You are paying for the R&D of the next competitor

31.07.2026

4 min read

You are funding the R&D of your next competitor and calling it AI transformation. Frontier Labs sell models into your most sensitive workflows while building the same verticals themselves. That is governance with deal-level consequences.

Key Takeaways

Sell, Embed, Map, Compete. The labs embed themselves into core processes and spin the same capabilities out as their own products.
Figma is a boardroom topic. After Anthropic’s CPO exited the Figma board, Claude Design arrived. Field’s verdict: “not consistently candid.”
Verticals chase ROCE. Legal, healthcare, science and support are precisely the markets with high margins and dense data.
Contracts and data classes govern. Multi-vendor alone isn’t enough. Without exit clauses and workflow sovereignty, the partner remains a silent M&A runway.

Related: Washington decides which AI models are allowed to run here · False security: when the cyber policy fails in a real crisis

The partner that steals your company

What is Frontier Lab cannibalisation? It describes how model providers like OpenAI and Anthropic first sell APIs and partnerships to enterprises, embed themselves in core processes, and later launch the same workflows as their own vertical products. The customer is effectively funding the product discovery of their future competitor.

Ole Lehmann’s 30 July 2026 thread puts it bluntly: OpenAI and Anthropic are cannibalising their biggest customers. The driver is profitability-doing nothing is not an option. Chamath Palihapitiya added that once models become commodities, vertical software is the natural resting place for margin.

For executive teams this isn’t a tech column-it’s a question of data sovereignty, IP, board conflicts and deal structure. Plugging Frontier APIs into contracts, clinical documentation, biotech R&D or support funds the lab’s product discovery and calls it transformation.

Figma: Trust is not a contract

The most public case is design. Figma and Anthropic collaborated; Anthropic’s CPO sat on Figma’s board. Just before Claude Design launched, the CPO stepped down. The product then appeared as an Anthropic offering. Dylan Field later said Anthropic had been “not consistently candid” in communications. The quote is documented via Upstarts Media and follow-up reports.

The lesson for CDOs and CPOs: a board seat and a press release do not replace product-roadmap transparency. When the lab targets the same job-to-be-done, the conflict was baked in from the start.

Why legal, health and science fall first

Lehmann’s list runs from Harvey and Claude for Legal through support agents to clinical documentation and biotech workbenches. Not every point is equally well evidenced, yet the pattern is stable: high ROCE, dense proprietary workflows, expensive specialist software with thin UI over third-party models. That’s exactly where Frontier Labs can capture the margin themselves.

Many “AI-first” SaaS firms are orchestration plus prompt plus compliance wrapper. When the lab builds the wrapper itself, the moat was never the model-the pitch was incomplete. That’s uncomfortable, but often true.

Risk Matrix for the Digital Chief

Rather than blanket lab restrictions, what’s needed is a control logic. The matrix is rough but decision-ready:

Data / Workflow	Frontier API OK	Only with Exit Path	Banned in Lab Tenant
Internal knowledge search, drafts	yes	–	–
Customer support L1 with masking	–	yes	–
Contract analysis, due diligence	–	yes	Raw files without policy
Clinical documentation, patient data	–	strictly regulated	unchecked cloud defaults
Core IP, pricing, roadmap raw data	–	–	yes

Five Clauses Before the Next Frontier Deal Is Signed

Product Launch Notice. The lab must inform you in good time when a competing vertical product is in the pipeline that addresses the same job.
Data Use and Training. Clear limits on logs, traces, fine-tuning, and human review. Default is rejection, not hope.
Workflow Portability. Export of prompts, tool schemas, and evaluation sets. Without these, multi-vendor setups are theater.
Board and Side-Letter Conflicts. Dual-role personnel must disclose conflicts and step down before products go live.
Exit in 90 Days. Technical and contractual switchability to a second model or your own stack. Otherwise, the deal is a bet on eternal friendship.

The Uncomfortable Second Read

Frontier labs act rationally. Whoever sees the world’s most sensitive processes builds the software that replaces them. Many customers have invited exactly that: cheap intelligence, quick demos, and no agent layer of their own. The provocation is clear: the lab is acting rationally. Often, the pitch was incomplete.

Digital chiefs who steer this now gain time. They segment workflows, demand clauses, and build portability. The others keep funding their competitor’s R&D-and wonder why the roadmap suddenly appears in the lab’s product catalog.

Frequently Asked Questions

Is multi-vendor protection sufficient?

It’s a helpful starting point, but inadequate as a standalone strategy. Without your own orchestration and data boundaries, you’re merely shifting the lab while retaining the dependency.

Should we avoid Frontier Labs?

No. Segmentation is the key. Commodity assistance can reside in the lab. Differentiating core processes require an exit path and strict data policies.

What makes the Figma case stand out?

Board proximity, partnership, and a competing product launched in quick succession. The quote “not consistently candid” transforms a product launch into a trust and governance issue.

How should I handle unclear cases from social threads?

Treat them as signals. Only the primary source counts as fact. Any line that enters contracts or board documents requires a primary source.

What’s the first concrete step in 30 days?

Compile a list of the top 10 workflows with Frontier dependencies, data classification, and exit capability. Then draft clauses and second-model pathways for the three most critical ones.