Saving five minutes on a task is valuable only if the result is usable and the saved time can be put to work. This guide shows how to account for review and correction time so an automation estimate does not count every model response as a saving.
Time saved is not automatically cash saved. It becomes economic value when the business reduces paid effort, absorbs more work without equivalent cost, improves a measurable outcome, or redeploys capacity to work with a credible value. The worksheet below keeps those claims separate so the business case can be challenged rather than merely admired.
Related service: AI & Automation services- 01Choose one workflow, evaluation period, owner, and unit of completed work.
- 02Measure current volume, active handling, queue time, quality, exceptions, and operating cost.
- 03Define which benefit categories are measurable and prevent overlap between them.
- 04Estimate one-time and recurring build and running costs with uncertainty scenarios.
- 05Calculate gross benefit, net benefit, ROI, payback, and cost per correct outcome.
- 06Test sensitivity to the most uncertain inputs and set pilot decision thresholds.
- 07Replace assumptions with actual usage, review, quality, adoption, and cost after launch.
SECTION 01
Use five formulas and define every input
Choose a consistent period such as a month or year and one completed-work unit. Use loaded labor rates only when the organization accepts that measure. Keep currency, tax, depreciation, and accounting treatment aligned with internal finance policy; this worksheet is an operating decision tool, not accounting advice.
| Measure | Formula | Interpretation |
|---|---|---|
| Gross capacity value | Volume × (baseline minutes − future human minutes) ÷ 60 × loaded rate | Value of released paid capacity; not automatically cash savings |
| Risk-adjusted benefit | Sum of distinct benefit × confidence factor | Expected value of benefits after uncertainty adjustment |
| Net benefit | Risk-adjusted benefits − one-time costs − recurring costs | Value remaining in the evaluation period |
| ROI | Net benefit ÷ total build and running costs × 100 | Return relative to the cost invested |
| Payback period | One-time cost ÷ positive monthly recurring net benefit | Months required to recover initial cost if recurring net benefit is positive |
| Cost per correct outcome | Recurring technical + human cost ÷ correctly completed units | Operating unit economics adjusted for quality |
SECTION 02
Measure the current process before estimating improvement
Sample enough representative work to capture ordinary cases and exceptions. Record incoming volume, correctly completed volume, active handling minutes, queue time, rework, escalation, error type, error consequence, and tools or subscriptions. Segment the baseline when different case types have materially different effort or risk.
Use system timestamps and time sampling where possible, then validate with the people doing the work. Avoid asking for a single remembered average. Document seasonal or growth assumptions. If the process changes during the study, keep the versions separate so the automation is not credited for unrelated improvement.
SECTION 03
Choose distinct benefit categories and prevent double-counting
Capacity value is the most common benefit but also the easiest to overstate. If ten minutes are released, name what happens to those minutes: reduced overtime, avoided hiring, more completed cases, faster response, or reassigned work. Do not count both full labor savings and the full value of output produced by the same hours unless the causal paths are genuinely separate.
Quality benefit may include avoided correction labor, fewer duplicate transactions, fewer missed follow-ups, or another observed consequence. Cycle-time benefit needs a link to an outcome such as conversion, service level, working capital, or customer retention. Revenue contribution should use contribution margin and an evidence-based attribution factor rather than total revenue touched by the workflow.
| Benefit | Calculation input | Evidence rule |
|---|---|---|
| Capacity released | Human minutes removed from correctly completed work | Name the redeployment, avoided capacity, or paid-cost effect |
| Quality improvement | Avoided errors × average incremental consequence | Use observed error categories and avoid speculative reputational values |
| Faster cycle | Time change × measured business effect | Demonstrate that speed changes the outcome, not only the timestamp |
| Additional throughput | Incremental correct units × contribution per unit | Prove demand and that capacity was the constraint |
| Revenue contribution | Incremental conversions × contribution margin × attribution | Use a control, comparison, or conservative attribution method |
| Risk avoided | Event exposure × probability reduction | Keep separate and scenario-based when evidence is weak |
SECTION 04
Include the complete build and running costs
One-time costs can include discovery, process and UX design, data preparation, application and AI engineering, integrations, security, evaluation, migration, deployment, training, and internal subject-matter time. Record a contingency or uncertainty scenario rather than hiding unknowns inside a precise estimate.
Recurring costs can include model and data processing, cloud infrastructure, databases, monitoring, subscriptions, human review, exceptions, support, content or source updates, security operations, regression evaluation, and improvement work. Include failure cost when unsuccessful runs create repair, duplicate actions, customer contact, or downstream reconciliation.
