Agentic AI ROI: The Process Intelligence Foundation

The number most boards receive for agentic AI return is an hours-saved estimate from the team that requested the budget, and it often does not reconcile to payroll or vendor invoices. Throughput may not support it either. McKinsey’s “State of AI 2025” found that more than 80% of organizations report no tangible enterprise-level EBIT impact from generative AI (McKinsey, 2025).

This guide gives a CFO a method for producing a dollar-denominated agentic AI ROI figure that survives audit committee review. It also explains why the per-seat and headcount models carried over from software and RPA produce figures a board should not accept.

The agentic AI ROI formula in one line

Agentic AI ROI = (risk-adjusted financial value delivered − total cost) ÷ total cost × 100, where total cost is one-time investment plus cumulative run costs. Measure it over a stated period against a pre-agent baseline captured from observed work. Every term has a definition, and the definitions decide whether the result is defensible.

The inputs on each side of the equation:

• Financial value delivered: Hours removed from human handling multiplied by loaded hourly cost, plus the change in cost per resolved interaction, plus cycle-time reductions converted to dollars (working capital released or penalties avoided), plus denials prevented, plus any revenue lift that can be traced to agent-handled cases.
• Risk adjustment: A discount applied to value delivered for agent accuracy and chain length, with rework rate also accounted for, so the numerator reflects what the agent completes correctly rather than what it attempts.
• Operational run costs: Tokens and inference, retrieval and vector storage, evaluation and observability, human oversight hours, rework on agent-handled cases, drift maintenance, and governance tooling.
• Total cost: One-time integration and evaluation harness costs, plus change management costs and cumulative run costs across the measurement period.
• Payback period: One-time investment ÷ (monthly value delivered − monthly run costs). The agent clears the hurdle rate when payback lands inside the period the finance committee set for the program.

Each term must trace to a ledger line or a metered invoice. An observed operating metric can also support it. An input that comes from a program owner’s estimate belongs in a separate column labeled unverified until an observed value replaces it.

Why RPA and generative AI ROI models mislead here

RPA models count bot-hours against human hours, and generative AI models count licensed seats against self-reported minutes saved. Neither measures whether a case finished without a person touching it, which is the outcome an agent is deployed to deliver. Both models were built for tools that assist a worker who remains in the loop, so their denominators assume the worker stays.

The headcount-reduction version of the RPA model is the one boards have the most reason to question. Operations claims a reduction, but finance finds the saved capacity reappeared as rework or manual fallback on exceptions.

A new oversight role can also consume that capacity, leaving the net position unclear. An agent that resolves a case end to end changes the structure of the claim. The saved hours can only become cash if the enterprise redeploys or releases the capacity, and the rework and oversight lines offset part of what remains.

Finance teams using the per-seat productivity model price access rather than outcomes. They score a user who drafts 30 ticket replies a day the same as one who drafts three, and users supply the minutes-saved input themselves.

For agentic AI ROI, finance needs the resolved case as the denominator, with autonomous resolution rate and cost per resolved interaction as supporting measures. Program owners who cannot report those measures struggle to defend the investment, a theme covered in why enterprise agentic AI initiatives stall.

How to calculate agentic AI ROI: Step by step

The calculation runs in four steps: baseline the workflow from observed data before any agent goes live, quantify hard and soft benefits separately, build the full cost side including recurring hidden lines, and risk-adjust the result for accuracy and chain length. Each step produces an input the board can test.

Step 1: Baseline the workflow before any agent goes live

Finance needs volume, handle time, loaded hourly cost, exception share, and rework rate for the workflow as it ran before the agent, and the window to capture them closes the day the agent goes live. Once agents handle part of the queue, the human baseline is contaminated, and any later comparison depends on reconstruction from memory or documentation.

The baseline gap finance usually faces is a measurement-source problem. Standard operating procedures describe the intended process, while interviews describe the remembered process. Event logs record the steps that instrumented systems happened to capture.

A claims adjuster opens the digital queue and verifies policy terms in a mainframe session, generating log events for those two steps. The adjuster then reconciles a spreadsheet against an email attachment without generating a log event. The logged handle time therefore understates the true cost per case.

Observed data closes that gap by measuring work at the point where the person performs it, across every application the case touches. The buyer vocabulary for this measurement layer is process intelligence. For a CFO, system records and self-reports differ from observation of the work itself as baseline sources. Record the source of each baseline input alongside the value.

