AI sales agents for small teams: lessons from OpenRouter’s Rasp

OpenRouter says its five-person sales team gets about 600 hours a month back from Rasp, an AI agent that researches leads, prepares calls and handles CRM paperwork. For a small team considering something similar, which of those jobs would be worth automating first?

In OpenRouter’s October 7 case study, Rasp deals with the administration around a sale, while people still run calls and negotiate. Start by testing lead research and routing. Give an AI sales agent permission to contact customers only after you know how often its decisions need correcting.

Follow a lead through the workflow

Rasp researches incoming enquiries, decides which belong with sales, sends initial emails and writes briefs before calls. Afterward, it drafts notes from the transcript and fills in CRM fields. OpenRouter says follow-ups and efforts to reopen quiet deals remain manual.

An outbound prospecting tool does a different job from an agent that prepares your team for people who already contacted you. Before buying either, decide where your salespeople lose time.

OpenRouter reports that pre-demo preparation fell from 30 minutes to 5. Its other task timings describe shorter note-writing and CRM updates. The individual reductions add up to 61 minutes per demo, while the overall reduction it gives is 59. Ask for the underlying timings before using the total in a staffing plan.

Consider a hypothetical consulting firm that receives a request for help with an inventory spreadsheet. An agent could find the company website, record what the prospect asked for and prepare a short brief for the person taking the call. It should distinguish the prospect’s own statements from its guesses. A company’s website mentioning three offices doesn’t establish its budget or readiness to buy.

Even a simple routing rule needs local judgment. OpenRouter says its intake form filters personal email domains. Copy that into a business serving sole traders and you could discard a perfectly good prospect because they use Gmail.

Qualification is a decision; sending is an action

An agent that recommends a reply and an agent that sends it need different permissions. OpenRouter’s tool-calling documentation explains that the model proposes a function call, and the application executes it. Your application can therefore require approval or refuse the action, regardless of what the model requests.

The Rasp case study says 93% of first-touch emails go out autonomously, but later says final sends always involve a person. Those descriptions don’t establish exactly which messages require approval. Ask a supplier to show the approval rule in the sending workflow, including what happens when the model suggests a discount or invents a commitment.

For an initial pilot, let the agent read the inputs it needs and write drafts in a review queue. Keep sending access separate. CRM access should also match the task: suggesting a company description doesn’t require deleting contacts or changing a deal’s value.

OWASP’s guidance on excessive agency recommends limiting tools and permissions and enforcing authorization in downstream systems. A prompt telling an agent to behave carefully cannot replace those controls.

Find out which customer information reaches each service, who can retain it and which staff can see it. OpenRouter’s privacy documentation separates its own handling of requests from the model providers involved; a CRM connection or meeting transcript brings further policies to check. Don’t assume one reassuring privacy statement covers the whole workflow.

What the $30-a-day figure means

OpenRouter reports Rasp’s cost at roughly $30 a day, including engineering time, compared with nearly $800 a day for earlier agents doing the same scope of work. It puts roughly $18 of Rasp’s daily spend on model inference.

OpenRouter’s internal cost figure doesn’t establish a public subscription price or what another business would spend building the system. The article doesn’t separate the initial build bill in enough detail to make that calculation. As of October 9, 2026, Ori’s product page also describes the Slack Intern product as waitlisted and in private testing.

Include setup and CRM integration in your budget, then measure the staff time spent checking answers and repairing mistakes. A cheap response that requires a rep to redo the research may cost more than a slower, accurate draft.

OpenRouter reports a 2.6-fold rise in close rate over the same period, but says pricing and market conditions also changed. Its comparison cannot isolate Rasp’s contribution to that increase.

The abstract of Generative AI at Work, a study of 5,179 customer-support agents, reports productivity gains that varied sharply with worker experience. It studied support assistance, not autonomous selling, so its results cannot supply a sales ROI forecast. Measure review time separately for the people who will use your agent, rather than assuming the team average applies to everyone.

Start with a pilot you can judge

Choose a batch of past enquiries with known outcomes. Ask a salesperson to record the correct routing and the facts a useful brief should contain before looking at the agent’s answers. Include support requests and ambiguous enquiries alongside qualified leads; testing only obvious prospects will hide the errors that cost you business.

Run the agent without sending messages. Track:

• Correctly routed enquiries, including genuine prospects it wrongly rejects.
• Unsupported claims in briefs and proposed replies.
• Rep time spent reviewing and correcting each result, compared with the existing process.
• Total cost per correctly handled enquiry, including human review.

Agree what would stop the pilot before connecting live actions. An invented contractual promise should block a send; an uncertain company match should ask for review. For a low-volume team, the integration effort may outweigh any time saved, even when the model performs well.

Build your first pilot around the task that consumes the most avoidable rep time. Expand its authority when your results justify it.