SECTION 05
Copy this worksheet into a spreadsheet
Create one row per case type if volume, handling, quality, or benefit differs. Keep every assumption, source, owner, and confidence factor beside its value. Separate pilot, expected production, and expansion scope. This structure makes it possible to replace estimates with actuals without rebuilding the business case.
| Input | Low | Expected | High | Source and owner |
|---|---|---|---|---|
| Monthly task volume | System report / process owner | |||
| Baseline human minutes per item | Time sample / operations | |||
| Future review and exception minutes per item | Pilot log / operations | |||
| Loaded labor rate | Finance or HR policy | |||
| Correct-completion rate | Evaluation and system reconciliation | |||
| Distinct quality or throughput benefit | Operational evidence / benefit owner | |||
| Benefit confidence factor | Decision team | |||
| One-time lifecycle cost | Approved estimate | |||
| Monthly technical operating cost | Architecture and billing | |||
| Monthly human review and support cost | Pilot log and staffing model | |||
| Evaluation period in months | Finance and project owner |
SECTION 06
Run scenarios and sensitivity before approving the business case
Low, expected, and high scenarios should vary the inputs that genuinely remain uncertain: adoption, correct-completion rate, human review, exception rate, volume, model usage, integration maintenance, and attributable benefit. Do not make every input optimistic at once. Create a downside case where quality is lower, review is higher, and benefit adoption is slower.
Then change one input at a time to see which assumptions control the result. Validate those first in the pilot. If the business case depends mainly on an unproven conversion increase or near-zero review, use that dependency as a decision risk rather than presenting the expected ROI as a fact.
- 01
Calculate the current process cost and quality by case type.
- 02
Calculate the low scenario using conservative benefit and higher review or operating cost.
- 03
Calculate the expected scenario using the best supported assumptions.
- 04
Calculate the high scenario without exceeding operationally plausible capacity or demand.
- 05
Change one assumption at a time and rank its impact on net benefit.
- 06
Turn the most influential unknowns into pilot measurements and stop criteria.
SECTION 07
Use pilot actuals for an expand, redesign, or stop decision
Measure eligible volume, attempted and correctly completed tasks, human review and correction time, exceptions, latency, model and infrastructure usage, failures, support, and adoption. Reconcile the agent's completion log with the system of record so abandoned or duplicate work is not counted as success.
Compare actuals with the scenario assumptions and explain material differences. Expansion should have its own cost and benefit case because another tool, data source, team, or authority level changes the system. The U.S. Small Business Administration's cost-benefit framing is intentionally simple: compare benefits and costs over a defined period. Keep the model understandable enough that decision owners will continue updating it.
SECTION 08
Frequently asked questions
How do you calculate AI automation ROI?
Subtract one-time and recurring build and running costs from risk-adjusted measurable benefits, divide the net benefit by total build and running costs, and multiply by 100. Use a defined period and document every input, owner, and assumption.
Should time saved be counted as money saved?
Not automatically. Report time released, then identify whether it reduces paid effort, avoids additional hiring or overtime, supports more contribution-producing work, or improves another measured outcome. Otherwise it is capacity, not cash.
What costs belong in an AI ROI calculation?
Include discovery, internal owner time, data, build, integrations, security, evaluation, deployment, training, model and cloud usage, subscriptions, review, exceptions, monitoring, support, maintenance, regression testing, and failure repair.
What is a good ROI for AI automation?
There is no universal threshold. The business should set its own hurdle rate, risk tolerance, strategic value, and payback requirements. Evaluate the downside scenario and non-financial safety or compliance constraints alongside the percentage.
PRIMARY REFERENCES
Sources and further reading
These references cover the standards, platforms, or published prices discussed in the guide. Worked examples and checklists are our editorial guidance.
- Cost-benefit analysis guidanceU.S. Small Business Administration
- Cost Estimating and Assessment GuideU.S. Government Accountability Office
- AI Risk Management FrameworkNational Institute of Standards and Technology
- Working with evalsOpenAI
- Cost optimizationOpenAI
EDITORIAL METHOD
About this guide
We use AI to assist with drafting and editing. Catapult AI Work is responsible for the published content. Examples illustrate possible approaches; they are not client case studies unless identified as such.
Budget examples are not Catapult package prices. Check linked provider pages for current fees and plan limits before making a purchase.
Read the editorial policy