Step 2: Quantify the benefit side

Hard benefits are the ones that appear in the P&L within the measurement period; soft benefits are capacity, quality, and risk effects that have value but no ledger line yet. The CFO signs off on the hard number and reports the soft number separately, with its conversion assumptions visible.

The hard benefit lines:

• Hours removed × loaded cost: Human handle time eliminated on agent-resolved cases, multiplied by the fully loaded hourly rate. This becomes cash only when the capacity converts to reduced contractor spend or overtime, or when attrition goes unbackfilled; until then, carry it as redeployed capacity.
• Containment and autonomous resolution rate: Share of cases the agent resolves with no human touch. Containment measures what the agent kept; autonomous resolution requires that the case closed correctly and did not reopen.
• Cost per resolved interaction: Total human and agent cost of the queue, including oversight, divided by resolved cases, compared with the pre-agent figure.
• Cycle time to dollars: Days removed from invoice approval or claim settlement, converted through working capital or late-payment penalties. Include prior authorization turnaround gains converted through denial write-offs.
• Cost avoidance: Penalties, rework, or compliance remediation that the agent’s procedure prevents, counted only when the pre-agent baseline shows the cost occurring.

Revenue lift belongs on the hard side only when a case-level link exists, for example a faster onboarding step that moves a measurable share of applicants to funded accounts. Satisfaction scores and employee time returned to higher-value work stay on the soft side. Audit readiness does too.

Step 3: Build the full cost side

Budgets usually capture the license fee and often under-model the lines that compound monthly. Build the cost side from the invoice up, with each line tied to the driver that makes it grow.

The cost lines to include:

• Platform or license fee: Subscription or per-case pricing, plus the renewal uplift written into the contract. Include consumption-unit pricing where applicable.
• Tokens and inference: Calls per case multiplied by tokens per call and the model’s price, including retries and reasoning steps. Include the context the agent re-reads on each turn.
• Compute and retrieval, including vector storage: Indexing and storage costs for the documents and records the agent consults. Include read units.
• Integration: Connectors and credentials for each system the agent reads from or writes to, including mainframe and legacy desktop access. Include maintenance for each system.
• Evaluation and observability: Test-case libraries and regression runs per release. Include the trace volume logged per case.
• Oversight: Reviewer hours per 1,000 agent-handled cases at the sampling rate the risk policy requires.
• Rework and prompt drift: Reopened cases handled by a person, plus engineering time spent when an upstream screen or policy changes. Include engineering time for model-version changes.
• Change management: Role redesign and training for every expansion wave. Include procedure documentation.

McKinsey’s June 2025 analysis notes that annual run costs for traditional IT typically represent 10-20% of initial build cost, while generative AI solutions at scale can incur recurring costs that exceed the initial build investment (McKinsey, 2025). A CFO who models agent run costs with the software convention will underestimate the denominator from month 1.

Two hidden lines deserve their own entries.

Agent sprawl: The recurring cost of agents that business units deploy outside the governed program. Each one incurs inference and oversight costs, with audit obligations that no budget owner booked.

Agency tax: The human time spent checking agent output and fixing what the agent got wrong on each agent-handled case. Budget it as a nonzero line from the first production month.

Step 4: Risk-adjust the number

Discount the benefit side for agent accuracy and chain length before the figure reaches a board deck, and account for rework rate, because the gross value assumes every attempted case completes correctly. The adjustment converts an attempted-resolution rate into a correct-resolution rate.

• Chain length: If each step in a 5-step procedure succeeds 90% of the time and the steps fail independently, the whole chain succeeds 0.9 to the fifth power, or about 59% of the time. The remaining 41% of cases need human recovery, which can cost more than handling them from the start. Laban et al. measured a related effect in multi-turn interactions: average performance fell to 65% in multi-turn settings against 90% in single-turn settings (ICLR, 2026).
• Verification trade-off: If a person spends v minutes checking an agent’s output and w minutes redoing a failed case, expected human time per case is v + (1 − p) × w, where p is the probability the agent succeeded. The agent clears the hurdle only when that expected time is below the original human handle time. A workflow with cheap verification therefore tolerates a lower p than one where every output needs a full review.

Set the risk discount per workflow from the observed p and the checking time v. Use the observed recovery time w as well, rather than a program-wide assumption.

