Enterprises are on track to spend $2.5 trillion on AI in 2026, yet 95% of GenAI pilots fail to scale, according to research from Gartner and MIT. Only 6% of enterprises are creating real business value from AI, according to McKinsey, although an additional 82% are using the technology.  

At Ai4 2026, Gururaj Bhat, Head of Evalueserve’s Data & AI Business, coached the audience to see this gap between AI adoption and ROI not as a model shortfall but a starting-point issue. 

Start with the workflow, not just the use case.

The common advice to “pick the right AI use case” isn’t wrong, Bhat said, but he encouraged the audience to think beyond use cases by picking the right first workflow. In his framing, a single workflow can contain several potential use cases. By thoroughly understanding the workflow first, you can best identify which use case inside it will move the needle for the business. Skip that step, Bhat explained, and you risk applying AI to the wrong problem. 

The good news is that more pilots are reaching production now. But the 6% of enterprises that achieve real business value tend to understand the workflow before they touch the technology. Knowing the process inside and out is what helps them choose a use case that’s both valuable and measurable.  

Measure outcomes, not activity.

Many companies today track AI activity, like user counts and the number of pilots, which Bhat called “the ROI illusion.” He encouraged leaders to move beyond that, since AI activity alone isn’t an outcome.  

Bhat outlined three things that businesses should be tracking:  

1. Adoption  
2. Proficiency: Is the AI changing how fast or how well people work? 
3. Financial Outcome(s): This is the main question a CFO will have – what changed in the business? 

Bhat said the time to consider AI ROI is before the use case is implemented, as the org structure needs to be designed with it in mind, with baseline metrics and a measurement plan in place. Once the use case is implemented and leadership asks to see results, you’re already behind.

Tokenomics: Not every task needs a frontier model.

A second theme of the talk was cost discipline. With open-weight models proliferating, Bhat explained that enterprises don’t need to route every prompt to the most expensive frontier model. Instead, they need a mechanism to match the right model to the right task, which requires a common orchestration layer and a unified data layer to feed it context. Bhat also pointed to knowledge graphs, not only data lakes and warehouses, as the structure agentic AI needs to ground its actions in enterprise reality. 

Contemplate the right diagnostic questions.

To find where AI can make the most difference, Bhat suggested considering the following questions: 

1. Where is cognitive load highest? 
2. Where do errors cost money? 
3. Where does delay create risk? 
4. Where do edge cases break automation? 
5. Where should humans stay in the loop? 

Bring agentic squads into the workflow.

Bhat said that effective AI deployment takes a blend of roles working together, not a single engineer or a single tool – domain experts who surface edge cases, someone managing model routing to control token costs, engineers building the actual agentic workflow, and change management focused on adoption at the workflow level. Forward-deployed engineers (FDEs) bring real technical skill, he noted, but without process knowledge, even they can run into limits.  

Bhat pointed out that, to achieve high adoption, it’s best to pitch AI’s benefits in the end users’ specific language, whether that person summarizes legal documents or closes the books each month.  

This is the thinking behind Evalueserve’s Agentic Squads, which we announced last week. These squads bring together domain experts, FDEs, and change and adoption specialists around a single enterprise workflow, so AI gets built into the way work really happens. Learn more about our Agentic Squads here. 

Bhat’s closing message? Pick the workflow, define what success looks like before you start, get the model and cost layer right, and bring the right mix of people to make the change stick.  

You can watch Bhat’s full Ai4 talk here on YouTube.