Skip to the cards
Filed
Loans
460

AI ROI Measurement: Track Full Costs, Agent-Level Value and Payback

AI returns are hard to defend without a complete cost baseline. Track usage, oversight and outcomes by initiative and agent to give finance a clearer payback view.

AI ROI Measurement: Track Full Costs, Agent-Level Value and Payback

AI investment is spreading faster than many organizations can prove what it returns. For IT and finance teams, the central task is to connect each initiative’s measurable business value to its full cost, including usage, oversight and the work needed to keep it running. A license price alone won’t do that job.

/articles/91/iW4f2Ve4zMZ1IqMLASwUSbCs5SzFHOyJ.webp

The gap is visible in reported results. McKinsey’s State of AI in 2025 found that 88% of organizations use AI regularly in at least one business function, while 39% say AI has affected enterprise-level earnings before interest and taxes (EBIT). BCG’s IT Spending Pulse reported an average realized three-year generative AI return of 11.2% in 2025, compared with an expected return of 14.1%. The realized figure had risen only slightly from 10.6% the year before.

Those figures point to a measurement problem as much as a technology problem. Adoption can be widespread while financial impact remains hard to attribute. If finance cannot verify the baseline, the costs or the outcome, an ROI figure may be a useful estimate for discussion, but it isn’t a number the organization can confidently defend.

Why AI ROI is harder to establish

Return on investment compares the value produced by an initiative with the cost of building, operating and governing it over a defined period. The basic calculation is familiar. The difficulty is that AI costs often change with usage, while the benefits can be hard to isolate from other work.

Traditional software costs are often easier to forecast: a fixed license fee can be compared with an expected productivity gain. AI adds variable expenses such as compute and token consumption. An agent’s usage may rise with demand, the model it uses or the number of other agents it calls. Meanwhile, benefits such as quicker decisions, faster service resolution or new revenue may be apparent to staff before anyone has measured them in financial terms.

Consider an organization piloting a customer support agent while rolling out Microsoft 365 Copilot. Copilot’s per-user license gives finance a relatively clear starting figure. The support agent’s usage cost can shift with ticket volume, model choice and its retrieval or handoff activity. Without tracking those factors, the team may still be estimating the agent’s cost and return months after launch.

That makes the cost baseline essential. It also explains why the CFO’s role is expanding: finance is being asked to assess AI spending against the same requirements used for other investments, including a defined payback period and a value model it can support.

Finance is taking a closer role in AI decisions

AI budgets are rising, but pressure to account for their results is rising too. In Gartner’s 2026 CIO and Technology Executive Survey, about 52% of CIOs said cost reduction would become more important over the next two years. BCG’s research on cost control found that cost management had become one of CIOs’ top three priorities, alongside growth, for the first time since 2023.

That combination changes the approval conversation. Finance needs to know what a proposed deployment costs, what outcome it is expected to produce and when the organization will reassess whether it is paying back. A budget allocation by itself answers none of those questions.

There is also a practical reason to involve finance early: a measurement model is easier to build before a pilot starts. Once employees are using a tool and costs are spread across cloud bills, subscriptions and agent consumption, reconstructing the baseline can be difficult. Agreeing on the costs and outcomes to track at the start gives both IT and finance a reference point.

Three common gaps undermine measurement

Spend is scattered across services and teams

AI costs can sit in cloud usage, software subscriptions, custom-agent consumption and expenses that individual teams manage. A central team may be able to retrieve a Microsoft 365 or Google Workspace invoice quickly, yet still lack a complete view of token usage for a custom agent, compute charges from a departmental pilot or subscriptions purchased outside the main IT process.

Each item may appear reasonable on its own. Together, the fragments make it difficult to establish one verified baseline. An organization comparing a pilot’s benefits with only its visible license fee may omit costs that belong in the calculation.

A license price is only one part of total cost

Total cost of ownership (TCO) should include the full set of expenses an AI initiative causes. Compute, storage growth, oversight and the labor required to monitor or retrain models can all affect the result. If those costs are left out, the apparent return improves while the accounting becomes less complete.

