Enterprise AI Is Entering Its ROI Era: Why Alex Karp Thinks Companies Are Losing Patience With AI Vendors

Enterprise AI is entering a more difficult era.

For the past two years, businesses have been encouraged to experiment with large language models, copilots, AI assistants and autonomous agents. Companies have spent heavily on subscriptions, cloud infrastructure, consulting and access to increasingly powerful models—often with the expectation that meaningful returns would follow quickly.

But the mood is beginning to change.

Palantir CEO Alex Karp has become one of the most vocal critics of the way enterprise AI is currently being sold. In interviews and earnings calls, Karp has argued that many enterprises are becoming frustrated with leading AI vendors. They are paying for tokens, pilots and impressive demonstrations, yet often struggling to convert AI capabilities into measurable business value.

Karp has also warned about a second problem: strategic dependency. If companies rely entirely on closed frontier models, they may increasingly hand over access to their workflows, proprietary data and institutional knowledge—the corporate intelligence Karp often refers to as “alpha”—to technology vendors they do not control.

This debate comes at an important moment for enterprise technology.

Businesses are no longer asking only which AI model is the smartest.

They are increasingly asking a more difficult question:

Which AI provider can actually create measurable value inside our business?

The AI Vendor Problem Is Becoming an ROI Problem

Enterprise AI is moving beyond the experimentation phase.

The first stage of generative AI adoption was largely about testing. Businesses deployed chatbots. Employees experimented with writing assistants. Developers tried coding tools. Marketing teams generated campaigns. Customer-service teams experimented with automated responses.

These experiments created enormous excitement because the technology was capable of producing impressive results.

But an impressive demonstration is not the same as business transformation.

A company may have thousands of employees using an AI assistant and still struggle to demonstrate meaningful improvements in revenue, operating costs, customer satisfaction or cycle times.

That is where Karp’s argument becomes important.

His central point is that enterprises need AI systems that work inside their businesses—not isolated models that simply answer questions.

According to Palantir’s approach, AI must be integrated into business processes rather than treated primarily as a question-and-answer tool.

The distinction matters.

An AI model can write a sophisticated email in seconds. But a large organisation often needs considerably more than that.

Enterprise AI may require:

• Access to organisational databases

• Awareness of internal rules, policies and procedures

• Controls over who can interact with the system

• Access to customer information

• Visibility into inventory and products

• Integration with accounting and operational systems

• Human approval for important decisions

• Records and audit trails of AI activity

• Automation of multi-step workflows

• Governance and control across the entire process

The challenge, therefore, is not simply whether an AI model can produce an intelligent answer.

The challenge is whether it can reliably operate inside a real organisation.

The Gap Between the Demo and Deployment

One of the biggest problems in enterprise AI is the gap between a successful demonstration and reliable production performance.

An AI system may answer a question perfectly during a demonstration.

But enterprise deployment requires much more.

The system may need to achieve acceptable levels of accuracy, reproducibility, traceability and confidence across thousands or millions of interactions. It must also perform reliably when connected to imperfect data, changing workflows and human decision-making processes.

This is where many AI projects become more difficult.

A vendor may demonstrate that its system can perform a task.

The enterprise must determine whether that system can perform the task reliably enough to become part of a critical business process.

That difference between can do and can be trusted to do repeatedly may be one of the most important distinctions in enterprise AI.

Why Karp Says Enterprises Are Becoming Frustrated

Token Economics Does Not Equal Business Profitability

One of Karp’s most controversial arguments concerns the economics of frontier AI.

Many AI vendors charge according to token usage, API calls, subscriptions or computing capacity.

Those pricing models are understandable from a technology-provider perspective.

But enterprises are not fundamentally interested in tokens.

They are interested in outcomes.

A manufacturer wants fewer production delays.

A bank wants to detect fraud faster.

An insurance company wants to process claims more efficiently.

A retailer wants to improve conversion while reducing customer-service costs.

A hospital wants to reduce administrative burdens while improving safety.

An AI system can consume millions of tokens and still fail to generate meaningful economic value.

