Employers are finding that AI returns often show up first as faster workflows, fewer avoidable checks and better client alignment. The harder task is connecting those operational wins to business value before projects scale.
For managers trying to justify AI spending, the return is not always a clean line on a profit and loss statement. The early evidence from several organisations suggests a more practical test: can AI remove a costly step, speed a team decision, or help employees deliver better work without weakening human judgment?
Alight Solutions, a benefits administrator, offers a useful example. The company was already familiar with AI, but in 2025 it began evaluating a newer type of tool. In September, Alight joined a beta test for Phenom's fraud detection recruiting agent, designed to spot suspicious job applicants earlier in the hiring process.
The test produced a quick signal. The agent found that one applicant had submitted for the same role twice, using a different name and email address on each application.
“Even by catching that one person that we ultimately didn't hire, that was enough for us to say, ‘It's going to work for us,’” said Julie Eagy, talent acquisition operations manager at Alight.
Avoiding a needless background check could save money. Eagy said the larger benefit, however, was time saved by the recruitment team. That distinction matters for employers. Many AI deployments will not immediately create new revenue, but they may reduce rework, prevent avoidable processes and free staff to focus on higher-value decisions.
Why the value case is still hard
AI returns remain uneven across large organisations. Arun Chandrasekaran, an analyst at Gartner, described the issue plainly: “It’s a very complex discussion.” He added, “Broadly speaking, I would argue that a lot of customers are still struggling to articulate clear business value from AI.”
According to Chandrasekaran, companies have seen productivity improvements in software engineering, customer service and back-office functions such as finance and HR. The harder part is proving how those gains translate into the outcomes boards usually care about, such as lower costs or higher revenue.
“Alignment between productivity gain and these value-driven metrics that businesses care about, that’s still a work in progress, I would argue, across a lot of organizations,” he said.
There are also cautionary cases. In May, a Pizza Hut franchisee sued the franchise, claiming that an AI delivery management platform had cost it millions of dollars instead of producing savings. The case underlines a basic governance point for leaders: a tool that optimises one workflow can still damage the economics of the wider operation if incentives, data and process design are wrong.
Gartner has also reported that around half of generative AI projects were abandoned after proof of concept in 2026, citing issues including poor data quality, rising costs and vague business value.
Smaller firms are tying AI to workflow gains
Some smaller organisations appear to be finding clearer paths from AI use to business results because the workflow and the commercial outcome sit closer together.
OBI Creative, a 25-year-old advertising agency in Omaha, Nebraska, has fewer than 50 employees and serves clients including TiVo and HD Radio. Founder and CEO Mary Ann O’Brien said agentic AI has helped the agency extend services from one industry to another more efficiently.
“Agentic AI has become a really great efficiency opportunity for us to scale some of our offerings from one industry to another,” she said.
The agency built tools for specific internal jobs. One monitors the health of websites created by its digital team and checks whether they remain current. The agency used models including Gemma and Qwen to code prompts that generate website health reports, and also used OpenAI GPT models and Anthropic Claude models to code prompts into its tools.
A second tool helps creative staff assess whether a campaign matches a client's stated goals. It began as an internal prototype for brand strategy and creative alignment, then became a product clients wanted to license.
O’Brien said the effect has shown up in margins. “We don’t have to go back to the client five times for changes,” she said. She also said that by using agentic AI tools, the agency “will have at least 20% year-over-year growth.”
The investment has not been free. O’Brien said overhead has increased, though efficiencies have quickly offset that. The agency started with models that had fixed costs and subscription services, giving employees a defined budget to work within.
The human side was equally deliberate. O’Brien told staff that generative and agentic AI tools were not a threat to their jobs. She also challenged one employee to become an efficiency ambassador and make the results part of his performance.
“There are still people in my agency who are very afraid of the tools, and so it's just really getting people comfortable and being able to trust that the thing between their ears is still the most valuable to our clients,” O’Brien said. “But the tools are there, and why wouldn't you use them?”
Start with the metric, not the model
Cornell University has taken a broader institutional approach. Ayham Boucher, head of AI innovations for Cornell Information Technologies, said the university treats AI as a tool. Faculty were given access to frontier models in a secure, private environment, but use is not required. Cornell uses AWS for creating, managing and deploying agents, alongside user-facing tools such as Microsoft Copilot and Claude Desktop.
Boucher said the objective is to extend human capability. “We are not superb creatures,” he said, pointing to limits in human memory and capacity. AI, in his view, is another technology that can expand what people are able to do.
Cornell is not measuring success through “tokenmaxxing,” the practice of tracking value partly by the volume of AI model tokens consumed. Boucher said outcomes are tied to the university's mission, including students served and scientific discoveries made.
For business leaders, the practical interpretation is clear: define value before choosing the use case. Chandrasekaran said companies should connect operational metrics to value metrics. In software, for example, counting lines of code is less useful than measuring the speed at which new features or capabilities reach an application.
He also warned against launching a use case first and deciding later how to measure it. For most enterprise applications, he said, organisations should expect to reach ROI within a year. The larger lesson is not that every AI project needs a narrow cost-cutting target. It is that human-led teams need a specific outcome, a measured workflow and a clear reason for using AI before the pilot begins.
Reported by Hybrion Insights with reference to AI Business.
hybrion
Better together. Built for what's next.
Stay updated
Get insights on AI, teams, and the future of work.