Customer-Controlled AI: The Future of Enterprise AI

Customer-controlled AI is at the forefront of the ongoing battle in enterprise AI, shifting from pricing per token to the ownership of data, models, and competitive advantages. Regulators are beginning to take sides, and the outcomes will hinge on infrastructure that clients can manage. India is uniquely positioned to lead this transformation.

The Shift from Boardrooms to Live TV

In the evolving landscape of enterprise AI, the emphasis is now on control rather than mere capability. Alex Karp made this argument dramatically during a live interview on CNBC last week. The CEO of Palantir noted that business leaders he interacts with privately feel frustrated with top model developers. They argue that while companies pay hefty fees for AI tools, the providers retain control of proprietary data that can enhance their own models, resembling a wealth tax for businesses.

According to Karp, enterprises are effectively paying for tokens that offer little tangible value while surrendering crucial data and process expertise that forms their competitive advantage. Following the interview, Palantir’s shares surged by approximately 9%, indicating that the market recognised the underlying concerns.

Karp’s perspective is not entirely impartial; the interview coincided with a partnership announcement, and he acknowledged that he benefits from the same practices he critiques.

When encountering a self-serving argument, the appropriate course of action is to scrutinise it rather than dismiss it. His assertions echo a growing consensus, supported by independent evidence outlined below.

In a prior discussion for StartupSuperb, it was argued that the future of AI billing should shift from effort-based to outcome-based metrics. Companies will not continue to fund systems that charge for activity rather than actual results.

This forms the first part of the challenge. The second aspect is more complex. Even a reasonably-priced token becomes a poor choice if it also entails the transfer of data, workflows, and proprietary advantages to another party.

Emerging Questions in Enterprise AI Contracts

Every significant conversation regarding AI procurement is coalescing around similar critical questions:

• Who holds ownership of the data?
• Where is it stored?
• Who has control over the models?
• How can clients be assured that their data isn’t being used to train external models?

Karp characterised these as questions that enterprises are now asking, stressing the need for control over computing resources, models, data stacks, and competitive insights. Within a year, these queries are likely to appear in standard procurement documents, just like security certifications and audit trails do today.

The underlying conflict is becoming increasingly apparent. A client’s proprietary data and process knowledge provide valuable raw material for enhancing foundational models. The vendor’s motivations often conflict with the interests of the customer, and no contractual agreement can entirely alleviate that tension if the system in operation remains a black box.

The result is straightforward. Any product requiring the client to relinquish visibility regarding the treatment of their data will face limitations on the market it can penetrate. The technology may function seamlessly; however, trust is essential for long-term contracts, particularly large, multi-year deals, which constitute the bulk of revenue in enterprise AI.

Regulatory Developments in Key Sectors

India’s regulatory framework has already made significant strides, being ahead of many other regions. The RBI’s FREE-AI framework clearly favours indigenous and sector-specific AI models rather than generic third-party solutions, highlighting data sovereignty and vendor concentration as critical risks. It foresees coordinated standards among RBI, SEBI, and IRDAI in the future.

Boards and senior management are ultimately responsible for validating third-party models, which are expected to be scrutinised with the same rigor as internal systems. A bank’s board cannot authenticate what it cannot review, making the regulatory framework incompatible with systems that operate as black boxes.

This trend is global. Certain categories of data are legally restricted from leaving the infrastructure controlled by the enterprise, such as sensitive health information and specific financial data. Zero-retention enterprise models can reduce risk, but do not eliminate it, and the EU AI Act will be fully in effect on 2 August 2026, with European data residency laws already prompting a shift toward self-hosted solutions.

Even in sectors without such regulations currently, like banking and energy, the likelihood is that they will emerge, as thoughtful regulators will adopt similar systemic-risk analyses. Each new guideline further narrows the obtainable market for unclear AI technologies in sectors that allocate significant resources to technology.

The Significance of Open Alternatives

The next phase of the market is taking shape, and it is developing more swiftly than many realise. In fields such as classification, extraction, summarisation, and coding, a robust open-weight model in 2026 will capably suffice, even as closed models maintain their edge in complex reasoning. However, most regulated enterprise tasks do not depend solely on these advanced capabilities.

Open-weight systems are now achieving coding proficiency comparable to top proprietary models while costing just a fraction per token, with one open model family surpassing one billion downloads by January 2026.

The market has clearly indicated its preference. Open-weight providers now account for over 45% of traffic on OpenRouter, the largest model aggregator, a substantial increase from less than 2% a year prior. The financial benefits are impressive, with Prosus reporting inference cost reductions of up to 26 times when using open models instead of proprietary systems.

What’s more significant is what open weights provide: three essential rights:

• The right to operate the model on infrastructure controlled by the client.
• The ability to fine-tune it using proprietary data.
• Freedom from a vendor’s pricing structures and lifecycle choices.

These rights directly address the procurement questions, explaining why this transition is likely to be long-lasting rather than temporary.

India’s Role in Model Deployment

India has been both deploying open-source models and developing its own, with a flexible design approach that is precisely what regulated buyers require. This flexibility allows for inspection, fine-tuning, and collaborative deployment within their own boundaries.

The broader opportunity lies at the application layer, which Karp identifies. This represents a service-oriented chance that resonates with India’s historical triumphs.

India has established a technology sector generating $315 billion (₹30 lakh crore) in annual revenue for FY26, according to NASSCOM’s Annual Strategic Review 2026, by implementing leading enterprise software according to client specifications. The AI sector can yield even more significant opportunities per engagement, as implementations now encompass model selection, fine-tuning with private data, and compliance evidence required by frameworks like FREE-AI.

The existing delivery framework consists of 2,117 global capability centers producing $98.4 billion (₹9.4 lakh crore) and employing 2.4 million professionals.

Indian enterprises demonstrate this potential effectively. H2LooP, based in Bengaluru, is crafting hardware-aware vertical AI models tailored for critical embedded systems across sectors like automotive, aerospace, and semiconductors, with a clear focus on intellectual property sovereignty and on-edge reliability.

Similarly, Smallest.ai, also out of Bengaluru, is handling foundational voice models that manage over one million inference calls for international enterprises, showcasing the capacity of Indian model firm to compete for significant enterprise workloads.

Recommendations for Founders

The roadmap for founders should be methodical:

• Establish model-agnostic deployment frameworks tailored for regulated Indian markets, ensuring that clients maintain control over model weights and audit trails.
• Approach compliance requirements under new regulations like FREE-AI as a product rather than a cost centre.
• Price services based on outcomes or directed inference delivered on client-controlled infrastructure, effectively aligning the two previous points.

Control is the next integral part of the economy focused on results. Earlier discussions underscored the need for enterprises to pay for outcomes rather than efforts. This current discussion adds the critical stipulation that outcomes must not require clients to relinquish their competitive advantages.

The emerging leading AI businesses of the next decade will deliver measurable results on models clients can review, all operating on infrastructure they manage. India possesses the necessary open models, engineering talent, and advancing regulatory clarity to cultivate this business first for its crucial industries. India for the World begins domestically, from within the controlled environment.