Analyzing ROI and AI
getty
Appetite for investment in AI continues to grow, but value remains elusive. In fact, Gartner finds that just one in 50 AI investments deliver transformational value and only one in five generates any measurable return on investment (ROI).
Despite billions in investment, many companies are struggling to see value from AI. Even the world’s largest AI startup, OpenAI, ending up shutting down its video generation app Sora in March which reportedly burned through $1 million each day.
Leaders who struggle to tie adoption to tangible business value will face increased scrutiny in the future. As a result, enterprises will need to be much more targeted and intentional with AI adoption if they want to demonstrate clear ROI.
Study after study shows business leaders are failing to connect AI adoption with value. A PwC survey released in January 2026 surveyed 4,454 chief executives across 95 countries and found that just 30% reported an increase in revenue from AI in the previous 12 months. At the same time, 56% of CEOs claimed AI delivered zero cost or revenue improvements.
AI initiatives and investment are under increasing scrutiny to provide results, and many are falling short. At the extreme end of the spectrum, OpenAI just missed its internal goal of reaching one billion weekly active users for ChatGPT by the end of 2025.
Sarah Friar, OpenAI’s chief financial officer, also reportedly raised concerns that the company may be unable to pay for future computing contracts if revenue growth fails to accelerate.
The fact that a leading AI startup is struggling to raise sufficient revenue highlights the need for business leaders to be cautious in managing the cost and implementation of AI initiatives and pilots.
Being able to measure the impact of AI adoption is critical to assessing a pilot’s value. Nathan Irby, principal enablement and innovation strategist at Snowflake who leads GTM enablement, has been working extensively on extracting practical ROI from AI pilots and initiatives.
He reportedly helped roll out AI roleplays to 3,000 sellers on Yoodli, an experential learning platform and roleplay provider for communication coaching. Irby tracked the outcomes from the project and claims it saved over 1,200 manager hours per quarter while offering approximately $700,000 in annual savings.
“With 94% of reps participating, we eliminated 1,600 hours of managing grading by shifting scoring to AI, enabling managers to focus on real coaching and driving an estimated 4-5x ROI on time and spend. We ran these roleplays on Yoodli’s AI platform, which handled automated scoring at scale across 3,000 reps,” Irby told me via email.
Irby claims Snowflake built a four-part AI roleplay series to help CAEs better prepare for conversations with CFOs and procurement teams, seeing improvements in key skills like discovery and negotiation, while focusing on deal growth and time to close in the long term.
But how does Snowflake approach the ROI for AI initiatives more broadly? Irby explains that he starts by matching a problem statement and a hypothesis with a goal, such as time savings, skill improvement or time to launch. From there, he reviews from a short-term and long-term perspective.
In the short term, he looks to assess whether employees are using it and improving. This takes into account completion rates, pass rates, attempt counts and CSAT. For the long term, he considers metrics such as upskilling, time to close and manager hours saved.
“The clearest signal is almost always time returned to people. Hours saved, mapped to spend, gives you a ratio you can defend. I like to validate with pilot groups, see what breaks, figure out fixes and then scale for launch. I lean hard into outcomes over completions. The point is to equip someone to do something better than before and/or to save them time,” Irby said.
He also shared that he is now using Snowflake Cortex Code as an automation assistant to help ship initiatives faster. More specifically, while it used to take 3-4 weeks to build AI roleplays, it now takes hours to build roleplays that anyone across the enablement and GTM teams can leverage.
Vanity AI adoption is on its way out, but those companies that have focused on streamlining core business processes have reported significant ROI. For instance, JP Morgan claims to have used AI to save the bank almost $1.5 billion through fraud prevention, personalization, trading, operational efficiencies and credit decisions.
Similarly, after Klarna launched its AI assistant powered by OpenAI, the company noted the assistant had automated two-thirds of Klarna’s customer service chats and estimated it would bring approximately $40 million in profit in 2024.
In 2025, business communication and contact center provider RingCentral generated $2.5 billion in revenue and achieved $100 million in ARR from new AI-led products, including its agentic voice product AI receptionist.
While Kira Makagon, president and COO of RingCentral, said the company doesn’t report on ROI in terms of dollars, she claims many downstream organizations have doubled their business through being able to handle more incoming calls.
For Makagon, experimentation has been a key driver for innovation. “As much as I would love to have a single point to be able to manage everything, it’s impossible. People experiment, and you've got to let them experiment,” she told me in a video call.
Makagon noted that while users can experiment, they can’t take data outside of approved tools. If the tool isn’t approved and a user wants to experiment, they can do so in a sandbox with synthetic data. This cuts down on potential risk when developing pilots.
We’ve seen a number of pilots completely collapse in recent years. One of the most high-profile examples occurred in 2024 when McDonald’s decided to remove its AI-powered ordering technology from drive-through restaurants in the US, after customers shared viral videos of orders being misinterpreted. It’s not just about rolling out AI, but doing it in a way that’s cost-effective and reliable.
When applied to narrow processes, AI has the potential to deliver significant ROI, but the value generation is dependent on whether the process is a fit for automation. McDonald’s pilot failed because it didn’t fully take into account the potential for voice recognition technology to underperform in external environments with lots of background sound.
When applying AI to a workflow that is a fit for automation, time taken to complete a workflow, number of tasks completed per employee, error reduction, accuracy and employee satisfaction can all be used to indicate performance. Headcount can also be another factor. For example, if a smaller team with AI can match the productivity of a large team without AI, this should be taken into account too.
Updated May 1st, 2026 with information about OpenAI’s challenges meeting internal goals.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。