Ramp, the corporate card and spend-management company, is opening a product called Glass to outside customers, and the story of why it exists comes entirely from its chief executive. Eric Glyman announced it in a post on X on 13 October. No press release, outside reporting or customer account accompanies it, so every figure below is Glyman’s, describing Ramp’s own operations.

Glass began as an internal system, which is the more interesting half of the post. Glyman says that in one week in May Ramp spent $1.5 million on tokens (the units AI models bill by), a figure he says his finance chief did not want shared. A small team then built a place where every employee started with the setup Ramp’s most enthusiastic AI users had built for themselves. By his account it became the way the company works within weeks. Only at the end does the post turn to customers: Ramp is “onboarding the first businesses personally” through a sign-up page at ramp.com/glass. It gives no price, no date and no list of who has access.

What it does is simpler than the label suggests. Glass connects to a company’s existing tools (Glyman names Linear, Notion, Google, Slack and Gong, “and dozens more”) and lets staff save a prompt or workflow so a colleague starts from it instead of from a blank chat. Sessions can be shared, so a support agent stuck on a ticket can hand the live session to an engineer who sees everything she did. Every project carries a running cost meter with limits. Glyman’s own example: he asked Glass for the number of British customers Ramp has signed since it launched, and the answer already excluded American firms that merely have an office in Britain, a rule the finance team had taught it months before.

The three numbers are not equally solid. The first is 6,500+ edits to “what Glass knows,” made, Glyman says, by 323 employees across product, support, sales, marketing and analysis. That counts inside Ramp only, with no period beyond “so far,” and it measures participation. It says nothing about whether the answers improved. The post also does not give Ramp’s headcount, so 323 cannot be read as a share of staff.

The feature-request figure is the one with an outcome behind it, and the post does supply numbers. Staff once keyed these into Salesforce manually, about 300 a month. After an account manager in May stored a prompt that extracts them from customer calls, Glyman says Ramp now logs 20,000 a month, roughly 67 times as many. Two more prompts prioritize them and email customers when something ships, with more than 5,000 of those emails sent. The post does not say how many of the 20,000 are distinct requests or how many a person acted on, and a prompt that logs every mention will log a lot.

The 30% saving is the claim hardest to weigh. Ramp Router chooses a model for each task before the work is sent off, and Glyman says that when Ramp moved its own work onto it “our bill fell 30% overnight.” Sending easy tasks to cheaper models is a sound, established way to cut spending, so the direction is credible. The baseline is unstated: the post does not say whether the starting point was the $1.5 million week, a typical month or something else. It also does not say whether Router is sold separately. Ramp is the company reporting the saving and the one selling the tool.

Glyman’s thesis is that handing every employee the best models is not the same as becoming an AI-native company, and that value leaks away when a good prompt sits in a single employee’s private chat log. That is a reasonable diagnosis, and the cost meter suggests Ramp felt it in its own budget first.

A buyer weighing Glass should ask for a pilot measured on the buyer’s own token bill and request volume, because the only 30% on offer so far is Ramp’s.

Based on a post by Eric Glyman, Ramp’s chief executive, on X, 13 October 2026.