto-coding rate has grown to >97% and continues to improve. Approvals are handled in real time throughout the month rather than compressed into a week of frantic reconciliation.
"Gone are the days of scrambling to chase receipts, manually coding really large data sets. All of that has been automatically addressed throughout the month in real time."
— Lauren Feeney, Controller, Perplexity
The hours that disappear from the routine reappear elsewhere: planning, partnership, and judgment. The biggest driver is the monthly close, which Ramp compresses by roughly 115 hours per month. When the Accounting Agent codes transactions continuously throughout the month, close isn’t a scramble. It's a review. That same Agent accounts for another approximately 28 hours of GL coding time saved each month. Patricia corrects exceptions in plain language, the model updates its understanding, and the exceptions list shrinks with every cycle.
The remaining hours come from structural changes upstream. The vendor-specific card program eliminates roughly 16 hours of monthly card administration. Each vendor has a dedicated card with preset controls, so there’s no manual categorization at month-end and no tangled offboarding when someone leaves.
On the bill pay side, another 4+ hours disappear because Ramp’s inbox surfaces only what needs human attention. Each item is pre-flagged with a reason, and large invoices route to the right approver without finance playing scheduler.
Add it up: the team automates an average of 163 hours of finance work every month—roughly the equivalent of a full-time hire’s workload, redirected to more strategic work.
Built for founders who think in systems
When Perplexity set up Ramp early on, the company made a deliberate decision that went beyond a platform choice. It was a bet that the right platform could let a small team operate like a much larger one. That decision has paid dividends. Perplexity’s finance function has managed millions in global spend across corporate cards, bill pay, expense reimbursements, and international operations without adding headcount to match.
"AI really multiplies the impact and efficiency of my team. It's freed us up to focus on more value-add tasks—partnering with the rest of the business to make sure they have the information they need to make strategic, informed decisions."
— Lauren Feeney, Controller, Perplexity
This is, in Lauren’s words, how finance should work: lucid and accountable at speed. Automation handles what automation should handle. Humans focus on judgment, context, and decisions that actually require a human mind. The team stays lean and constantly gets stronger by design.
Build finance that scales (without hiring as fast as you spend)
Key takeaways from Perplexity for other startups and hyper-growth teams:
- Design for leverage. Put controls at the point of spend, automate the coding, and keep humans in an exceptions-only loop.
- Move controls upstream. Set the rules when spend is created, not at month-end—now close becomes review, not reconstruction.
- Teach the system your context. When the Accounting Agent is wrong, correct it in plain language. Those corrections compound.
- Make signal obvious. The win isn’t just speed—it’s surfacing why something needs attention, so people spend time on judgment, not hunting.
