Beyond Time Savings: A Business Value Framework for Evaluating Agentic Automation

Why RPA-Era ROI Models No Longer Work

For the past decade, enterprises have evaluated process automation through almost a single lens: how many work hours were saved. This ROI model, rooted in the RPA (Robotic Process Automation) era, reduces automation to a simple equation — "labor cost × hours saved." But as the primary driver of automation shifts from rule-based script robots to reasoning-capable AI agents, this formula systematically undervalues what's actually being created.

The core problem is that RPA-era ROI models miss the majority of the value that agentic automation generates. RPA is fundamentally a mechanical replication of deterministic, repetitive tasks — its value ceiling is defined by the human hours it replaces. Agents, by contrast, can handle exceptions, exercise judgment, and continuously learn. The value these capabilities deliver simply cannot be measured in hours saved.

Four Dimensions of Agentic Value

To truly quantify the business value of agentic automation, organizations need to move beyond the single dimension of time savings and adopt a more complete evaluation framework. For leaders at AI Centers of Excellence (CoE), the following four-dimensional framework offers a more rigorous basis for assessment.

Time Savings

This is the most traditional and intuitive dimension — and the only one that was ever seriously calculated in the RPA era. Agents take over tasks previously performed by humans, directly freeing up work hours. But this should be treated as the starting point of value assessment, not the endpoint. Evaluating automation on this dimension alone is like navigating a new continent with an old map.

Exception Handling

This is one of the key capabilities that distinguishes agents from traditional RPA. When RPA encounters situations outside its predefined rules, it typically halts and requires human intervention. Agents, by contrast, can understand context and flexibly handle edge cases. The value of exception handling lies in reducing process interruptions and decreasing the frequency of human escalations — value that is almost entirely invisible in legacy models.

From a technical standpoint, RPA's limitations in exception handling stem from its rule-based automation nature — developers must anticipate every possible input scenario when writing scripts, and anything outside those parameters triggers an error or suspension. Industry data suggests that exception rates in mature RPA deployments typically run between 5% and 15%, meaning 5 to 15 out of every 100 transactions require human intervention. Agents, leveraging large language models' contextual understanding, can autonomously determine the "most reasonable next step" within a certain range — internalizing exception handling as part of the process itself rather than offloading it to human fallback. This capability can be quantified from two directions: reduced human intervention hours (which connects back to the traditional time savings model), and the harder-to-measure but more important value of process continuity. Exceptions-related delays often trigger customer complaints, compliance risks, or downstream system backlogs — with hidden costs far exceeding the labor cost of handling a single exception manually.

Decision Quality

Agents don't just execute tasks — they make judgments within processes. Higher-quality, more consistent decisions reduce errors, decrease rework, and improve compliance. Improvements in decision quality often generate cascading downstream benefits — such as lower bad debt rates or more accurate customer responses — none of which can be captured by a time savings model.

Maintenance Economics

A major pain point with traditional RPA bots is fragility: when an underlying system's interface changes, scripts break, and maintenance costs soar. Because agents possess a degree of adaptability, they are theoretically more economical to maintain over time. Evaluating the full lifecycle value of automation requires factoring in long-term maintenance overhead — otherwise, short-term gains get systematically overstated while long-term burdens get underestimated.

Maintenance costs are among the most commonly underestimated hidden expenses during the business case phase of RPA projects. Research from Gartner and similar organizations indicates that maintenance work typically accounts for 30%–50% of total cost of ownership (TCO) for RPA, driven primarily by selector failures from upstream system UI changes, script redevelopment when business rules change, and version upgrade compatibility issues. This maintenance burden multiplies at scale — when an enterprise is running hundreds of RPA bots, the maintenance team can sometimes match the size of the development team. Agents' potential advantage here is that their comprehension capabilities allow them to adapt to interface or process changes to some degree without requiring line-by-line script modifications. That said, this advantage is still being validated, and agents introduce their own new maintenance demands — model version updates, prompt maintenance, and so on. Evaluating maintenance economics therefore requires a full lifecycle perspective, not just an assessment of initial deployment costs.

From Value Assessment to Prioritization

Building a complete value framework is only the first step. The more practical challenge is: when facing dozens or hundreds of candidate processes, which one gets automated first?

The framework's secondary function is helping organizations prioritize workflows. The processes most worth investing in first are typically those with the highest combined value across all four dimensions — significant time savings, frequent exception handling, sensitivity to decision quality, and manageable maintenance costs. Focusing narrowly on the processes with "the most hours" often means missing opportunities with far greater returns in exception handling and decision quality.

This shift in prioritization logic fundamentally requires organizations to develop a new decision-making vocabulary. CoE leaders need to explain to business and finance stakeholders why a process that appears to "save only a few hours" might actually deserve to be automated first. The answer lies in the value dimensions that legacy models ignore entirely.

Implications for AI Adoption

Agentic automation is ushering enterprise automation into a new era — and outdated evaluation methods risk becoming a hidden obstacle to adoption. If organizations continue applying RPA-era time savings logic to evaluate agentic projects, many high-value automation opportunities will be rejected simply because "the numbers don't add up."

For organizations actively driving AI adoption, building a multi-dimensional evaluation framework that encompasses time savings, exception handling, decision quality, and maintenance economics is not merely a financial refinement — it's a strategic capability. It determines whether an organization can identify genuinely high-value automation opportunities and make the right investment prioritization decisions under resource constraints.

The shift from "how many hours were saved" to "how much total value was created" may well be the defining inflection point that determines whether agentic automation can scale across the enterprise.

The AI Center of Excellence (CoE) is the critical organizational vehicle for driving this perspective shift. CoEs typically function as cross-functional internal teams responsible for setting automation standards, evaluating tool selection, building internal capabilities, and managing the automation portfolio. In the RPA era, the CoE's core work was maintaining a bot inventory and ensuring script stability. In the agentic era, the CoE's function needs to evolve into that of a "value portfolio manager" — one that not only understands the technology, but can also establish a common language with business and finance stakeholders, using a multi-dimensional value framework to inform investment decisions. Organizations without a CoE, or with an underpowered one, often fall into the trap of fragmented departmental pilots that fail to generate scaled value — and the absence of a robust evaluation framework is one of the root causes of that trap.