Introducing Stack.
The AI operating system that lets accounting firms take on more clients without hiring. Learns your firm's process, runs the close, posts the journals. Fully auditable.
We’re living through the biggest shift in accounting since the spreadsheet.
We launched the Software Adoption workflow in Perplexity Computer. It uses aggregated industry-wide Ramp credit card spend to provide data on the adoption, growth momentum, churn, and market share of SaaS products.
We are pleased to partner with @tryramp to connect Ramp Rate API data to Computer in this workflow, at no additional charge to users, and no setup required. Just run the workflow and ask for whatever SaaS categories or vendors you want to analyze.
Ramp Data is now LIVE in Claude, ChatGPT, Bloomberg, Perplexity, and Grok. Ask what 50,000+ businesses are paying for software or where adoption is shifting and you’ll get answers based on real spend data.
It's free, public, aggregated, and anonymized. Enjoy, data lovers.
The problem Ramp solved with RL is actually a very real agentic retrieval problem that most people building agents run into.
General purpose models are extremely eager and tend to over fetch. They keep making tool calls, trying to retrieve more and more context "just in case", and a lot of the retrieved data ends up being irrelevant noise. Ramp’s own traces showed ~17.8% of all tool calls were wasted on exploration, and ~75% of retrieval calls were immediately followed by another read because the model failed to fetch the right information the first time.
That’s why their decision to train a specialized retrieval subagent is interesting.
Instead of using a giant reasoning model to navigate spreadsheets and filter irrelevant information, they fine-tuned a smaller Qwen model specifically for retrieval using RL. The result was not just lower latency, but actually higher accuracy than Opus 4.6 on their evals while running at much lower latency.
The hard part of agentic systems isn’t always reasoning anymore, it’s knowing what NOT to retrieve and preventing context from turning into garbage. Context is a scarce resource, and over-fetching hurts latency, cost, and even accuracy because the model starts anchoring on irrelevant information.
We partnered with @PrimeIntellect to build Fast Ask, a small RL-trained subagent that helps our Sheets agent find answers in spreadsheets. It scores +4% over Opus on exact match accuracy at Haiku latency.
This man has three Olympic golds and we somehow convinced him to speak at a conference about the future of autonomous finance.
Ramp customer @shaunwhite will be our keynote speaker at OnRamp 2026, in partnership with @thesnowleague.
98% of companies don't have a procurement team. The ones that do are stretched thin. Today, they all get backup.
Introducing a suite of AI agents to run your entire purchasing process, saving you 46 hours of manual work per month and 16% on yearly vendor spend.
I've noticed "most" companies founded in 2022 are quite different in how they organize/build than the ones founded in 2024.
If you were founded pre-2024, Ramp built a really thoughtful playbook on how to get your whole org ai pilled.
Ramp is building toward zero-touch AP.
OCR recognizes invoices, but your team is still doing manual corrections. Our AP agent now remembers how you process them and applies that knowledge to every bill.
Over the past three months, weekly active users on the Ramp MCP has grown 10x as more customers reach into the product through Claude, ChatGPT, and other agents.
Sharing some learnings as we’ve scaled this product below (thread 👇)
AI token spend across Ramp customers is up 13x. Bills are spiking overnight and no one is noticing until the invoices hit. So our team built a solution.
Now with Ramp AI Spend Intelligence, you can track your spending down to the token, team, and model, all in real time.
Last week @sebgoddijn shared Glass, Ramp's internal AI productivity tool, with the world. Nearly a million views later, everyone's asking the same thing: how did you actually build this?
Read the full story here:
https://t.co/gzO1X4V2uN
The short of it: we built an app that builds itself
A small team of us pretty much vibe coded the entire thing in about a month. The trick was teaching Glass how to improve its own codebase.
Turns out if you discipline your AI agents well enough, they grow up to be high-functioning, self-sufficient adults in no time.
we gave agents cards, a CLI, and now telepathy
today agents share context by converting everything to tokens — slow, expensive, lossy. our research team built a way to skip that entirely. agents share relevant memory directly, cache to cache. 31% cheaper, no accuracy loss.
Introducing Latent Briefing, a way for agents to quickly share their relevant memory directly. Result: 31% fewer tokens used, same accuracy.
Multi-agent systems are powerful, but can be wildly inefficient. They pass context as tokens, so costs explode and signal gets lost. We built an algorithm that allows agents to communicate KV cache to KV cache.
Introducing Ramp AI Spend Intelligence. A way to track and control spend all the way down to the token level across providers.
Tokenmaxxing requires tokentracking. This is how your business will get ahead (or fall behind) with AI.
The best companies are investing aggressively in AI. And it’s working. Our data proves it. Since January 2025, average monthly AI token spend across Ramp customers increased 13x. Not 13%. Thirteen *TIMES.
But visibility into that spend has remained virtually nonexistent. No one knows how fast it’s adding up, where it’s going, who’s purchasing what. It’s a financial mess.
And AI spend is fundamentally different from every other cost category. A single prompt template change can triple your bill overnight. An agent stuck in a loop can burn through $50k before anyone notices.
No tool in the market connects token-level usage, invoice data, and card-level spend. So we built one.
Ramp customers can get early access today!
https://t.co/fFQApIZ3Pz