This is why enterprise AI buyers are beginning to ask a more difficult question:

What did we actually achieve from this AI investment?

That question is more complicated than it appears.

Imagine a company spends $2 million deploying an AI system and employees become 15% more productive. That sounds like a success.

But if those employees were not operating at full capacity before the AI deployment, the productivity improvement may not translate into equivalent financial value.

Similarly, an AI customer-service system might answer 80% of incoming enquiries. But if every response still requires human verification, the organisation may achieve little or no reduction in operating costs.

The value of AI cannot therefore be measured simply by whether the technology works.

It must be measured by what changes in the business.

Karp’s Argument for AI Sovereignty

The second major issue raised by Karp concerns control.

Who ultimately controls an organisation’s intelligence?

Companies are increasingly providing AI systems with access to highly valuable internal information:

• Customer data

• Product information

• Engineering knowledge

• Financial information

• Operational procedures

• Research documents

• Internal workflows

• Proprietary business practices

Karp argues that companies should be cautious about allowing third-party AI providers to become deeply embedded in their information infrastructure.

This argument is often described in terms of AI sovereignty.

The idea is that companies should retain meaningful control over their data, AI infrastructure and strategic knowledge.

The issue is not simply cybersecurity.

It is also strategic dependency.

Consider a pharmaceutical company that provides an AI platform with access to decades of research documents, production knowledge and market intelligence.

Even if contractual protections prevent the vendor from directly using that information, the company may still become dependent on a third-party technology layer for accessing, interpreting and operationalising some of its most valuable knowledge.

That creates a strategic question:

What happens if the company wants to change providers?

The Risk of Becoming Locked Into an AI Vendor

Vendor lock-in is not new.

Businesses have experienced it with databases, cloud platforms and enterprise software for decades.

AI may create an even more complicated version of the problem because a single provider could potentially control several layers of the technology stack:

• The foundation model

• Cloud infrastructure

• The AI development environment

• The application layer

• Data processing technologies

• Workflow orchestration

That concentration can give vendors significant strategic power.

Karp’s preferred alternative is a more modular environment in which companies can use different models while retaining control over their broader enterprise infrastructure.

This is one reason Palantir emphasises a model-agnostic approach: customers should be able to use different AI models without rebuilding their entire enterprise AI architecture.

The fastest or smartest model today may not necessarily be the best model tomorrow.

Enterprises may therefore benefit from preserving the ability to change.

Palantir’s Answer: Build Around the Model

Karp’s broader argument is that the model itself may eventually become increasingly interchangeable.

The AI industry has spent years competing around:

• Larger models

• Larger context windows

• Better reasoning

• Faster performance

• Lower latency

• Higher benchmark scores

These improvements matter.

But Karp’s argument is that they may not represent the most durable source of value for enterprise customers.

The more durable technology may be everything surrounding the model.

That includes:

• Data integration

• Security

• Workflow management

• Governance

• Application development

• Human oversight

• Operational deployment

This is the philosophy behind Palantir’s AI Platform.

Rather than simply providing employees with a chatbot, Palantir attempts to integrate AI capabilities into operational workflows.

The focus is not just on whether the model can generate intelligence.

It is on whether that intelligence can move through a controlled business process and create a measurable outcome.

Palantir’s Bootcamp Model

One of Palantir’s more distinctive approaches is its AI bootcamp programme.

Rather than spending months discussing potential AI opportunities, Palantir brings customers into intensive working sessions designed to build real applications around specific operational problems.

The philosophy is straightforward:

Show the business result, not just the benchmark.

Palantir has highlighted cases in which customers developed production-oriented AI applications within short periods of time. One example cited by Karp involved an AI-based disruption-management system that Palantir said could generate approximately $10 million in savings.

Whether every organisation can achieve results of that scale is another question.

But the underlying philosophy is important.

Enterprise AI vendors may increasingly need to demonstrate value quickly and in concrete business terms.

The conversation is shifting from:

“Look what our model can do.”

to:

“Show me what this changes in my business.”

