Why CEOs Fear AI Costs as Firms Burn Through Budgets

When a corporate board pours millions of dollars into generic AI licences hoping for magic to happen, it usually only takes a few quarters for the lack of return to set in.

“People went out and bought 50,000 Copilot licences, threw them at every person in their company and hoped for the best,” says Jed Dougherty, Senior Vice President of AI and Platform at Dataiku.

“That burned a whole bunch of money, didn’t go anywhere and no one saw any gains from it. Now, the shift has come from generic tools to personal accountability.”

At the executive level, that accountability has translated into outright anxiety. According to Dataiku’s Global AI Confessions Report: CEO Edition, 2026, 78% of CEOs fear AI could cost them their jobs. 

They find themselves personally accountable for AI outcomes, despite not fully trusting or controlling the underlying systems.

Shifting from hype to unit economics 

For non-technical C-suite executives trapped in the black box dilemma, evaluating an AI system can feel impossible. 

However, Jed argues that leaders do not need a computer science degree to govern the technology safely – they can apply foundational business logic to the unit economics of AI.

“A good AI strategy is one that replicates or replaces an aspect of your existing business,” he says.

“If I augment an insurance claims processing workflow with generative AI, does the operational cost go down or does the revenue go up? As long as you aren’t just building little toys, if you are measuring your business, you are measuring your AI.”

This shift means moving away from viewing an agent as a single entity and instead treating it as an agentic process. 

In a standard workflow – like claims auditing or actuarial analysis – a series of specialised agents sit inside a larger, deterministic system to support human teams rather than replace them.

Failing to build these boundaries leads to capital inefficiencies. Jed highlights the astronomical cost of unconstrained compute: “We are all burning compute. Uber burned through its entire AI annual budget in the first three months of this year alone.

“We’re all absolutely burning through it and I think most organisations are still in the ‘I hope this pans out stage’ because everyone using it [AI] feels so productive that you can’t turn it off.

“But the places where we’re seeing real value right now is largely in software development. It’s funny, the first thing it got really good at was the thing that created it.”

The reality of shadow AI

The democratisation of frontier models has brought a familiar threat back to the forefront: shadow AI. 

“Shadow AI is here because every single software vendor has decided to inject AI everywhere,” says Jed. “Whether you log into Salesforce, Snowflake, AWS or Workday, every platform lets users build agents. 

“It’s a reality we can’t put back in the box because we aren’t going to stop offering it, because users want it.”

The solution requires a centralised agent management system capable of scanning internal infrastructure to index exactly how many autonomous tools are running under the hood.

Piercing through ‘code slop’

The primary bottleneck preventing enterprises from moving past the proof-of-concept phase isn’t raw data infrastructure – it is trust. 

While gen AI is remarkably efficient at writing front-end user interfaces, verifying the underlying backend data processing remains highly complex.

Jed warns of an emerging industry crisis known as ‘code slop’:

“Gen AI can generate thousands or millions of lines of code to build a backend pipeline.

“For a human developer to trust what it’s doing, they have to read every single line of that code, which is an insane burden. 

“Code contains a lot of information per word. Gen AI lies and hallucinates in code, just like it does in text, so we have to be vigilant.”

To bridge this trust gap, Dataiku believes AI needs to communicate via visual architecture rather than blocks of text. 

“When I ask an agent to build something in Dataiku, instead of generating thousands of lines of code, it generates a visual description of what it’s done.

“I can look at that description and understand in minutes, rather than hours, days or never. That is how we move from all these toys to true production grade workflows.”

When an enterprise user asks an agent to build a data pipeline on Dataiku, the platform generates a visual blueprint of the workflow. 

Users can audit, understand and verify the data pathways in minutes rather than spending days reviewing raw code.

Bridging the gap with conversational building

This visual, trust-first approach is the driving force behind Dataiku Co-build, which was released last week. 

Jed describes the tool as the “ChatGPT for data with a visual layer that describes how it came to its answers”.

Rather than forcing users to blindly trust text-based outputs or sift through complex code, Co-build allows teams to interact with their data systems through natural conversation without losing oversight.

“It’s really the culmination of a lot of what we’ve talked about here,” Jed says. “It’s a way to conversationally build visual workflows in Dataiku.”

Ultimately, shifting from “toy” applications to production-grade workflows relies on tools that make human-AI collaboration faster, safer and cleaner. For Jed, the proof is already in the productivity gains.

“I’ve been using it for the last four months and it has dramatically sped and cleaned up the way I work.”