Rachel Simon headshot

Some of these expenses are easy to miss because they sit inside existing budgets. Compute may be grouped with general cloud spending; storage growth may be assumed to be negligible if the original user data volume appears unchanged. Yet AI-generated content, logs, embeddings and duplicated data can add storage demand. Likewise, staff time spent reviewing agent access, enforcing policies or checking performance may never have been scoped as a project cost.

These costs don’t need to be attributed with artificial precision. They do need a consistent method. If the same categories are included for each initiative, finance can compare like with like and see where assumptions need further investigation.

Without ownership and activity records, value is hard to attribute

For a result to count toward an agent’s return, the organization needs a way to identify the agent, its owner, its permissions and the work it performed. Without that inventory, teams may be unable to tell which agent handled a workflow or what data it accessed. That weakens both cost tracking and attribution of outcomes.

The potential consequences include paying for agents that no longer serve a useful workflow and overlooking the expense of those that remain active. Gartner projects that more than 40% of agentic AI initiatives will be canceled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls as leading reasons. These problems are harder to address when spend and outcomes can’t be tied back to specific deployments.

Build an ROI method before the pilot begins

A useful process starts with the cost baseline, then defines the value measures and payback expectations for the initiative. Keeping those decisions consistent makes it easier to compare results later.

  1. Consolidate the costs. Bring license, compute, token and agent expenses into one view across the platforms in use. Record the period covered and the initiative or workflow each cost supports.
  2. Include the full TCO. Account for storage, monitoring, retraining, auditing, policy oversight and staff time spent reviewing access. Where a cost is shared, use a documented allocation method rather than silently excluding it.
  3. Define outcomes before deployment. Decide which business measures the initiative is intended to change. Depending on the use case, these could include cost per resolved outcome, payback period, attributable revenue, cost avoided, productivity gains or variance from the baseline.
  4. Keep hard and soft value distinct. Cost avoided, revenue directly attributable to the initiative and headcount reallocated can be reported separately from less direct measures such as faster decisions or improved employee sentiment. Combining both into a single figure can make the result appear firmer than the evidence allows.
  5. Set a payback expectation that fits the initiative. A small departmental pilot and a platform-wide deployment may have different timelines. Set the standard with finance using the organization’s own cost of capital and investment targets; a generic industry figure may not fit either case.

A complete TCO view should cover the following categories:

Cost categoryCommon omissionWhat to account for
ComputeGrouped with ordinary cloud infrastructureUsage-based inference and training costs linked to the initiative
StorageAssumed to stay flat because source data volume looks unchangedAI-generated content, logs, embeddings and duplicated data
Governance and oversightAbsent because no project cost line was createdStaff time for monitoring, retraining, auditing and policy enforcement
Agent licensing and token usageReduced to the platform’s seat licensePer-agent consumption, token use and any relevant license or credit pool

Compare costs across Microsoft 365, Google Workspace and SaaS

AI pricing varies across platforms, so an organization needs a way to compare costs that don’t appear on the same kind of invoice. Microsoft 365 Copilot combines per-user licensing with other services and agent usage that may have separate or usage-based costs. Google’s Gemini offerings can also combine seat and usage charges. Salesforce Agentforce relies primarily on consumption credits.

The seat price is often the easiest part to identify, which can make it tempting to stop there. But custom agents built in Copilot Studio, agent work using Google Cloud services or sales workflows built with Agentforce may add costs below that platform-level price. Those expenses should be measured against the workflow they support.

Platform-by-platform accounting can miss overlap. For example, separate teams might build agents for related support, data and sales workflows across different services. Each team could report a defensible local result, while no one can reconcile whether the organization is paying for overlapping work or whether the benefits add up at the enterprise level.

A common framework should therefore normalize the different cost models and connect each expense to an initiative or workflow. When pricing structures change, the framework also needs to reflect those changes in the period being measured. That gives finance a clearer view than adding together seat licenses while leaving consumption charges out of the comparison.

