Case Study

The Commoditization Trap

Why customer support is Anthropic's make-or-break move

Kevin Nguyen · · 8 min read

Anthropic makes Claude — the AI brain powering customer support, code generation, and research for 300,000+ businesses. But what happens when every AI brain becomes equally smart? A synthesis of the strategic case for moving up the stack before the window closes.

Key Takeaways

The strategic fork

Every infrastructure company eventually faces the same question: do you stay invisible under the hood, or do you own the product that customers actually use?

Think of it like this. Anthropic is a brain surgeon — the best in the world. But right now, other companies are hiring the surgeon, putting their own name on the practice, and keeping the patient relationship. Anthropic does the hard work; everyone else captures the value.

This isn't hypothetical. It's happened before:

The pattern: infrastructure companies that don't move up the stack get commoditized when the technology matures. For AI, the maturation is happening faster than anyone expected — because every competitor (OpenAI, Google, Anthropic) is racing toward the same finish line.

The core tension: When Claude, GPT, and Gemini all score similarly on benchmarks, the model stops being the differentiator. What matters is who owns the workflow, the data, and the customer relationship.

Why customer support is the move

There are dozens of enterprise categories Anthropic could chase. Customer support is the right one for five compounding reasons:

1. It's the largest proven AI use case

Customer support is a $50 billion market heading to $95 billion by 2031. These aren't projections based on hype. Companies are already spending:

Enterprise buyers are already comfortable writing 7-figure checks for AI customer support. Anthropic doesn't need to educate the market — it needs to redirect spending that's already flowing.

2. The economics expose the gap

This is the most important comparison in the analysis:

Revenue per customer interaction
$0.03
Anthropic
(API tokens)
300×
Revenue gap
$10
App Layer
(per resolution)
Anthropic provides the intelligence but captures pennies. Application-layer companies like Sierra and Decagon capture dollars on the same interaction.
Dimension API Layer (Anthropic) App Layer (Sierra, Decagon)
Revenue per interaction $0.03 $10
Switching costs Low — easy to swap models High — deep integration
Pricing power Declining Increasing
Data ownership None Full customer data

3. The data flywheel is the real prize

Every support interaction generates valuable signal: what customers complain about, what products fail, what language works. This data compounds in three ways:

  1. Model improvement: Fine-tuned models trained on a company's support data outperform general models on that company's tickets
  2. Cross-functional intelligence: Support data feeds product roadmaps, sales forecasting, and marketing
  3. Switching costs: An AI agent that's processed 2 million of your interactions and learned your brand voice is extraordinarily hard to replace

Today, Sierra and Decagon capture this flywheel. Anthropic provides the base intelligence but gets none of the compounding value.

4. The window is closing

Application companies are actively building "model swap" layers — architectures designed to make the foundation model interchangeable. Decagon already uses multiple models. Sierra's Agent OS can swap between providers. Zendesk and Salesforce support multiple AI backends. Their customers never know or care which model powers the response.

5. Cowork + MCP makes it possible now

Before January 2026, Anthropic had no credible path to the application layer without building a standalone product. The Cowork plugin architecture changes this: a CS plugin is a set of markdown skill files and MCP connectors — not a multi-year product build. Companies already paying for Claude Enterprise can activate CS workflows for zero incremental cost.

What happens if Anthropic doesn't act

Commoditization isn't sudden. It's a gradual erosion that compounds quarter over quarter:

The four phases of commoditization
Phase 1
Now
Model Abstraction
Apps route queries to whichever AI is cheapest. Anthropic's brand becomes invisible.
Phase 2
6–18 mo
Data Moats
Apps fine-tune proprietary models on customer data. General Claude loses its edge.
Phase 3
12–24 mo
Pricing Squeeze
With model-swap capability, apps negotiate prices down. Margins compress.
Phase 4
24–36 mo
Ecosystem Lock-Out
Buyers evaluate the app, not the model. Anthropic competes on price alone.
Each phase makes the next one harder to prevent. By Phase 3, the leverage has shifted permanently to the application layer.

Three advantages nobody else has

Anthropic has three asymmetric advantages. The strategy should be built on these — not on imitating what Sierra or Decagon have already built.

Anthropic's asymmetric advantages
1
The Trust Brand
Anthropic's safety reputation uniquely positions Claude for HIPAA, SOX, PCI-DSS — regulated industries that represent ~40% of the CS market
PwC, Accenture, and Epic are already partnering with Anthropic for compliance
2
Bundled Distribution
300,000+ businesses already use Claude. Adding CS is usage expansion, not a new sale — near-zero acquisition cost
Same wedge that Slack, Zoom, and Figma used to displace incumbents
3
Intelligence Compounds Globally
One model upgrade benefits every customer at once. Sierra's bespoke integrations require custom work per account
Margin trajectory: 50% in 2026 → 77% by 2028 — integration-first can't match this
The trust brand opens the door. Bundled distribution eliminates the sales cycle. Intelligence compounding wins the long game.

The whitespace is specific: healthcare, financial services, insurance, and government customer support is ~40% of the total market — and the most underserved by AI-native startups like Sierra and Decagon, whose customer bases skew toward internet-native companies (Notion, Duolingo, Eventbrite).

The playbook: three horizons

The strategy has to navigate a partner tension. Anthropic can't alienate Sierra and Decagon (who drive significant API revenue) while competing with them. The answer is market segmentation:

Sequenced execution plan
Horizon 1
Next 90 days
CS Plugin
Ship ticket triage, response generation, and brand voice customization via the existing Cowork plugin architecture. Build MCP connectors for Zendesk, Salesforce, Intercom.
Horizon 2
6–18 months
Regulated Wedge
HIPAA/SOX/PCI compliance templates. Voice-native CS. "Powered by Claude" partner certification. Proactive support: AI initiates before issues escalate.
Horizon 3
18–36 months
Intelligence Layer
Unify CS data across product, sales, and marketing. Per-resolution pricing (300x API revenue). Agent-to-agent protocol via MCP.
Horizon 1 requires no new product build and targets mid-market self-serve — deliberately avoiding Sierra and Decagon's enterprise segment.

The beauty of H1 is the math. Cowork + plugin architecture already exists. The target buyer already has a Claude Enterprise account. The procurement cycle is weeks, not months. There's no partner conflict because mid-market self-serve isn't Sierra's or Decagon's segment.

The decision

Every infrastructure company faces a moment where staying in the infrastructure layer stops being a choice and starts being a constraint.

For Intel, it was mobile. For IBM, it was cloud. For Twilio, it was AI. The companies that recognized the shift and moved up the stack survived. The ones that didn't became commodity inputs in someone else's value chain.

Anthropic is at that moment now. The numbers are extraordinary — $14 billion ARR, 10x annual growth, $380 billion valuation. But 80% of revenue comes from API consumption, which means 80% depends on other companies' decisions about which model to use.

Customer support is where the vulnerability is most acute — and where the escape is most clear. The question isn't whether to act. It's whether to act now, while the window is open.

Kevin Nguyen is a product manager exploring consumer, fintech, and AI. He writes about product strategy, design process, and the things he's curious about.

Connect on LinkedIn →