Anthropic Customer Research at Scale: How Claude's Maker Learns from Enterprise AI Buyers

TL;DR#

Anthropic, the maker of Claude, has become the canonical example of an AI lab that systematically researches its own enterprise buyers using AI — not just builds models. Public statements from CEO Dario Amodei and CPO Mike Krieger place enterprise revenue at the center of Anthropic's trajectory: the company reportedly crossed roughly $5 billion in annualized revenue in mid-to-late 2025, with Claude for Enterprise, Claude Code, and the Agent SDK as the three product surfaces selling into Fortune 500 buyers. In October 2025, Anthropic publicly announced its own "AI Interviewer" use case — a flagship example of using Claude to conduct customer interviews at scale, the same approach Perspective AI ships as a product. Anthropic's research practice combines Project Glasswing (its internal customer-listening effort referenced in 2025 communications), structured Claude-moderated interviews, and a deep partnership-style account model with anchor enterprise customers like Bridgewater, GitLab, Pfizer, and Notion. The meta-story is unmistakable: an AI lab whose largest customers are enterprises is using its own AI to learn what those enterprises actually want — and publicly inviting other companies to do the same. This post unpacks Anthropic's public research stance, what we know about how they run AI customer interviews at frontier scale, and what other AI-product companies can copy.

Anthropic's 2026 product surface and why enterprise research matters more than ever#

Anthropic in 2026 is no longer "the company that makes Claude" — it is an enterprise software vendor whose go-to-market depends on understanding very specific buyer requirements, from CISO security expectations to platform-team agent SDKs. The product surface buyers see today spans at least six distinct entry points, each with a different ICP, a different buying committee, and a different evaluation pattern.

Five public data points anchor why customer research at this surface area is non-trivial:

1. Revenue scale: Reuters and The Information reported Anthropic's annualized revenue crossing roughly $5B in mid-2025, up from $1B at the start of the year — a more-than-quintupling driven primarily by API + enterprise.
2. Customer count: At its 2025 developer event, Anthropic disclosed that more than 300,000 businesses use Claude across the product surfaces above.
3. Big-customer concentration: A handful of named anchor customers — including Bridgewater Associates, Pfizer, GitLab, Notion, Asana, and Zoom — drive material API consumption.
4. Model cadence: Anthropic shipped Claude 3.5 Sonnet (June 2024), Claude 3.7 Sonnet, Claude 4 Opus + Sonnet (May 2025), Claude 4.5 (late 2025) and is now on Claude 4.7 — a release cadence faster than any historical enterprise software vendor.
5. Agent surface: The Agent SDK and Model Context Protocol (MCP), both shipped in 2025, opened a third buyer (platform/agent teams) on top of the existing chat + code buyers.

That's a wider surface than most enterprise SaaS companies cover in a decade. The research question — "what do our enterprise buyers actually need next?" — is no longer answerable by 20 customer calls a quarter. It needs to be answered continuously, across thousands of accounts, by something that scales like the product itself does. For the deeper structural argument on why old research methods break at this scale, see our analysis of why the sample-size problem is finally solvable.

The "AI Interviewer" announcement and what it signaled#

In October 2025, Anthropic publicly highlighted "AI Interviewer" as a flagship use case for Claude — explicitly framing AI-conducted customer interviews as a category of work Claude was built to do. This is the announcement that put the AI customer interview category on the map and remains the single biggest external validation of the approach Perspective AI ships.

The signal mattered for three reasons. First, when the maker of a foundation model names a use case as flagship, the entire developer + buyer ecosystem treats that use case as endorsed. Second, the announcement implicitly admitted that Anthropic itself runs structured Claude-moderated interviews — the use case wasn't theoretical. Third, the framing — interviews, not surveys — was a deliberate contrast with the forms-and-surveys paradigm dominant since the 1990s.

We covered the announcement in detail and the follow-on reaction in our most-read post, Anthropic Launches AI Interviewer. That post is currently our #1 paid-conversion blog by velocity — itself evidence that buyers are searching for exactly this story. The piece you're reading is the deeper case-study companion: less about the announcement itself, more about what Anthropic actually does with AI for customer research and what other AI-product companies should copy.

The headline takeaway: Anthropic publicly endorsed the workflow of running AI-moderated customer interviews as a legitimate, citable category — and the rest of the AI industry now has air cover to do the same.

How Anthropic researches enterprise AI buyer requirements#

Anthropic's public communications describe a layered customer-research stack that combines design-partner-style anchor accounts, structured Claude-moderated interviews, model-evaluation feedback from API customers, and direct executive interviews — not a single research method. Based on what Mike Krieger (CPO) and Dario Amodei have said publicly across Stratechery, the Lenny's Podcast appearance, the Dwarkesh interview, and various Anthropic blog posts, the practice appears to have at least four distinct layers.