Worked example: Service desk agents over 12 months

A service desk handling 4,000 tickets a month, with observed inputs replacing every estimate, reaches payback at the end of the first year and a 30% return at the end of the second. The figures below are illustrative assumptions for the method; replace each input with the observed value for your queue.

The 12-month walkthrough, with the baseline captured before the agent went live:

Service desk agent economics, 12-month illustration

• At month 12: Cumulative value is $201,600 and cumulative cost is $200,112, so the 12-month agentic AI ROI rounds to 1%.
• At month 24: Cumulative value is $403,200 against $310,224 of cost, a 30% return.

The same deployment scored with a headcount model would claim $201,600 of annual savings against license and one-time costs alone and report a 60% first-year return. That figure evaporates when the audit committee asks where rework and oversight were booked, along with inference.

Two inputs in the table decide the outcome. The 40% containment is an observed post-pilot rate. The 15-minute handle time includes the spreadsheet and ticketing-tool steps that the ticketing log alone did not record.

A desk that baselined from log data at nine minutes would have understated the gross value by 40% and reported a payback beyond 24 months.

A line-item TCO table for agentic systems

Total cost of ownership for an agentic system is a set of metered and recurring lines, most of which scale with case volume rather than with seats, and several of which trigger on every upstream change. The table assigns each line its driver and its billing frequency so finance can model growth instead of discovering it at renewal.

The cost components a 24-month model needs:

Agentic AI cost components over 24 months

The frequency column is the one that catches budget owners. Integration and change management look like one-time capital items, but each upstream screen change or policy revision reopens them as operating expense. Model deprecation does too.

What costs do I need to include beyond the license fee?

Beyond the license fee, book tokens and inference, compute and retrieval, integration maintenance, human oversight, and rework as separate operating lines, because each is metered or recurring and none appears on the platform invoice.

• Consumption-metered pricing: Some platforms bill agent actions in units, such as assists or credits; tokens are another billing unit, and at least one vendor’s units expire monthly. An agent that re-reads context on every turn consumes units far faster than the pilot team estimated in a chat-style test. Treat the pilot’s unit consumption as a floor and model overage at production volume before signing.
• Renewal escalation: Contracts often carry annual uplift clauses, and some vendors expand consumption pools at renewal “to accommodate growth” whether or not prior usage filled the pool. Finance will understate year two if it holds the license fee flat in a 24-month model.
• Oversight and rework: Operations pays for reviewer hours and reopened cases through the payroll that the agent was supposed to reduce. Finance can therefore report savings in a license-only model while finding none in the operating budget.
• Capital versus operating treatment: Confirm with your controller which one-time items, such as integration and the evaluation harness, qualify for capitalization. Inference, retrieval, oversight, and drift maintenance recur as operating expense for as long as the agent runs.

Which KPIs should I track to prove agent value?

Track a fixed set of metrics in three phases, with the pre-deployment set captured before go-live so the live and post-deployment sets have a comparator that no program owner can restate. Each metric below maps to a term in the formula.

Pre-deployment, from observed human handling:

• Case volume by type and channel
• Handle time per case, including cross-application steps
• Loaded cost per case
• Exception share and variant count
• Rework rate and reopen rate

Live, for the agent-handled share of the queue:

• Containment rate, the share of cases the agent kept
• Autonomous resolution rate, the share closed correctly with no reopen
• Escalation rate and the reason code for each escalation
• Cost per resolved interaction, human and agent costs including oversight
• Decision latency, the elapsed time from case arrival to the agent’s action
• Average handle time (AHT) for the human-handled remainder, which rises as easy cases leave the queue
• Tokens per resolved case

Post-deployment, for the finance committee:

• Net value realized against plan, by workflow and business unit
• Payback months, recalculated monthly
• Oversight hours per 1,000 agent-handled cases
• Drift rate, the share of cases where the agent’s path diverged from the approved procedure
• Write-backs to the system of record, the count of cases the agent completed inside the ledger, claims, or ticketing system rather than handing a draft to a person

The AHT line is the one that misleads when read alone. An agent that absorbs simple tickets leaves harder ones for people, so human AHT rises even as cost per resolved interaction falls, and a dashboard that reports only AHT will show the program failing while the economics improve.

Benchmarks and payback periods, and why the surveys disagree

Published agentic AI ROI figures range from 10% of organizations reporting significant returns to 88% reporting positive ROI, and the spread comes from who was asked and what counted as ROI. The measurement window also matters. Read each benchmark with those three questions answered before comparing it to your own number.

