Anthropic's Sales AI | Conversion Rate Doubles, Deal Cycle Cut by 5 Days
機械翻訳 / Machine-translated
A
AI Friends
3時間前・22分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever submitted a contact form on a company's website and waited three days for a reply?
As it turns out, AI company Anthropic faced exactly the same problem. This article breaks down how the company rebuilt its sales intake with AI, what results it achieved, and what Japanese companies can borrow from the playbook.
On September 30, 2026, Anthropic published a case study from its own sales team on the official blog.
The author was Carl Johnson, Head of Sales Development.
The subject was a full rebuild of inbound sales — handling inquiries that come in from prospective customers.
The old process worked like this: someone who wanted to use Claude at their company filled out a "Contact Sales" form on the website. That submission was routed to a sales rep, who replied by email.
The problem? Inquiries numbered in the tens of thousands every month. Responses took time — sometimes several days.
And most of the questions were simple ones: pricing, minimum seat counts, security requirements.
Imagine a popular restaurant with a long checkout line where every customer just wants to ask, "What time do you close?" That was essentially the situation.
The company's answer was a Buying Agent.
An AI agent — an AI that reasons and acts on its own to achieve a goal — responds to inquiries on the spot. It guides prospects from an initial "I want to use Claude at my company" conversation all the way through to completing a purchase. It now handles thousands of conversations per day.
Each interaction follows this sequence:
From there, one of three things happens: the prospect proceeds directly to purchase; a complex case is handed off to a human rep with a summary of the conversation; or the interaction ends after the question is answered.
Imagine a 20-person startup. The founder opens the pricing page at 11 p.m. and asks, "Which plan works for 20 people?" The agent asks about their use case and security preferences, then recommends a plan and seat count. The founder can complete the purchase that same night.
Under the old system, they would have submitted a form and waited days for a reply.
The agent is available in four places: the sales contact page, the pricing page, inside the Claude product itself, and via email.
Crucially, the system is opt-in — customers choose upfront whether they want to talk to AI or a human rep.
Here are the results the official blog reported:
| Metric | Result |
|---|---|
| Rate of advancing to a qualified meeting | More than 2× vs. traditional form |
| Time to close | ~5 days shorter |
| Share of conversations requiring human assistance | Cut roughly in half |
| Email back-and-forth for one rep | ~10 exchanges → ~6 |
| Closed deals for that same rep | 2.5× |
The top two rows compare prospects who spoke with the agent before being handed to sales against prospects who came through the traditional form.
Prospects who went through the agent arrive at the rep's desk already understanding what they're buying — which is why conversations move faster.
Some coverage led with "deals closed at 2.5×" as the headline.
But reading the original blog post, this is the result for one inside sales rep, Ojas — not the team overall.
He used to exchange roughly 10 emails to close a single deal. Now it takes about 6. As a result, his personal deal count is 2.5× what it was before the tool launched.
The blog does not say the entire team's closed deals rose 2.5×. Keep that in mind if you cite this internally.
It's also worth noting that all of these figures come from Anthropic's own announcement — they have not been independently verified by a third party.
The Buying Agent runs on Claude Managed Agents.
This is a service where Anthropic handles the infrastructure work required to run AI agents. The public beta launched on April 8, 2026.
Pricing is the standard API usage fee plus $0.08 per active hour. Idle time is not billed.
According to the blog, the agent is made up of "a prompt, a handful of tools, and Claude."
The first version was built by one engineer in a few weeks.
The knowledge base draws on internal documentation and support articles, enabling the agent to answer questions about pricing, security, data handling, and more.
The improvement process is worth noting.
Engineers manage the code. But the prompt is edited directly in a management interface by sales and content leads — no engineering required.
Changes are tested in a staging environment before going live.
Within about a week of internal testing, the agent had reached version 7. The team reviews and updates the prompt every week even after launch.
Providing a single high-level objective worked better than listing out detailed steps.
The actual line used: "Your goal is to understand customer requirements, qualify prospects, and recommend the best plan."
It's like onboarding a new hire. Telling them "find the best fit for the customer" leads to more adaptive behavior than handing them a thick manual to memorize.
Simpler prompts with supporting context outperformed complex, heavily constrained ones.
In other words, the more "don't do this, don't do that" rules you add, the harder it becomes for the AI to act effectively.
When the agent hands a conversation to a human rep, it explains why.
The team reviewed each reason individually and treated it as a signal for improvement.
As a result, the share of conversations requiring human help dropped by roughly half.
The agent frequently recommends the Team plan over the Enterprise plan for smaller companies.
Anthropic describes this as "a feature, not a bug." The philosophy: giving customers the answer that actually fits them is what builds trust.
Several services already offer AI-powered inquiry handling.
| Service | Key Feature | Best For |
|---|---|---|
| Anthropic Buying Agent (in-house) | Custom-built on Managed Agents; guides from inquiry to completed purchase | Companies that can build in-house |
| Qualified "Piper" | Handles chat, voice, and email with site visitors; automates meeting booking | Salesforce-heavy companies |
| Salesforce "Agentforce" | Deep CRM integration; pricing around $2/conversation | Large enterprises running on Salesforce |
| HubSpot "Breeze" | AI agent suite bundled into HubSpot subscriptions | SMBs and mid-market on HubSpot |
| immedio (Japan) | Automates rep assignment and meeting scheduling immediately after inquiry | Domestic B2B companies |
Most products are designed to get prospects to book a meeting. Salesforce, for example, has published that it deployed Piper on its own site in 30 days and started booking 60+ meetings per week.
Anthropic's Buying Agent goes further — it guides prospects all the way to completing a purchase. In some cases, the deal closes without a human rep ever getting involved.
Anthropic has another relevant case study. A sales rep profiled in August 2026 was spending about five hours a day responding to inquiries.
She automated first-draft replies using Claude Cowork (a feature that lets Claude perform tasks directly on a computer) — while keeping a firm rule: a human reads every reply before it's sent.
The contrast is instructive: AI assisting a rep means "humans always review before sending." AI talking directly to customers means "opt-in, with complex cases handed to a human."
Claude Managed Agents is accessible from Japan via API. Rakuten is among the early adopters named publicly.
Claude handles Japanese natively, so a Japanese-language inquiry interface is entirely feasible.
At roughly ¥12 per active hour plus API costs, it's easy to start small.
Consider a common B2B scenario in Japan. At 6 p.m. on a Friday, a procurement manager at a manufacturing company submits a pricing question to a SaaS company's contact form.
The reply arrives Monday afternoon. By then, the manager has already been in talks with a competitor who responded right away.
With an AI agent at the front door, pricing and eligibility questions can be answered Friday night. Monday morning starts with a prospect who already understands the offering.
There is a real risk that AI will quote incorrect pricing or terms. A safe starting point: let AI handle Q&A only, and keep contract finalization in human hands.
No. It was built for Anthropic's own sales intake and is not sold as a standalone product.
However, the underlying Claude Managed Agents platform is available in public beta, so you can build something similar yourself.
The blog describes a shift in how reps spend their time — more on educating prospects in early stages of consideration, in-person conversations, and event participation. Headcount reduction is not mentioned.
It's accurate for one individual rep. It is not a team-wide figure.
The team-level outcomes reported are: qualified meeting rate more than doubled, and time to close shortened by roughly 5 days.
That hasn't been disclosed. The tools used internally and any external system integrations are also not described.
Building it from scratch requires developers. If you don't have engineers in-house, an off-the-shelf solution like Piper, Breeze, or immedio is more realistic.
Start by looking at your own contact form and measuring how many hours it takes to send a first reply.
This article is a cross-post from AI Friends.
