Token costs a concern for many customers, says TCS CEO K Krithivasan

Krithivasan and Seksaria discuss artificial intelligence revenue, rising token costs, and increasing AI spends

Tata Consultancy Services (TCS), India’s largest IT services firm, saw its share price surge due to better-than-expected show during a tepid growth environment in Q1FY27. K Krithivasan, chief executive officer (CEO) and managing director (MD), is confident that FY27 will be better than FY26. In an interview with Shivani Shinde in Mumbai, Krithivasan and chief financial officer (CFO) Samir Seksaria discuss artificial intelligence (AI) revenue, rising token costs, and increasing AI spends. Edited excerpts:

 

TCS chairman N Chandrasekaran made an interesting remark that the company will eventually have an equal number of AI agents and employees. Operationally, what does that mean?

 

Krithivasan: Today, we already have more than 2,500 AI agents deployed across customer engagements. Each agent can have multiple instances running across different use cases. We have also opened the infrastructure to all our associates.

Many employees are creating their own productivity agents. Some of these may eventually become part of client delivery, while others remain personal productivity tools. As the volume of work increases, we do not believe headcount will grow linearly because agents will perform an increasingly significant portion of the work. The idea is that, as the model matures, we will be able to do more work with the same number of people, or at least avoid increasing headcount at the same pace as business growth.

  The principle is “human plus agent in the loop.” Agents improve productivity, help clear backlogs faster and deliver better outcomes. That, in turn, expands the addressable market and creates more business opportunities.

 

As you use more AI agents, wouldn't your costs increase?

 

Seksaria: There are two aspects to this. First is productivity. Our objective is that the first right of refusal should be to the agent. Only if the agent cannot complete the task should it go to an employee. This leads to efficiency. 

Second is investment. We have invested in multiple frontier models. We have Copilot, Claude, Gemini and other models available across the organisation, allowing employees to choose the model that best suits their work. 

Add to this is our investment in the AI ecosystem and our partnerships. This is one of the costs. These investments are increasing, but much of the spending is reflected as operating expenses in the profit and loss (P&L) account rather than as capital expenditure. Some part of cost is token-related. 

In some cases, the investment is customer-led because we use the customer's infrastructure, especially when deploying solutions in their environment. 

In other cases, where we manage end-to-end service delivery, we incur those costs ourselves while also putting the necessary guardrails in place.

 

As you increasingly use models such as Anthropic and Mistral, will TCS itself require more large language model (LLM) capacity and compute — not just for clients, but for internal use as well?

 

Krithivasan: We take a very pragmatic approach. We evaluate what type of LLM is required for a particular workload. Even within a single model family, there are multiple versions. We select the model that is most appropriate for a specific use case. Wherever suitable, we also use open-source or on-premise models. We don't rely on a single model.

  It may be tempting to use the latest frontier model for every workload. But if a much cheaper model — or even a free one — can deliver 80-90 per cent of the required outcome, there is no reason to use the most expensive option.

  This requires continuous calibration of the models available in the market. We follow the same approach with customers. Most customers do not depend on a single model; they use multiple models. As system integrators, our job is to advise them on the most efficient way to optimise costs.

 

There has been concern around rising token costs. We have also heard of companies scaling back AI projects because of this. Is that becoming a concern among your customers, and how are you addressing it?

 

Krithivasan: The cost of tokens is becoming a concern for many customers. That is why model FinOps, or token FinOps, is becoming a significant opportunity for system integrators. We can help customers determine where they should use an LLM, where a smaller language model is sufficient, what should run on Cloud, what should run on-premise, and which model is best suited for each workload.

  Since we understand both the customer's environment and their objectives, we are able to deliver greater value.

I don't think customers are pulling back from AI projects. Most customers are seeing genuine value from AI. The discussion is about how to use AI efficiently, not whether to use it at all.

 

This quarter TCS signed some large AI-led deals, such as SKF. At the same time, you also said AI deals continue to be relatively small and need to be replenished continuously. Can you explain the difference?

 

Krithivasan: The answer lies in how we define AI revenue. If we deliver a project using AI coding assistants, it does not automatically become AI revenue. However, if we build an application that uses AI — for example, to make credit decisioning faster — that would qualify as AI revenue. We have a sharp definition of what is AI revenue. 

When we refer to AI revenue, we are primarily talking about deals that are either transformational deals where AI is used for either tech modernisation or business process transformation. 

Many of these transformation projects are relatively short in duration — sometimes lasting one, two or three quarters. Their ticket sizes are often smaller than traditional outsourcing contracts, although we are seeing deal sizes gradually increase. Unlike annuity deals, these projects have to be continuously replenished once they are completed. 

SKF is an AI-led deal because AI will be embedded throughout the engagement. We will deliver services ranging from SAP S/4HANA transformation and application development, while using AI to deliver those services better, faster and more efficiently. 

At the same time, we are helping SKF become an AI-led organisation. That is why we describe it as an AI-led deal, even though our reported AI revenue is based on a much narrower definition.

 

Can you give us an update on HyperVault?

 

Krithivasan: We are at a very advanced stage of discussions with a couple of hyperscalers. I am hopeful that by Q2, we will be able to announce definitive agreements, along with the locations where work will begin. Overall, progress has been good. 

We are in advanced discussions with several potential users, most of whom are hyperscalers or AI-native companies.