• Google Cloud’s “The ROI of AI 2025” reports that 88% of agentic AI early adopters see ROI from generative AI on at least one use case, against a 74% average (Google Cloud, 2025).
• Deloitte’s October 2025 report “AI ROI: The paradox of rising investment and elusive returns” found that 10% of agentic AI users currently realize significant, measurable ROI, across 1,854 senior executives in European and Middle Eastern markets (Deloitte, 2025). The population includes organizations at every stage, the bar is “significant and measurable,” and the report’s respondents expected multi-year horizons rather than first-year payback.
• PwC’s May 2025 AI Agent Survey of 308 US executives reports no numeric ROI figure at all; its outcome measures are shares of adopters reporting productivity and cost benefits, with speed measured separately (PwC, 2025). PwC’s figure on agentic budgets is a spending-intent measure and is sometimes misquoted as an ROI result.
• Forrester Total Economic Impact (TEI) studies produce the payback figures vendors quote. Forrester’s TEI of Microsoft’s agentic AI solutions models 120% ROI with a 15-month payback for a composite organization built from 8 interviewees across 6 organizations, with the study commissioned by the vendor (Forrester, 2026). TEI composites measure organizations that completed deployment and agreed to participate, so they describe the successful tail.

These figures represent success rates based on self-reported data, but that’s not necessarily reflective of agentic AI ROI as a whoe. On the other side sits Gartner’s June 2025 prediction that over 40% of agentic AI projects will be canceled by the end of 2027 “due to escalating costs, unclear business value or inadequate risk controls” (Gartner, 2025).

Build, buy, or partner: How the choice moves the ROI curve

In-house teams incur costs up front and reach breakeven later, while enterprises buying packaged platforms pay metered run expenses. Partners charge for services on top of either route.

Model all three on the same 24-month timeline before choosing.

Cost concentration by sourcing route over 24 months

In-house build: An in-house team must establish the evaluation harness before the first case runs, along with observability and governance tooling. None produces value on its own, so finance models the in-house route with the deepest trough.

Packaged platforms: Enterprises using packaged platforms reach returns sooner but pay the consumption and renewal costs from the previous section. Their year-two cost depends on contract terms they negotiate before go-live.

Partners: Enterprises using partners follow a similar cost curve, with services fees up front. Their long-run return depends on whether the partner transfers procedures and outcome data to them at the end of the engagement.

The deciding question for a CFO is portability: the share of outcome data and procedures the enterprise can export in a vendor-neutral format. Include audit records in that measure. If an enterprise leaves accumulated operational context inside a provider’s environment, its teams must re-baseline during a future migration, a cost none of the three curves shows.

Where returns concentrate

Returns concentrate in bounded, high-volume workflows with a clear completion criterion, dense rules, a small exception set, and a system of record the agent can write to. In those workflows the chain is short, verification is cheap, and the value lands in a ledger.

The workflows that fit the profile, described at the step where the agent acts:

• Service desk and ITSM: A password reset or access request arrives in the ticketing tool, the agent verifies entitlement against the directory, executes the change, and closes the ticket; the completion criterion is binary and the write-back is to the ticketing system.
• Claims: A low-value, first-notice claim enters the digital queue, the agent verifies coverage in the policy system, checks the loss against the deductible and exclusions, and either settles within authority or escalates to an adjuster with the reason coded.
• Accounts payable: An invoice arrives by email, the agent matches it to the purchase order and receipt, posts a clean match to the ERP, and routes a price or quantity exception to the buyer with the variance stated.
• Prior authorization: A request arrives from the provider portal, the agent checks the plan’s medical policy and the member’s eligibility, approves within policy, and routes clinical judgment cases to a nurse reviewer.
• Supply chain: A delivery exception arrives from the carrier feed, the agent reconciles it against the order and the demurrage clock, and triggers the standard remediation or escalates.
• Software development lifecycle: A dependency update or test failure arrives in the pipeline, the agent proposes the fix, and the merge waits for a human review, which the cost model must price per change.

Workflows that depend on tacit knowledge sit at the other end of the range. A complex liability claim or a Know Your Customer (KYC) escalation involves judgment that formal documentation does not capture, so an agent grounded only in documents faces a long chain with expensive verification. A practical starting point in each function is a workflow where the case closes inside a system of record; see how programs produce real results.

