Synopsis
As AI adoption grows, IT service providers and clients are grappling with calculating the cost-benefit of token usage. With a shift towards usage-based models, companies are focusing on measuring the actual business impact of AI spend, acknowledging that direct one-to-one benefits aren't always guaranteed during this innovation phase.AI usage costs are rising and with companies such as Anthropic and Github moving to a usage-based model rather than a subscription model, this would require IT companies and their clients to get a better idea of what the value of the token usage works out to in the end.
At Infosys, while the company is encouraging ‘massive token usage’ at present, the company also accepts the fact that there won’t necessarily be a one-is-to-one benefit from the spend on tokens, CEO Salil Parekh told ET.
“We absolutely want to measure and encourage token usage because today it's an innovation phase. Some of that will result in impact, some of that will be learning. But the eventual impact also needs to be kept in mind,” Parekh said in a recent interview.
He added that token usage was an ‘intermediate measure’ and that the company was putting in place processes to study the end-impact of the use.
While IT companies can set token guidelines for their own internal projects, delivering IT projects to clients who have cost-and-productivity goals is more complicated.
“From a business standpoint, the model is still evolving. Token consumption is not always predictable—sometimes we underuse, and sometimes we exceed estimates. As an organization, we are continuously improving how we measure the relationship between token consumption and the business outcomes,” NS Kumar, chief digital officer, Hitachi Digital Services, told ET.
He added that operationally the company tracked and provided visibility to its teams on token usage. For example, on a large legacy platform modernization engagement, it reserved tokens for each scrum team every month and provided consumption dashboards to the project leads.
While large clients typically provide the tools and define a pool of licenses and token limits, the way the deals are structured can impact where the cost falls.
“Most pure AI deals are currently getting structured either in a fixed fee model, or as a fixed fee for the implementation plus a variable fee that is tied to the volume of events that trigger the AI. In the former, service providers are estimating the costs of different elements of the AI system such as compute, API, tokens, and baking them into the deal up to a certain level of consumption,” Rahul Gehani, Partner at Everest Group, said.
He added that in some cases, for highly competitive deals, clients have been able to push providers and lock-in a price for the entire deal term.
“For service providers, this does create a structural risk as the price is getting locked in but the actual AI costs are uncertain and could spike over the deal term,” Gehani said.