Layer 1 — Anchor design partners. Anthropic ships product features in deep partnership with named enterprise accounts. The June 2024 Claude 3.5 Sonnet release explicitly thanked Bridgewater, Pfizer, GitLab, Notion, and others for early-access feedback. This is the high-touch, partnership-style research layer that resembles a traditional enterprise software design-partner program — except the feedback loop is much faster because the partners are also Claude-API customers running their own evaluations daily.

Layer 2 — Structured Claude-moderated interviews. This is the layer Anthropic surfaced with the AI Interviewer announcement. The mechanics map cleanly onto what we describe as AI-moderated interviewing: a research outline, an LLM acting as moderator, automated follow-up to vague answers, and machine-driven synthesis. Anthropic doesn't have to invent this from scratch — they have Claude. The interesting question for the rest of the industry is whether you build it on Claude or buy it from a vendor that already runs the workflow.

Layer 3 — Eval and usage telemetry as continuous research. Every enterprise customer running Claude in production is generating live signal — latency thresholds breached, prompts retried, tools that fail, outputs that get edited. Anthropic's evaluation infrastructure (publicly described in posts about Claude's Constitutional AI training and post-deployment monitoring) functions as a permanent customer research channel that doesn't require a single human interview. This is closer to telemetry than research, but in practice it shapes roadmap decisions in the same way our analysis of continuous discovery habits describes.

Layer 4 — Executive interviews and on-the-record conversations. Dario Amodei and Mike Krieger regularly do long-form public interviews — Stratechery's "Anthropic and the Three-Speed Race," Dwarkesh's two-part series with Dario, Lenny's Podcast with Mike. These are, functionally, public-facing voice of customer artifacts. The audience hears Anthropic's view of where AI is going; Anthropic hears back through the questions asked and the reactions on social media. Several of these long-form interviews have been cited by Anthropic team members as influencing internal product priorities.

The point of stacking these layers is not redundancy — each layer answers a different question. Anchor partners answer "what should we ship?" Claude-moderated interviews answer "what do hundreds of customers we don't have time to meet think?" Telemetry answers "what's actually breaking right now?" Executive conversations answer "where is the buyer's mental model going next year?" Most companies running voice of customer programs try to collapse all four into one survey. Anthropic doesn't.

Project Glasswing and the meta-research story#

Project Glasswing is the internal name Anthropic has used in public communications for the practice of using Claude to study its own customers — the meta-research story of an AI lab eating its own dog food. The name surfaced in 2025 internal posts that became public, and we've written about the Glasswing principle as a customer feedback design pattern — the idea that most feedback tools have the same blind spot because they're all built on the same form-based foundation.

The meta-story is the most interesting part of the case. Three observations:

1. The maker of the foundation model is also the heaviest user of it for non-product work. This is unusual. Microsoft does not use Word to run its customer research. Salesforce does not use Service Cloud to interview its own enterprise buyers. Anthropic uses Claude to interview the people buying Claude — and publishes the result.

2. The endorsement effect compounds. When Anthropic publicly states that AI-moderated interviews are a flagship use case, every PM at every AI-product company in the world has air cover to propose the same internally. The CPO no longer needs to argue from first principles; they can cite Anthropic.

3. The buyer signal is bidirectional. Anthropic's own customer-research practice generates content (announcements, interviews, blog posts) that itself becomes research data — telling Anthropic what enterprise AI buyers are paying attention to. This is the kind of compounding research advantage we explored in the hidden advantage of the fastest-growing companies.

The Glasswing framing also matters because it gives a name to the gap most B2B SaaS companies still operate inside: they have telemetry, they have NPS, they have win/loss notes — but none of those surface intent, context, or the "why now." That gap is exactly what conversational data collection is built to close.

What other AI labs and AI-product companies can learn#

The replicable lesson from Anthropic isn't "use Claude" — it's "treat customer research as a continuous, AI-native workflow and pick a layered stack that matches your product surface." Five concrete moves any AI-product company can copy in 2026:

1. Name your equivalent of Project Glasswing publicly. The endorsement effect only fires when you put a name on it. Naming the practice (the way Anthropic named AI Interviewer) gives your team, your buyers, and your investors a shared phrase to point at. We see the same dynamic with Figma's customer research approach, which we covered in Figma's AI customer research strategy and with Notion's $10B buyer-led roadmap practice.

2. Adopt the layered stack, not a single tool. Anthropic doesn't choose between anchor partners, AI interviews, telemetry, and exec conversations — it runs all four. The stack should be matched to your product surface. A pure-API company can weight Layer 3 (telemetry) heaviest. A horizontal SaaS product like the ones we cover in the state of customer research 2026 will weight Layers 1 and 2.

3. Run AI-moderated interviews on the buyer journey, not the post-purchase survey. Most companies still use AI to summarize post-purchase NPS responses. The higher-leverage move is to run AI-moderated interviews during evaluation — when prospects are still deciding. The mechanics are the same as our JTBD interviews guide for product teams but pointed at the buying committee rather than current users.