Governance and oversight as an ROI variable

Governance is a recurring run cost with its own lines (AgentOps staffing, a kill switch, human-in-the-loop review, and continuous controls monitoring). A model that treats it as a one-time compliance project understates the denominator every month the agent runs. Put each line in the TCO table and the ROI figure stays honest as the agent estate grows.

AgentOps is the operating function that deploys, evaluates, and retires agents, and its staffing scales with the number of agents in production rather than with case volume. A kill switch is a documented procedure for halting an agent mid-run and returning the queue to human handling; it needs testing, and the test is a cost.

Article 14(4)(e) of the EU AI Act requires that those overseeing a high-risk system can “interrupt the system through a ‘stop’ button or a similar procedure that allows the system to come to a halt in a safe state,” with the Annex III high-risk obligations now applying from 2 December 2027 (EUR-Lex, 2026). A bank or insurer with EU exposure budgets that control whether or not the deployment is in scope today.

Human-in-the-loop review is the oversight line from Step 3, and its rate is a policy decision: the sampling share, the escalation thresholds, and the authority limits an agent may act within. Continuous controls monitoring detects deviations from the approved procedure on every case rather than on a sample, which can lower the sampling burden on reviewers but adds its own tooling and log-retention cost.

Governed autonomy is the operating state where an agent acts within defined authority and policy boundaries, with explicit escalation limits, and agentic operating procedures (AOPs) are the artifact that defines them. An AOP encodes the workflow steps, decision rules, exception handling, escalation thresholds, and success criteria the agent must follow, which gives the oversight function something concrete to audit against. Without an AOP, a reviewer is judging each case from scratch, and the oversight line never declines.

How observed context strengthens the ROI case

Finance can only defend an agentic AI ROI figure when the baseline and the live measurements come from a source independent of the team requesting the budget, and that source has to see the work across every application the case touches. Documents describe intent, while event logs record what instrumented systems captured. Self-reported metrics describe what program owners believe; none of these sources gives an audit committee a baseline it can test.

Skan AI builds a Context Graph of Work from desktop-level observation of process activity, so the baseline in Step 1 comes from how work ran rather than from how it was documented. The Skan Virtual Assistant (Skan VA) observes process activity across mainframes, VDI, legacy desktop applications, web applications, spreadsheets, and office tools without backend integration with every application. That captures the cross-application handoffs and manual workarounds that event logs miss. The unit of analysis is the work and the process, never the individual worker.

The Context Graph of Work is a continuously updated model that connects people, applications, tasks, decisions, policies, and outcomes as they are observed. The closest plain-language description is a digital twin of operations: a current, queryable picture of the workflow that updates as the work changes rather than a point-in-time assessment that is stale by the time the board deck is built.

• Skan AI Blueprint: Finance teams use that observed picture as the pre-agent baseline and live scoreboard. Blueprint maps operational activity and produces a prioritized roadmap tied to measurable outcomes. Teams receive initial operational insight in weeks, not months.
• Skan AI Intelligence: Documented on the Skan AI process intelligence platform page, Intelligence converts desktop-level observations into end-to-end process flows, variant analysis, execution metrics, bottleneck detection, and drift detection. Finance teams capture cost per case and handle time before go-live, along with rework rate. They can then detect drift on the live agent-handled share when its path diverges from the approved procedure.

Skan AI Agents uses the Context Graph of Work and AOPs to define what an agent can do and when it must stop. They also define when a human must decide, and the observation-to-agent pipeline includes evaluation against human baselines. That evaluation supplies observed inputs for the risk adjustment in Step 4, so the discount can reflect the workflow rather than a program-wide assumption.

The 4-phase roadmap describes the path from observation to governed execution.

Regulated enterprises can keep raw screen data inside their own perimeter. Skan PRISM is an on-premises privacy and information security gateway that applies pixel-level redaction before protected information can leave the customer environment. Email content and keystrokes are never captured. Customer-hosted deployment supports data residency requirements in banking and insurance; it also supports healthcare requirements. Privacy by design is the operating principle for the whole observation layer.

Re-baselining is where the continuous view earns its place in the formula. HFS Research’s 2026 analysis of agentic productivity KPIs found reported efficiency gains averaged 60% for single-agent deployments and 52% for large multi-agent ones. It recommends re-baselining at each maturity threshold because, without the single-agent baseline, the decline “looks like decay rather than progress” (HFS Research, 2026). A baseline that updates as agents expand lets finance separate a maturing program from a deteriorating one.