Artificial intelligence is entering a new phase of enterprise adoption, where success is increasingly measured not by the number of models deployed or pilots launched, but by tangible business outcomes and economic value. As organisations move beyond experimentation, executive priorities are shifting towards AI governance, cost optimisation, token efficiency and return on investment. In this evolving landscape, concepts such as AI unit economics, token governance and context-aware AI are emerging as critical enablers of scalable and sustainable AI adoption.

Against this backdrop, Celonis has introduced its Context Model, designed to provide AI systems with operational context that improves decision-making, reduces token consumption and minimises hallucinations while helping enterprises optimise AI costs. The company's vision reflects a broader industry shift towards grounding AI in business processes rather than relying solely on increasingly powerful foundation models.

In an interaction with AI Spectrum, Kaushik Mitra, Vice President and Head of India Go-to-Market, Celonis India, discusses why the future of enterprise AI will be defined by measurable business value rather than model volume, the growing importance of token governance and AI economics, the role of contextual intelligence in improving AI performance, and how organisations can build sustainable competitive advantage through process-aware, context-driven AI.

After several years of rapid AI experimentation, many organisations are now facing increasing scrutiny over AI spending. How do you see the conversation evolving from AI innovation and adoption toward AI economics, accountability, and measurable business outcomes?

The enterprise conversation is shifting from unconstrained experimentation toward strict capital discipline and token austerity. CFOs now scrutinize AI spending to ensure deployments generate clear economic value against rising compute costs. Grounding AI in operational logic establishes a predictable cost-per-outcome, transforming it from an unpredictable investment into a defensible line item. Ultimately, future AI success will be measured by the business value generated per token rather than the total volume of models deployed.

Celonis has introduced the Context Model to help enterprises optimize AI deployments. What role does operational context play in improving AI performance, reducing token consumption, and minimizing costly errors such as hallucinations and unnecessary retries?

Frontier models are generalists that guess when they lack operational context, causing hallucinations and errors. The Celonis Context Model (CCM) acts as an intelligent layer and translator, providing AI systems with structured access to process data and business logic. This optimization shortens prompts and filters out redundant data. Consequently, early customer pilots cut average output and cache read tokens by more than 50% while doubling task completion success.

You have suggested that future AI success will be measured not by the volume of AI deployed but by the business value generated per token. How should enterprises begin thinking about AI unit economics, and what metrics should leaders track to evaluate return on AI investments?

Enterprises should adopt the strategy to "experiment broadly, optimize purposefully" to manage AI spend effectively. To evaluate return on investment, leaders must track a new core metric: "Revenue per Token". This requires agentic cost tracking to tie token usage directly to concrete business outcomes. Leaders should monitor reductions in token consumption alongside completion success to justify corporate AI expenditures.

The concept of "token governance" is gaining traction as organisations seek greater visibility into AI-related costs. How do you expect governance frameworks around token usage, model selection, and agent performance to evolve over the next few years?

Governance frameworks will evolve from managing unlimited pilots toward strict budget scrutiny and automated model routing. Systems will automatically route simple process steps to lower-cost models, reserving expensive frontier reasoning for tasks that strictly require it. Frameworks will also prioritize open-standard architectures to successfully decouple core business logic from specific AI vendors. This independence ensures that an organization's operational context remains fully auditable even when technology providers change.

Many enterprises are exploring agentic AI at scale. What are the biggest challenges organisations face when balancing advanced AI capabilities with rising compute costs, and how can context-aware systems help address this challenge?

Scaling autonomous agents creates a "tokenmaxxing" problem, where continuous queries and bloated prompts rapidly drive up inference costs. Without guidance, agents waste compute repeatedly trying to learn core workflows from scratch. Context-aware architectures solve this by acting as a control layer that extracts only relevant data and eliminates redundant reasoning loops. This provides a calculator for business logic, which significantly reduces error rates and eliminates costly silent retries.

There is growing discussion around the emergence of internal "AI class systems," where access to premium AI models may vary across departments, functions, or geographies. How do you see this trend developing, and what implications could it have for enterprise productivity, innovation, and workforce equity?

Rising compute costs may prompt organizations to restrict premium model access, creating operational inequalities across different departments or geographies. This tiered access threatens to negatively impact overall enterprise productivity, workforce equity, and localized innovation. Instead of rationing usage, organizations should focus on structural token efficiency. Grounding smaller, less resource-intensive models in deep operational context allows all teams to achieve highly precise results affordably.

Looking ahead, how do you envision the relationship between process intelligence, contextual data, and AI agents evolving, and what will distinguish the organisations that achieve sustainable competitive advantage from those that struggle to realise value from AI investments?

Process intelligence, contextual data, and AI agents will merge into a complete operational loop providing hindsight, insight, and foresight. Winning organizations will integrate AI directly into their operational reality using a system-agnostic digital twin. They will treat operational context as a compounding balance sheet asset that makes each subsequent deployment more efficient – both in terms of results and costs. Conversely, companies relying on context-blind agents will struggle with compounding errors and fail to realize meaningful value.