4. Treat eval scores as customer research, not just model science. API customers vote with their evals. Watching enterprise eval scores move week-over-week is, functionally, a continuous discovery habit. Anthropic's research team and model-evaluation team are not strictly separable; the rest of the industry should stop pretending they are.

5. Publish the meta-story. Anthropic's case-study peers — companies like Shopify, Stripe, Linear, Miro, Loom, Duolingo, and Canva — all share one trait: they publicly describe their customer-research method. That publication is itself a recruiting and buyer-signal tool. Don't keep the practice private.

Perspective AI exists because of the same insight Anthropic articulated when it announced AI Interviewer: an AI-first company cannot start its customer relationships with a static web form. Forms flatten buyers into dropdowns; AI-moderated interviews let buyers speak in their own words, with follow-up, on the way to a purchase decision. The product surface Anthropic now sells — Claude API, Claude for Enterprise, Claude Code, Agent SDK, MCP — is too wide, too fast-moving, and too high-stakes for survey-based research to work. The same is true of any AI-product company building in 2026.

External grounding for the claims above: Anthropic's own announcements page is the authoritative source for product timing and customer counts; Mike Krieger's Lenny's Podcast appearance in 2025 covered the customer-research practice in some detail; Dario Amodei's interviews on the Dwarkesh Patel Podcast are the most-cited public source for Anthropic's strategic worldview.

Frequently Asked Questions#

Does Anthropic actually use Claude to run its own customer interviews?#

Yes — Anthropic publicly named "AI Interviewer" as a flagship use case in October 2025, and CPO Mike Krieger has confirmed in podcast appearances that Anthropic's product and research teams use Claude to study customer needs at scale. The exact mechanics of any internal Glasswing-style program have not been published in full detail, so it's correct to treat the specifics as inferred from public statements rather than as confirmed internal process documentation.

What is Project Glasswing?#

Project Glasswing is the internal name Anthropic has used in public communications for its practice of using Claude to research its own customers and product surface. The name has surfaced in Anthropic statements and has been picked up by industry press; the public details are limited, but the framing — an AI company using AI to study its own buyers — has become a reference point for the broader category. We cover the design principle in our Glasswing principle post.

How does Anthropic's enterprise revenue compare to its consumer revenue?#

Public reporting (Reuters, The Information) indicates that Anthropic's revenue mix is heavily enterprise-weighted, with API and Claude for Enterprise accounting for the majority of the $5B-plus annualized revenue figure disclosed in mid-to-late 2025. Claude.ai (consumer) and Claude for Work (team) contribute, but the model is fundamentally a developer-and-enterprise business — closer to AWS than to ChatGPT in its revenue concentration.

What can a 20-person AI startup learn from Anthropic's research practice?#

Three things copy cleanly at small scale: name your customer-research practice publicly, run AI-moderated interviews on prospects (not just current users), and treat your evaluation infrastructure as customer research. None of the three require Anthropic-scale resources. The full layered stack — anchor partners + AI interviews + telemetry + exec conversations — also works at startup scale; the relative weights just shift.

Is "AI customer interview" a real category or just a marketing term?#

It's a real category, anchored by Anthropic's October 2025 endorsement and now sized in the hundreds of millions of dollars across vendors and internal programs. We cover the category's adoption patterns in detail in the state of AI customer interviews mid-year 2026 update and the original state-of-the-category post. The category replaces survey-based research for buyer evaluation, ongoing user research, and continuous discovery — three use cases that combined represent a multi-billion-dollar segment.

How do I run AI customer interviews on my own buyer journey?#

Start by replacing one form with a conversation — usually the demo-request form or the post-signup onboarding form — and instrumenting an AI interviewer to run follow-up on the answers. The full mechanics are in our JTBD interviews methodology and our practical AI moderated research guide. Most teams running their first AI customer interview project see a usable insights deck within two weeks — a cycle time that's not possible with survey-based research.

Conclusion#

Anthropic, the maker of Claude, is the canonical 2026 example of an AI lab using AI customer interviews at frontier scale — and the public endorsement effect of that practice has reshaped how the rest of the industry talks about customer research. The layered stack (anchor partners, Claude-moderated interviews, eval telemetry, executive conversations) is replicable. The meta-story (an AI company using its own AI for research) is replicable. The naming move (call your practice something — Glasswing, AI Interviewer, anything memorable) is the easiest of all to copy.

If you build an AI product and your customer-research practice still starts with a survey link, you have the same blind spot Anthropic publicly named in 2025. The fix is to run AI customer interviews on your real buyer journey — at the demo request, at the trial-to-paid moment, at the renewal. Perspective AI is the product version of the practice Anthropic endorsed: AI-moderated customer interviews that follow up, probe, and capture the why behind every answer. Start a study now, explore the agents that run them, or see how teams like yours are operationalizing the workflow.