Anthropic's sales team built a buying agent on Claude Managed Agents, more than doubling lead-to-opportunity conversion and closing about five days faster

The buying agent replaces the form-queue-rep inbound flow with an always-available conversation that can end in purchase, hand-off, or a quick answer, and it answers questions about pricing, seat minimums, and HIPAA contract requirements instantly. Previously there were tens of thousands of inbound requests a month that the BDR team could not keep up with, and customers sometimes waited multiple days; the agent now meets customers where they already ask questions and lets them choose at the start whether to talk to an agent or a sales rep. The text states deployment locations, the volume ("thousands of conversations a day"), the three conversation endings, and the opt-in design; this is a first-party operational account without a controlled experiment or statistical test.

Leads escalated by the agent are better qualified: they turn into opportunities more than twice as often as leads from the old form and close about five days faster, while the share of conversations needing a person to close has fallen by about half. Under the old process reps spent their days answering questions the docs already covered and could not reach everyone in the queue; now a rep starts the conversation already knowing what the customer needs, what they have been told, and how best to help. The text gives direction and magnitude through phrases such as "more than twice as often," "about five days faster," and "fallen by about half," plus one named rep example (Ojas: about 10 emails down to about six, 2.5x output on closed won deals); all are first-party self-reported figures with no stated sample size or statistical method.

On the engineering side the agent is simple, "a prompt, a handful of tools, and Claude," running on Managed Agents, which handles hosting, session management, and tool orchestration; one engineer built the initial version in a few weeks. The team spent its time on the prompt, the tools, and the knowledge base rather than infrastructure, and sales and content leads could edit the system prompt directly in the Console, with changes going to a staging agent first. The text lists concrete reasons for choosing Managed Agents: fast to production, tech and non-tech contributors, focus on the domain rather than the harness, versioning that lowers iteration cost (v7 about a week into internal testing, weekly prompt changes after launch), and scheduled runs as a path to other engagement types.

The team distilled transferable lessons for building agents: give a goal rather than rules, keep prompts minimal ("less is more"), put subject-matter experts in the development loop, optimize for what is right for the customer (the agent often points small teams to the Team plan instead of Enterprise), and treat every escalation to a rep as feedback. These lessons came from live iteration; for example, early escalation reasons were mostly things customers could not yet do on their own with self-service products, and those reasons shaped improvements to the experience. This is a first-party practice summary illustrated with specific examples, such as the goal statement "Your goal is to understand customer requirements, qualify prospects, and recommend the best plan" working better than a flowchart of qualification rules, but no controlled comparison is reported.

The article is aimed at sales, support, and engineering teams designing customer-facing agents, especially organizations handling large volumes of repetitive inbound inquiries that still want humans on complex deals. The described approach fits a setting where customers opt in to talking with an agent or a sales rep, the agent is deployed where customers already ask questions, and it can either guide checkout directly or hand off to a rep with the full conversation details. Scheduled runs are mentioned as a path to exploring other forms of customer engagement.

The text does not state the measurement basis, time window, sample size, or control group, so figures such as "more than twice," "about five days," "about half," and "2.5x" should be read as this team's directional observations rather than causal conclusions that generalize. The agent is in beta, and the results are tightly coupled to Anthropic's own product, documentation, and sales organization, so whether other organizations would see similar outcomes under their own conditions remains an open question. In addition, the claims about customer preference and better understanding of what they were buying come from team observation, and the text does not describe how they were measured.