From purchase requests to 409 API keys: How AngelList put spend under owner-level control
“AI is moving faster than the finance context around it. Prices change, models change, and the value is not always obvious from an invoice. We needed enough detail to know which bets deserved more investment — and which ones did not.”
Controller, AngelList
Greg Cooley could see AngelList’s AI bill. That was the easy part. What he could not see was what mattered: which key, team, agent, or model had created the cost and whether anyone still needed it running. The invoice arrived each month as a single number, but behind it sat hundreds of decisions finance had not made.
AngelList was not watching the AI shift from the sidelines. When the company launched Link in July 2026, it made live AngelList fund data queryable from any AI tool. Fund managers could ask questions in plain language and receive current, source-backed answers about capital calls, LP commitments, portfolio positions, distributions, financials, and documents. The product built on behavior already underway: leading fund managers were drafting LP letters with AI and building dashboards without an engineering
team.
Inside AngelList, that shift had already happened. It had also created a new kind of finance problem: spend that moved quickly, lived close to engineering, and arrived in accounting without the context needed to govern it.
Every dollar had a vendor. Not every dollar had an owner.
AngelList had solved a version of this problem before, at a much smaller scale.
The company offers employees generous reimbursements for expenses such as home internet, home-office equipment, and wellness. With roughly 155 employees, those benefits generated nearly 600 reimbursement requests each month.
Travel and other reimbursable expenses added another 500, bringing the monthly average to roughly 1,1100 requests, each one paid out of pocket, submitted with a receipt, reviewed, and reimbursed.
Finance inspected a steady stream of receipts and manually reconciled submissions that looked nearly identical month after month. Meanwhile, employees waited to recover money they had already spent.
Corporate-card spending sat in a separate system, with its own feed of transactions to review, leaving finance to reconstruct employee spend across two systems before it could see the full picture.
Procurement followed a similar pattern. A request could pass through the appropriate approvals in their procurement workflow tool, only to stall when finance needed to turn that decision into a purchase order, and ultimately, a payment.
“A purchase started in our procurement workflow tool, but finance had to recreate it and all its supporting documents in our payments system before it could be paid. One purchase became two workflows, two sets of approvals, and a reconciliation nightmare.”
— Greg Cooley, Controller, AngelList
AngelList learned that a patchwork of systems could turn ordinary finance work into a chain of reconciliations and follow-ups.
That lesson became more consequential as AI experimentation spread across the company. Finance could see the total bill, but not what AngelList was actually buying: which team, key, agent, or model was driving the spend, and which experiments had quietly become recurring costs.
“As token pricing and model capabilities climb, we’re hitting an inflection point: not every task needs to run on the most expensive model. Not every task is a frontier-level problem. We needed a way to match the cost to the job.”
— Greg Cooley, Controller, AngelList
One workflow from request to payment
AngelList started by changing the employee experience. Finance issued virtual Ramp cards with preset limits for internet, home-office, and wellness expenses. Purchases were automatically coded, turning a recurring employee benefit into a controlled company spend instead of an out-of-pocket reimbursement.
Next, AngelList brought Bill Pay into the same view. Finance could manage vendor invoices and payments alongside employee card activity, with data syncing directly to NetSuite instead of requiring a separate reconciliation.
Procurement was the next layer. AngelList had sophisticated requirements for how purchases should be reviewed: depending on the request, legal, security, and engineering might all need to weigh in alongside the requester’s manager.
As an early Ramp Procurement partner, AngelList worked directly with Ramp to bring that level of control into the platform. The teams tested workflow iterations until approvals could run in parallel, appear only when a purchase required them, and carry through to payment and accounting. The result was a procurement process built around AngelList’s operating model—not another approval tool sitting beside the rest of finance.
The final layer was AI spend. AI Token Spend Management gave finance and engineering a shared vocabulary: key, owner, model, purpose, and cost. Instead of asking whether the AI bill was high or low in the abstract, they could ask whether a specific workflow justified the model—and the spend—behind it.
“We weren’t looking for another approval tool. We needed one workflow from request to purchase order to payment. Ramp worked with us to fit complex, cross-functional approvals into that workflow. Now our AI spend can sit alongside it.”
— Greg Cooley, Controller, AngelList
The bill was never the problem. Attribution was.
The first find was prompt caching, surfaced through a spend briefing rather than a review meeting.
"The briefing surfaced prompt caching, which wasn’t on my radar as a Controller. I sent it to engineering immediately. We found we could save up to $10,000 a month with the same quality and performance, and the fix was implemented the same day.”
— Greg Cooley, Controller, AngelList
Within a month, cache-read shares rose from 51% to 92% and began immediately generating recurring monthly AI spend savings. Cache read shares have remained above 80% since.
Two questions for every key
The second discovery was a key with spend no one could explain. Greg took it to the engineer who owned it and asked what it did. The answer: it was an early experiment, still running, but no longer used. Roughly $5,000 a month was continuing on autopilot.
“AI has made it possible for anyone to build powerful experiments and automations. Finance’s challenge is making sure the spend does not outlive the value. A workflow can continue consuming tokens long after its original purpose has disappeared. So we asked: who owns it, what value is it creating, and are we still using it?”
— Greg Cooley, Controller, AngelList
The team shut the workload down. Together, prompt-caching improvements and the retired workload reduced AngelList’s monthly AI spend by 17.5%. Across Ramp customers, the average reduction is 12%.
What discipline buys: more bets, not fewer
AngelList now applies the same discipline before other company spending ever becomes a bill. Since the company adopted Ramp Procurement, 575 purchase requests have produced 507 approved purchase orders. Roughly one in three workflow steps clears without anyone touching it, automating an estimated 19 hours of routine procurement approval work each month.
Ten requests representing more than $63,000 in proposed spend were declined before becoming a purchase order or payment. Two were for software AngelList was already paying for through an existing plan.
With cards, bill pay, procurement, and AI spend in one place, Greg’s team can devote more attention to experimentation.
When AngelList noticed an unexpected increase in accounts receivable, accounting used its internal AI tools to investigate. Historically, finding the root cause would have required a series of meetings across finance, engineering, product, and customer service over several days or even weeks.
Instead, the team traced the problem to the underlying code, drafted a fix, and submitted a pull request to engineering for final review—all within hours of discovering the issue.
The value of AI spend management was not that every experiment became cheaper. It was that every experiment became accountable enough to keep funding.
For AngelList, that distinction matters. A company building for an AI-native version of fund operations cannot afford to slow its teams down. But it also cannot allow hundreds of keys to become hundreds of unexamined commitments. Greg’s job is to keep those two truths in the same system.