Enterprise AI Is Entering a Reality-Check Phase

The broader market appears to be moving in the same direction. AI adoption is increasing rapidly. But successful transformation is proving more difficult.

This creates an important paradox.

More companies are using AI and more employees are experimenting with AI.

More enterprise software products now include AI capabilities.

Yet widespread adoption does not automatically translate into widespread business transformation.

The next stage of competition may therefore not be about persuading organisations to adopt AI. It may be about making AI actually work.

Research on enterprise AI adoption also shows that usage remains uneven. Organisations use AI differently depending on their data environments, regulations, workflows and risk tolerance.

There is no single enterprise AI playbook.

A successful deployment in one company may fail completely in another.

Real-World Example: Manufacturing

Consider a manufacturer dealing with production disruptions.

A basic AI assistant could analyse incident reports and inform the operations team about potential problems.

That is useful.

But an integrated AI system could potentially do much more.

It could:

• Interpret production alerts

• Analyse current inventory levels

• Check supplier status

• Assess available machine capacity

• Review delivery commitments

• Propose alternative production schedules

• Evaluate the operational impact

• Request human approval

• Trigger approved downstream actions

The difference is significant.

The first approach is AI-assisted analysis.

The second is AI integrated into a business process.

This is central to Palantir’s argument.

Enterprise AI requires more than connecting corporate data to an intelligent model.

It requires building the systems around the model that allow intelligence to become controlled action.

Real-World Example: Financial Services

The same principle applies in financial services.

Consider an AI assistant used by an insurance company. In its simplest form, it might summarise claims documents. In a more advanced implementation, the system could also need access to:

• Policy databases

• Customer records

• Claims-processing systems

• Fraud-monitoring systems

• Regulatory requirements

• Pricing systems

• Human approval processes

At that point, the AI system is no longer simply summarising information. It is becoming part of an operational decision-making environment. 

That creates greater opportunities for value. However, it also creates greater risks.

If an error occurs, the organisation may need to understand:

• What information influenced the decision?

• Why did the AI produce that recommendation?

• Who approved the action?

• Was the decision compliant with regulations?

• Can the organisation reconstruct what happened?

This is why enterprise AI cannot be treated purely as a model-quality problem.

The intelligence of the model is only one part of the system.

Why AI Vendors May Need to Change Their Sales Pitch

For years, AI competition has largely revolved around a relatively simple message:

Our model is smarter.

That message made sense when the primary technological challenge was demonstrating what AI could do.

But enterprise buyers are becoming more sophisticated.

The question is increasingly moving from:

“Does Model A outperform Model B?”

to:

“Which system will generate the best return for my specific business?”

The answer may depend on:

• Cost

• Security

• Reliability

• Integration

• Governance

• Deployment speed

• Human oversight

• Vendor flexibility

Benchmark performance still matters. However, benchmark performance alone is no longer enough.

What If AI Pricing Became Outcome-Based?

This shift may eventually change how AI vendors price their products.

Instead of relying entirely on tokens, API calls or software licences, some AI services could increasingly be priced according to business outcomes.

For example:

• Cost per successfully resolved insurance claim

• Cost per automated customer-service interaction

• Cost per completed software task

• Revenue generated through an AI-enabled sales process

• Savings created through procurement automation

Such models would transfer more financial risk from the buyer to the technology provider.

But they would also force vendors to demonstrate the effectiveness of their systems.

And that may be exactly what enterprise customers increasingly want.

The Irony: Palantir Is Also an AI Vendor

There is, of course, an important contradiction in Karp’s argument.

Palantir is itself an AI infrastructure and software vendor.

Its criticism of frontier-model companies is therefore also a form of competitive positioning.

Palantir is competing for the same enterprise AI budgets.

Its argument is effectively that enterprises should not build their AI strategies primarily around model providers—but around the infrastructure and operational layer that Palantir itself provides.

That does not necessarily make the argument wrong.

A commercially useful argument can still describe a genuine market problem.

But corporate leaders should recognise the competitive context.

