Enterprise AI is entering a more skeptical phase. After months of frontier-model launches and agent platform announcements, the sharper story today is that large customers are starting to ask harder questions about the economics: high token bills, unclear ROI, and the risk of handing valuable business logic to outside AI labs.
Palantir CEO Alex Karp has become the loudest voice in that backlash. His complaint is not simply that OpenAI, Anthropic, and other frontier labs are expensive. It is that customers may be paying for model access while giving away the context, workflow data, and operational patterns that make their own businesses valuable.
That criticism lands at a delicate moment. OpenAI, Anthropic, Meta, xAI, and other AI companies are moving faster than most buyers can evaluate, with new models, coding systems, enterprise agents, voice interfaces, API tiers, and paid developer platforms arriving in the same week. The industry is selling velocity. Buyers are starting to ask for proof.
The buyer mood has changed
The first phase of enterprise AI adoption was defensive. Nobody wanted to miss the platform shift. CEOs approved pilots, departments bought seats, engineers added copilots, and vendors promised that agentic systems would turn knowledge work into repeatable automation.
The second phase is less forgiving. Finance teams are watching token spend. Security teams are watching data movement. Product leaders are asking whether the AI feature improves retention, margin, response time, or customer satisfaction. Operations leaders want systems that survive policy changes, model outages, pricing shifts, and regulatory reviews.
Karp's argument is resonating because it gives language to a frustration many companies already feel. A model can be impressive and still fail as an enterprise product if it cannot prove value inside the buyer's real workflow. A demo can look magical while the production system quietly becomes expensive, hard to audit, and dependent on a provider's roadmap.
Token pricing is becoming a boardroom issue
Token pricing used to sound like developer plumbing. Now it is a budget question. Once AI moves from a chat window into code review, customer support, document analysis, research, voice calls, and background agents, usage can multiply quickly. Agentic systems do not just answer once. They plan, browse, call tools, retry, summarize, inspect, revise, and produce long outputs.
That can be worth paying for when the workflow creates measurable value. The problem is that many buyers are still paying for activity before they can prove outcomes. The vendor charges for input and output. The enterprise has to prove whether those tokens reduced labor, improved quality, increased sales, shortened cycles, or opened a new product line.
This is why the debate is shifting from "which model is best?" to "which system creates controlled economic leverage?" Cheaper model tiers help, but they do not solve the full problem. A company also needs routing rules, spend limits, human approvals, task queues, cached context, evaluation loops, and a way to stop agents from turning every problem into a long expensive reasoning session.
Control is the deeper issue
The cost argument is only the surface. The deeper enterprise concern is control. If an outside AI lab sits between a company and its workflows, the lab may learn which tasks matter, which documents drive decisions, which prompts create value, which data structures are useful, and which internal processes are ready to automate.
Even when contracts protect customer data, the strategic discomfort remains. Businesses want AI capability, but they do not want to outsource their "alpha" to a provider that may later sell adjacent tools into the same market. They want the intelligence layer to strengthen their internal operating system, not replace it with someone else's platform.
That is why AI sovereignty language is becoming commercially useful. For enterprises, sovereignty does not only mean running every model on-premise. It means keeping ownership of workflow state, audit trails, user permissions, domain logic, evaluation data, and the final customer relationship. The model can be external. The business system should not become hollow.
The labs are also becoming platforms
The pushback is happening because the labs are not just model suppliers anymore. They are building work platforms. OpenAI is pushing deeper into Codex, ChatGPT Work, Programmatic Tool Calling, multi-agent workflows, and tiered model routing. Anthropic is expanding Claude into coding, Slack, research, and specialized work surfaces. Meta is turning model access into a paid developer API. xAI is moving Grok from consumer attention into infrastructure and developer distribution.
That creates a strategic tension. Enterprises want the best intelligence available, but the most capable providers increasingly want to own the workflow layer around that intelligence. The more complete the lab's platform becomes, the more convenient it is for buyers. It also becomes harder to separate the enterprise's own product, data, and process advantages from the provider's stack.
The next AI platform winners may not be the labs with the biggest model alone. They may be the companies that help businesses keep control while still using strong models. That includes model routers, private deployment layers, evaluation tools, workflow builders, observability systems, secure data connectors, and agent frameworks that make provider switching realistic.
What enterprises should demand
Enterprise buyers should stop treating frontier AI as a simple subscription line item. The stronger approach is to define the operational contract before scaling usage.
That starts with economics. What is the expected cost per completed task? Which tasks are allowed to use the premium model? What is the fallback when the model is too expensive, too slow, or unavailable? Which outputs need human approval before they touch customers, code, legal documents, or financial decisions?
It also requires data boundaries. Which documents can leave the company environment? Which fields must be redacted? Which prompts, outputs, tool calls, and intermediate artifacts are logged? Who can inspect the agent's reasoning path, and how long should records be retained?
The final question is ownership. If the company learns that an AI-assisted workflow improves margins, does that learning stay inside the company? Can the workflow be exported, rerouted, or reproduced with another provider? Or has the business accidentally turned its most valuable process into a tenant inside a lab-controlled platform?
The product lesson
The backlash against the labs does not mean enterprise AI is failing. It means the market is growing up. Buyers are moving from experimentation to procurement discipline. They are learning that model intelligence is only one ingredient. The full product has to include governance, cost control, security, reliability, user trust, and measurable operational value.
That is the same shift SunMarc has been tracking across agent-native tools. The winning product is not the one that simply mentions the newest model. It is the one that turns model capability into a workflow users can understand, control, and trust.
For SunMarc App Labs, the practical takeaway is direct. AI-enabled products should keep the workflow visible and owned by the product. QR Remix should make transformations inspectable before a code is generated. PDF utilities should preserve local processing, previews, and reversible steps. Navigation and utility tools should make data use obvious. Future AI-enabled web properties should expose sources, actions, permissions, and costs in plain language.
Enterprise buyers are not rejecting AI. They are rejecting vague AI. They want systems that prove value, protect data, respect boundaries, and keep business advantage in the hands of the business. That is a healthier market, and it will reward builders who design for trust instead of hype.
Relevant links
- WSJ: Alex Karp Is Saying What Every Angry CEO Is Thinking About AI
- Axios: The revolt against U.S. AI labs
- Business Insider: Alex Karp rips into AI labs
- Business Insider: New AI model announcements in the latest rush
- SunMarc archive: GPT-5.6 Turns OpenAI's Model Launch Into a Work Platform
- SunMarc archive: Meta's AI Push Is Turning Into a Compute Business Story