Short answer
Bebop is an AI sales agent built for small and medium-sized businesses by Blackpearl Group, a Wellington technology company listed on the NZX and ASX. It turns a description of what a business sells into researched lists of likely buyers, with playbooks for approaching them. Levy Street engineers work with Blackpearl’s AI team on Bebop, and Levy Street’s CEO is Blackpearl Group’s Vice President of AI.
- Owned and sold by Blackpearl Group; not a Levy Street product.
- Launched in March 2025; Bebop 2.0 rebuilt the prospecting agent in August 2026.
- Real Bebop.ai queries became the task set for the GTM-Bench agent benchmark.
- Product owner
- Blackpearl Group (NZX and ASX: BPG), Wellington
- What it does
- Finds in-market buyers, builds prospect lists and writes sales playbooks
- Launched
- March 2025; Bebop 2.0 released 26 August 2026
- Levy Street’s role
- Engineers working with Blackpearl Group’s AI team
- Related research
- GTM-Bench, built from 59,881 real Bebop.ai queries
What Bebop is
Bebop is the AI sales product of Blackpearl Group, a Wellington-headquartered data and AI company listed on the NZX and ASX that sells into the United States small and medium business market. Blackpearl describes it as “an AI sales agent purpose-built for small and medium-sized businesses”, one that generates tailored outbound sales strategies and helps businesses find and convert new customers.
A seller describes, in plain language, what they offer and who they want to reach. Bebop’s agent researches the market, builds a list of likely buyers with the evidence behind each one, and writes a playbook covering who to contact, what to say and the objections to expect.
How Levy Street is involved
Levy Street engineers work with Blackpearl Group’s AI team on Bebop, and Levy Street’s CEO, Max Polaczuk, is Blackpearl Group’s Vice President of AI. The product, the brand and the customer relationship are Blackpearl’s. This arrangement is a version of the forward deployed model we run with other organisations: engineers working inside the operating team, on its systems, against its numbers.
Timeline
| Date | Milestone (Blackpearl Group’s public statements) |
|---|---|
| FY25 Q4 | Bebop in beta |
| March 2025 | Official launch |
| June 2026 | GTM-Bench published, using Bebop.ai queries as its task source |
| 26 August 2026 | Bebop 2.0: “a deep rebuild of the prospecting agent at its heart” |
The engineering problems underneath
Sales prospecting agents fail in predictable ways, and building Bebop is a running education in them:
- Volume versus precision
- An agent can always return more rows. The hard problem is returning fewer, better ones, each with evidence a salesperson can check.
- Grounding
- Every claim about a company or contact must trace to a source, or it becomes a hallucinated lead that wastes a sales call.
- Intent
- Understanding what the seller actually offers, and who genuinely buys it, matters more than search breadth.
- Speed and cost
- Research agents can spend minutes and dollars per request. Showing progress and controlling spend are product features, not afterthoughts.
Those lessons fed directly into GTM-Bench, which scores weak leads as negative, and into how we design AI agents for other clients.
What it means for other businesses
If you are considering an AI agent for sales research, the useful question is not “which model?” but “what data can it reach, and how will we know when it is wrong?” Our guide to designing reliable AI agents and the analysis in AI agents for go-to-market research cover how to evaluate one before you trust it with your pipeline.
Written by the Levy Street team in Wellington. Last updated 19 October 2026.