The hardest part of evaluating an AI agent isn’t the demo — it’s the math after the demo. So it’s useful when a company publishes its own numbers, including the ones that complicate the story. That’s what OpenRouter did on October 7, 2026, in a case study of Rasp, the AI sales agent running on its own five-person sales team, built on their Ori platform.
What the agent actually does
Rasp is deliberately unglamorous. It researches and qualifies inbound leads, sends first-touch emails (93% without human help, with the tricky 7% flagged to a person), writes pre-call briefs, drafts post-call notes from transcripts, and fills in most CRM fields. Reps still run the calls, negotiate, and decide anything sensitive. As the case study’s authors put it, Rasp is a “street sweeper” — it clears the admin debris so reps can move faster.
The measured result: a demo that used to take about 103 minutes of total work now takes 44. Prep dropped from 30 minutes to 5, notes from 15 to 2, CRM updates from 30 to 7. Across the team, that’s roughly 600 hours a month, or 28 hours per account executive.
The numbers worth copying into your own ROI model
A few figures here are more useful than the headline:
- Two more calls per rep per day, at the same self-reported preparation level — that’s the mechanism behind the hours, not just a correlate.
- Deal cycle compressed 34%, and about 57% of the current pipeline is attributed to capacity the team didn’t previously have.
- Close rates rose 2.6x, but the authors flag this honestly: pricing changes and market conditions moved in the same window. Not all of it is the agent.
- Cost fell from ~$800/day to ~$30/day — but note the comparison is against Rasp’s predecessors, Ace and Dove, not against human labor.
That last number deserves scrutiny. The $30 includes engineering time plus outputs, with roughly $18 of it being GLM 5.2 inference. The routing layer keeps swapping in cheaper models that hold quality, and OpenRouter validated three options that could push routine inference down to $0.50 a day.
Lessons from two failed-ish predecessors
Rasp is the third sales agent OpenRouter built, and the design choices that matter came from watching the first two break:
- One pipeline, task type as a parameter. Per-task code paths made the earlier agents drift and become unwieldy.
- A feature flag on everything, off by default. They had shipped things before that couldn’t be cleanly turned off.
- Metrics logged from day one. Every number in the post exists because the logging existed before launch.
- Slack as the interface. Dashboards store metrics; reps consume them where the work happens.
If you’re planning a GTM agent, that list is a better checklist than most vendor feature matrices.
Where routing does the compounding
The cost story here is quietly the most builder-relevant part. Ace and Dove cost nearly $800 a day; Rasp does the same scope of work for $30, mostly because the layer underneath keeps searching for a cheaper model that maintains quality. That’s a real argument for treating model selection as infrastructure rather than a one-time decision — the same reasoning behind routing benchmarks that measure the layer above the model itself (see A Leaderboard for the Layer Above Your Model).
The honest limits
Rasp doesn’t do follow-ups, dropped-deal detection, or negotiation judgment — some by omission, some by design. Every send, deal classification, and compliance call keeps a human in the loop. And the authors note Rasp “is likely not the final state,” which is the right posture for anyone shipping agents.
The takeaway for your own build: target the always-on drudgery first, instrument before launch, and put routing under the agent so costs fall after you ship instead of rising.