Karp is not a neutral observer of the enterprise AI market.

He is promoting a particular vision of how that market should develop.

The Bigger Industry Trend: From Experimental AI to Accountable AI

The first era of enterprise AI was experimentation.

The next era may be accountability.

Corporate leaders will increasingly need answers to questions such as:

• How much money has this AI saved us?

• How reliably does it work?

• How much human oversight is required?

• What happens if we change models?

• Where is our data stored?

• Can we switch vendors without rebuilding everything?

• Who is responsible when an error occurs?

• Can we demonstrate compliance?

These are considerably more difficult questions than:

“Is the model good at writing paragraphs?”

But they are the questions that will increasingly determine whether enterprise AI investments survive.

What Corporate Leaders Should Take From Karp’s Argument

The lesson is not that companies should stop working with OpenAI, Anthropic, Google or other AI providers.

The lesson is that the AI model itself cannot be the entire strategy.

Enterprise leaders should evaluate the complete system.

1. Start With the Business Problem

Do not begin with:

“Where can we use AI?”

Start with:

“Which business process is expensive, inefficient or slow?”

Then determine whether AI can meaningfully improve it.

2. Measure Economic Outcomes

Measure results that matter to the organisation.

Productivity is useful, but the most important measures may include:

• Revenue

• Cost savings

• Cycle time

• Quality

• Customer satisfaction

• Risk reduction

3. Protect Critical Data and Knowledge

Leaders should understand exactly what information enters AI systems, how it is processed and what contractual and technical protections exist.

4. Avoid Unnecessary Dependence

Where practical, organisations should preserve the ability to use different models and providers.

The best AI provider today may not be the best one tomorrow.

5. Keep Humans Where They Matter

High-impact decisions should not be automated simply because AI is technically capable of making them.

The appropriate level of human oversight should depend on the consequences of failure.

The New Enterprise AI Vendor Test

Karp’s argument ultimately leads to a new test for enterprise AI.

The winner may not necessarily be the company with the largest or most impressive model.

It may instead be the company that can answer five straightforward questions:

• Does it work?

• Can it be integrated into our organisation?

• Can we measure the ROI?

• Do we retain control over our data and knowledge?

• Can we change vendors or models if necessary?

That is a much more demanding test than outperforming a benchmark.

And it may explain why enterprise customers are becoming less patient with AI vendors.

For years, companies have heard about the revolutionary potential of artificial intelligence. They have seen impressive demonstrations, attended AI conferences and launched pilot programmes.

Now they are asking for results.

Karp has turned that frustration into one of the most powerful narratives in enterprise technology.

Frontier AI companies may be exceptionally good at building intelligent models. However, enterprises ultimately need systems that can convert intelligence into measurable business value without forcing organisations to surrender control over their data, processes and strategic knowledge.

Enterprise AI is moving from a race for intelligence to a race for accountability. The winners in the next phase of enterprise AI will not necessarily be the companies offering the most impressive chatbot. They will be the providers that can connect AI to business reality, demonstrate economic value, preserve corporate control and give customers enough flexibility to remain in charge.

Frequently asked questions

What is causing frustration among enterprises regarding AI vendors?

Enterprises are frustrated because they often pay for AI capabilities but struggle to convert these into measurable business value, leading to impatience with their vendors.

Why is integration of AI into business processes important?

Integration is crucial because it allows AI to provide actionable insights and improve operational efficiencies rather than only serve as a standalone tool that answers questions.

What does Alex Karp mean by AI sovereignty?

AI sovereignty refers to the control companies should retain over their data and AI infrastructure, ensuring they do not become overly dependent on third-party technology providers.

How should companies measure the value of their AI investments?

Companies should focus on economic outcomes such as revenue, cost savings, and customer satisfaction rather than simply productivity metrics to evaluate the true impact of AI.

What shift in AI vendor strategy is suggested by the article?

The article suggests that AI vendors may need to move towards pricing models based on business outcomes instead of traditional metrics like tokens or API calls to align better with enterprise needs.