The AI blank cheque is over: Customers want partners to prove the ROI

Enterprise customers are asking harder questions about AI costs and returns, says Infobip

Enterprise customers are taking a more pragmatic approach to AI, and demanding clearer evidence of its business value, according to Infobip chief innovation officer Krešo Žmak.

Speaking at Infobip Shift in Zadar, Croatia, Žmak said businesses have become more knowledgeable about AI over the past year, while finance leaders are scrutinising its costs.

“A year ago, every middle management in the world got a KPI: implement AI,” he said.

“20 per cent knew what AI could be used for. 80 per cent just had the KPI.”

Customers now have a clearer idea of what they want to achieve, he said, while boards and CFOs are increasingly prepared to stop funding initiatives that cannot demonstrate a return.

As a global cloud communications platform provider, Infobip is embedding AI across its products and internal operations as part of what it describes as an “AI-first” transformation.

The shift comes as businesses face growing pressure to translate AI adoption into measurable outcomes.

CloudZero research found that 87 per cent said they needed to connect AI spending with business outcomes within the next year, but only 22 per cent could currently do so.

That doesn’t mean AI is failing to produce value.

A Dun & Bradstreet survey of 10,000 businesses found that more than three-quarters reported some measurable ROI.

However, almost half described this as “pockets of ROI”, while 28 per cent reported broad or strong returns across multiple projects.

Not every problem needs the biggest model

Elsewhere, Benoit Schillings, VP of technology at Google DeepMind, told software developers attending Shift that he expects AI systems to become more selective about where and how they process requests.

Rather than sending every query to a large cloud-based model, he predicted a hierarchy in which smaller models running locally handle straightforward tasks and escalate more difficult problems.

“You’re going to have a hierarchy of intelligence, where your local device, your phone, can do pretty mundane things,” said Schillings.

He also expects systems to become better at matching computational effort to the difficulty of a task, with implications for latency and energy consumption.

“You can probably cut your power consumption by 5x just by tuning correctly the complexity of the process that you use to solve the problem,” he said.

Infobip chief financial officer Damir Prusac similarly expects businesses to become more selective about the models they deploy.

“LLMs are becoming commodity,” he said.

“If you are a manufacturer of a specific type of machinery, you don’t need the LLM. You need something specific.”

To that end, Prusac anticipates greater use of specialised models running on local infrastructure rather than relying exclusively on large, general-purpose cloud-based models.

A new role for partners

For enterprises, these developments add another dimension to the build-versus-buy decision.

Žmak said customers are increasingly looking to technology partners for AI solutions, including agents and hyper personalisation, while simultaneously using AI development tools to build more software themselves.

For channel partners, the opportunity could therefore extend beyond selling AI products to helping customers determine which problems warrant AI, what technology they need and whether the resulting value justifies its cost.

That includes identifying use cases, deciding whether to build or buy, integrating AI into existing systems and establishing how outcomes will be measured before projects are scaled.