Measure agents individually as their numbers grow

Agent-level tracking helps distinguish deployments that are returning value from ones that consume resources without a clear result. A blended figure for a platform can hide wide variation: some agents may support valuable work, while others may have little activity or no current owner.

Gartner projects that the average Fortune 500 organization will run more than 150,000 AI agents by 2028, up from fewer than 15 in 2025. At that scale, a single portfolio-level ROI figure is unlikely to explain which deployments are paying back. A named owner and an individual cost trail give teams a basis for reviewing usage, outcomes and ongoing need.

A practical maturity model has three stages:

StageMeasurement stateExample
UnmeasuredNo reliable cost baseline or outcome attribution; spend is visible only at platform level, if at allA broad Copilot or Gemini rollout that began before a cost model was set
TrackedCosts are recorded for an initiative, but value remains qualitative or estimatedA departmental pilot with a named budget owner but no attributed outcome measure
ProvenCosts and outcomes are measured at the agent or workflow level, with a finance-supported payback figureA deployment with a named owner, full cost trail and business outcome measure

Moving from one stage to the next takes more than adding a dashboard. The records must link usage and costs to an owner, a workflow and an outcome measure. That makes it possible to reassess a deployment and decide whether its current cost still makes sense.

Give finance and leadership one reviewable picture

Reporting cost and value together gives decision makers a direct view of whether an initiative is producing its expected result. If spending appears in one report and value in another, reviewers have to assemble the comparison themselves, often from figures gathered on different schedules.

A useful review should separate a pilot from a mature platform deployment. The pilot may still be testing whether a workflow can produce a measurable benefit; a broader deployment may have a different payback timeline and evidence base. Combining them can make early work look worse than it is, or hide weak results behind mature initiatives that are already returning value.

Review ROI on a consistent cadence rather than waiting for a subscription renewal. A quarterly initiative-level review, paired with more frequent cost monitoring for deployments whose usage changes quickly, can reveal a widening gap between spend and value sooner. The right cadence depends on the initiative and how quickly its costs can change.

Bring governance and ownership records into the same review. A cost total is easier to assess when the organization can show who owns the agent, what access it has and how the measured outcome was attributed. Assign an owner to every figure presented for review, so someone can explain its assumptions and evidence.

Questions to settle in your own environment

How does ROI differ from adoption? Adoption measures how many people or workflows use AI. ROI measures the value returned relative to cost. High usage can coexist with weak or unmeasured financial returns.

Is there a standard enterprise AI ROI benchmark? There isn’t one figure that fits every initiative. BCG reported an average realized three-year generative AI return of 11.2% against an expected 14.1%, and said high-maturity adopters reported more than double the return of early-stage adopters. Those results provide context, but an organization should also consider its own maturity, costs and finance targets.

How does agent sprawl affect ROI? Agents can continue to consume tokens, compute or license capacity even after a workflow stops using them. Without ownership and activity records, their costs may be hard to identify and their weak results can disappear inside a portfolio average.

What is the difference between AI TCO and AI ROI? TCO is the complete cost of an initiative, including items such as compute, storage, oversight and agent consumption. ROI compares that full cost with quantified value. Leaving costs out of TCO makes the resulting return look stronger than the evidence supports.

What evidence supports an ROI figure? Teams should be able to connect the cost baseline to a defined outcome measure and show which agent or workflow produced it. Ownership, access and activity records help explain how the figure was built and give finance a basis to challenge its assumptions.

The most useful measure depends on the work being evaluated. Cost per resolved outcome, payback period, attributable revenue, cost avoided, productivity gains and cost variance against baseline can all help when they match the use case and have a clear method behind them.

The next step is straightforward: establish a verified cost baseline, agree on the outcome measures with finance, and assign an owner before the initiative scales. That gives IT a practical way to spot rising costs, weak attribution and deployments whose value no longer justifies their spend.

Filed alongside