AI Agent ROI Framework

“The agent saves seven minutes” is an incomplete business case. Seven minutes on which tasks, for which users, after how much review, and compared with what alternative? A benefit becomes investable only when it survives those questions and has an owner who can turn the improvement into an operational result.

This framework is for sponsors deciding whether to fund, expand or stop an agent workflow. It connects a task-level measurement to a business outcome without treating every generated answer as a completed job or every released hour as cash. All numbers below are illustrative; they demonstrate a method and are not reported client results.

Write the benefit claim as a testable chain

For a supplier-onboarding team, the proposed chain might be: approved documents are available; staff use the agent on eligible applications; the agent prepares a correct evidence pack; staff spend less time assembling it; the team uses the released capacity to reduce a backlog without weakening review. Each step can fail independently.

A faster draft is not necessarily a faster onboarding decision. Missing documents may still take days to obtain. If the business objective is supplier activation time, measure both active staff effort and elapsed time between states. The agent may improve one without changing the other. If the objective is reducing overdue applications, record the queue size, arrival rate and completed reviews as well.

Define eligible work in advance. For this example, eligibility requires supported document types, an existing supplier record and a named reviewer. Unsupported applications remain in the manual process and stay visible in total demand. The scoping guide helps establish that boundary before the benefits model is built.

Keep three value ledgers

Different benefits require different evidence

Ledger	Possible benefit	What establishes it
Capacity	Staff hours available for other work	Measured net effort reduction and a plan for redeployment
Cash	Avoided paid overtime or a removed external charge	A corresponding change in expenditure, approved by finance
Quality and service	Fewer incomplete packs or shorter queues	Consistent outcome definitions and comparable observations

These ledgers may reinforce one another, but they must not be added blindly. If reduced overtime is the way released capacity becomes cash, count that benefit once. If faster onboarding is claimed to create revenue, show the link to actual commercial activity and use the appropriate incremental margin, not the face value of every supplier transaction. A plausible story is a hypothesis until measured.

A worked monthly calculation

Assume 1,200 eligible applications per month. Manual evidence preparation takes 12 minutes per application. With the agent, 70% produce an accepted pack that takes five minutes of staff handling, before a separate quality-review step. The accepted-task handling reduction is therefore seven minutes. All 1,200 applications require an additional 1.5-minute quality review. The remaining 30% require two extra minutes of exception triage beyond the manual baseline, and service oversight takes eight hours per month.

Illustrative net capacity released each month

Component	Calculation	Hours
Accepted packs	1,200 × 70%	840 packs
Handling time released	840 × 7 minutes ÷ 60	98
Additional quality review	1,200 × 1.5 minutes ÷ 60	−30
Exception triage	360 × 2 minutes ÷ 60	−12
Service oversight	Monthly allocation	−8
Net released capacity	98 − 30 − 12 − 8	48

The original manual preparation workload was 240 hours. This scenario releases 48 net hours, not the 98 hours implied by looking only at accepted packs. Its review and exception assumptions must be measured independently; if the five-minute handling figure already includes quality review, deducting review again would double count it.

At an illustrative loaded labour value of SAR 120 per hour, 48 hours represent SAR 5,760 of monthly capacity value. Subtract an assumed SAR 3,000 of monthly technical operating costs and the net economic contribution is SAR 2,760. With an initial build cost of SAR 66,000, simple undiscounted payback is approximately 23.9 steady-state months. It excludes ramp-up, financing, tax and future changes.

Under those same assumptions, first-year capacity value is SAR 69,120 and first-year build plus technical costs are SAR 102,000. Using ROI = (benefits − costs) ÷ costs gives approximately −32.2% for that year. This is a capacity-valued scenario, not a cash-return forecast. A different horizon or treatment of labour changes the interpretation, so label both.

The capacity-to-cash conversion can overturn the result

If salaries and paid hours do not change, released time does not automatically reduce expenditure. The business may still benefit by completing more useful work, but it should identify that work. Ask who receives the capacity, when it becomes available and whether it is sufficiently concentrated to be redeployed. Scattered seconds between unrelated tasks may be less usable than a complete review block.

Suppose only half of the illustrative capacity value can be converted into a measured financial benefit. That is SAR 2,880 per month. Against SAR 3,000 in technical costs, the monthly cash-like contribution is negative SAR 120 before the initial build. There is no positive payback in this simplified scenario. Reporting the 23.9-month figure without this conversion assumption would be misleading.

A sponsor can respond by improving adoption, reducing review effort, narrowing the supported workflow or choosing a simpler tool. The agent-versus-SaaS comparison provides a way to test whether an existing product achieves the same outcome with less ownership cost.

Find the assumption that deserves a pilot

For the worked example, let p be the proportion of eligible applications producing accepted packs. Monthly net hours equal 140p − 30 − 40(1 − p) − 8, or 180p − 78. To cover SAR 3,000 in technical cost at SAR 120 per hour, the model needs more than 25 net hours: p must exceed approximately 57.2%. This is a mathematical threshold for these assumptions, not a universal quality target.

Recovering the SAR 66,000 build over 24 months adds SAR 2,750 per month to that simplified requirement. The corresponding acceptance threshold is approximately 70.0%, before ramp-up or further change. The example therefore has little room for a weaker result if the intended horizon is two years. A pilot should concentrate on whether acceptance and review effort are credible, rather than merely showing that the agent can write a fluent pack.

Also test adoption separately from technical acceptance. If eligible volume exists but only 900 applications reach the workflow, using the same proportions yields 73.5 saved handling hours, 22.5 review hours, nine exception hours and eight oversight hours: 34 net hours. Fixed oversight makes the value fall faster than a simple proportional headline might suggest. Use the pilot-value diagnostic when the gap is in use rather than capability.

Measure the improvement against a credible alternative

The UK government’s Magenta Book explains why impact evaluation needs a credible view of what would have happened without the intervention. That methodological point is useful here; it is not a Saudi regulatory requirement. A before-and-after dashboard alone cannot separate the agent’s effect from staffing, seasonality, a backlog cleanup or a simultaneous policy change.

Where appropriate, compare contemporaneous groups of similar applications, assigning the workflow in a way that avoids selecting only easy cases. Keep the review standard and outcome observation window consistent. If random allocation is unsuitable, document the comparison design, known differences and limits on attribution. Avoid claiming a precise causal effect from a design that cannot support it.

Record the full task journey: eligibility, assignment, actual use, accepted result, review time, correction, reopening and final business outcome. Let someone other than the builder review a sample against the same rubric. Preserve failures in the dataset. Sample size should follow the variability and importance of the outcome; a convenient number of demonstrations does not establish general reliability. The evaluation guide covers behavioural evidence that complements this economic assessment.

Set expansion and stopping rules before the results arrive

Define which conditions must all hold: acceptable quality, no unresolved severe action failures, manageable review demand, meaningful use and a defensible economic contribution. A positive average time saving cannot compensate for a critical unauthorised action. Conversely, a safe but unused workflow may be an adoption or process problem rather than an engineering success worth scaling.

Give the benefit owner authority to act. For the onboarding example, the procurement operations lead might commit released review blocks to a named backlog and report completed applications. Finance validates the cash ledger; the service owner tracks costs using the TCO framework. Expansion should be based on the next workload’s evidence, since a second document family may have different acceptance and review rates.

For an Open for AI development discussion, prepare the current task volumes, a measured handling-time sample and the cost or service outcome you want to change. Bring that baseline to define a testable agent business case. The most useful first deliverable is an explicit value hypothesis and a way to disprove it.