In CX, artificial intelligence (AI) is often sold as a cost saver. But whether due to a lack of governance, expensive mistakes around data readiness or the absence of a framework by which to measure returns, it can easily fall short of delivering the anticipated financial returns.
A 2025 report from the Massachusetts Institute of Technology (MIT) concluded that 95 percent of enterprise AI and generative AI pilots fail to deliver measurable financial returns. At a time when 52 percent of CX practitioners have told CX Network the pressure to prove ROI is increasing, figures like this are worth taking note of.
"Not every AI deployment achieves the return organizations expect," says Steve Blood, VP of market intelligence, Five9. "But in most cases, underperformance isn't a technology problem. It's a planning and measurement problem."
There are common failure patterns, Blood explains, such as:
- Deploying AI without the clear definitions of success.
- Misaligned expectations around timelines.
- Underestimating the importance of integration.
Drawing on insights from the CX Network report Balancing the books on AI, this article explains why the ROI of AI is about more than money spent and labor saved, the importance of knowing when to measure ROI and how to understand billing structures and "drastically reduce" token use.
To find out more about this topic, download Balancing the books on AI here.
ROI is about more than money spent and labor saved
To generate a return on investment (ROI) that meets expectations, practitioners must first understand the full range of indirect costs associated with AI.
In the customer service suite, this often involves the need to upskill or retrain teams, the cost of reducing headcount, time and resource implications of robust testing and governance, or the need to ensure vigorous data security.
During CX Network's 2026 research into the state of CX practitioners were asked to select which changes they expect generative and agentic AI will bring to their organizations. The results showed:
- At 36 percent, the requirement for teams to upskill was the most selected.
- Paving the way for digital employees to work alongside human employees was selected by 32 percent of respondents.
- The need for new operating systems and back-end processes was selected by 30 percent of respondents
Here is how the cost of such additional considerations can be anticipated and managed:
Workforce readiness
"The most commonly overlooked factor is workforce readiness," says Blood. "AI changes how agents work, not just how much they work. That shift requires investment in training, change management, and in some cases, role redesign. Organizations that treat this as an afterthought tend to see slower adoption, lower utilization, and ultimately weaker returns."
Blood adds that the 36 percent of practitioners in the research who expect AI to require team upskilling, "are right to flag this".
"Building that cost into the business case from the start leads to more accurate projections and better outcomes," he says.
There's also the question of "organizational momentum". Blood says: "AI deployments that lack executive sponsorship or cross-functional alignment tend to stall. The indirect cost of slow decision-making, competing priorities, and siloed ownership is hard to quantify but very real in its impact on time-to-value," he explains. In a similar way, the resource cost of involving multiple departments in deployment and roll out changes the calculation again.
Data quality
Data quality is another factor that is often overlooked in initial ROI calculations.
"AI is only as good as the data it operates on. If an organization's customer data is siloed, inconsistent, or incomplete, there will be a cost, in time, resource, and sometimes third-party support, to get it to a standard where AI can perform effectively," Blood warns.
Governance and compliance
Governance and compliance carry real resource implications, particularly in regulated industries. This means testing, auditing, and maintaining oversight of AI systems isn't free, "and in sectors like financial services or healthcare, the requirements are substantial," Blood says.
When AI is deployed without the correct guardrails and testing it leads to trust erosion, operational friction, or weaker decision quality. For organizations operating in the EU, there could also be financial penalties under the EU AI Act.
"The indirect costs of AI deployment often sit on the loss side of ROI," says Chicago-based AI governance advisor Dawn Kristy JD.
"They may not appear in the initial business case, but tend to surface later through trust erosion, operational friction, or weaker decision quality," she says.
Documentation is another potential cost related to governance and compliance. If a customer, regulator, insurer, or board later asks why a decision was made, the final output may not be enough.
Kristy says: "The organization may also need to show the path behind the decision. What data was used? What assumptions were made? Who reviewed the result? How were exceptions handled? Was human intervention possible?"
Clear decision-making authority
One major cost is unclear decision-making authority. Kristy says: "When AI outputs move quickly through a business, organizations need clear rules for reliance, verification, escalation, and intervention. Who can rely on the output? Who must verify it? Who can stop or escalate a decision? Without that clarity, speed can create exposure."
In short, the organizations that achieve the strongest returns on their AI investments "are those that account for all of this upfront; treating the indirect costs not as surprises to be managed, but as inputs to be planned for," Blood concludes.
Kristy adds: "The real ROI question goes beyond speed. What might the business stop seeing, questioning, or documenting during AI integration?"
AI billing structures and tokens
Many AI vendors structure their billing around metered use and subscriptions. Providers may also apply outcome-based pricing structures, where users pay when AI successfully resolves an issue, saves a cancellation, or drives an upsell.
These structures can add a new layer of complexity to ROI calculations. If a business has a churn reduction or upsell target to achieve, the practitioner must balance the cost of using AI to achieve this target against the revenue generated by achieving the target.
Ebrahim Hyder, VP of customer care for Michael Kors and a CX Network Advisory Board member, says CX leaders must proceed carefully.
"During my own agentic AI research, I encountered structures ranging from per-minute billing with added fees when the agentic bot resolves an inquiry, to flat per-resolution pricing, to mandatory annual minimum commitments tied to projected usage," he explains.
"Understanding how each model scales as adoption grows is essential to avoiding surprises."
Token costs quickly add up, too. Mrunal Gangrade, who also sits on the CX Network Advisory Board is a researcher and engineering VP who has worked for Larsen & Toubro Infotech, Citibank and JP Morgan Chase & Co. She says that building knowledge locally can drastically reduce overall token usage and therefore, costs.
This method of cost control can be applied within and beyond the contact center.
Step one is to monitor how employees are using AI tools to assist in their work. By way of example, Gangrade says that by understanding the prompts and questions employees use, leaders have the opportunity to build locally accessible knowledge bases to avoid incurring additional token costs for the same queries.
"If today I search for ABC and again tomorrow I search for ABC there would be two token hits going outside. But this is static data.
"If you maintain a repository of the knowledge internally, on the second day when I come and search for ABC then it shouldn't go out and utilize your token," she explains.
"Internal employees should be trained on how to maximize the use of AI with fewer tokens because that builds up your own agents locally. This means you can fetch the information from your local agents rather than utilizing the tokens which is going to exponentially increase the cost," she adds.
The bigger picture on AI returns
Organizations that measure ROI in isolation are missing the bigger picture. Blood explains: "The impact [of AI] depends on how well it connects with existing systems, data sources, and human workflows. A well-designed AI solution sitting on top of fragmented or incomplete data will always underperform."
It can also depend on when – not how – ROI measurements are made.
According to Blood, knowing when to measure ROI is just as important as how. This is because AI solutions involving large language models (LLMs) or agentic workflows, often require a period of tuning and learning before they deliver at full capacity. "Organizations that measure ROI too early, before the solution has had time to mature in their specific environment, can draw the wrong conclusions," Blood says.
That said, when deployments are well-planned, the returns are real and measurable.
"We see organizations consistently achieving meaningful reductions in average handle time, increases in first contact resolution, and significant deflection of routine queries; all of which translate directly to cost savings and improved customer satisfaction," Blood explains.
"The key is approaching AI deployment with the same rigor you'd apply to any major operational investment: clear objectives, robust data, a defined measurement framework, and the patience to let the solution perform before drawing conclusions," he adds.