AI 自動化投資回報：建置前如何計算成本效益與回本期

Originally published on EpiFive, where it is kept up to date.

To calculate the ROI of an AI automation before you build it, estimate the yearly value of the time it saves, subtract the yearly cost of running it, and divide the one-off build cost by the monthly net benefit to get the payback period. The arithmetic is simple. The work is in getting honest inputs: measured task times, realistic exception rates, and every licence and usage charge. This guide gives a template, a worked example, and the pricing facts you need to fill it in.

Who this is for and what you will have

This is for team leads, operations managers, developers and small-business owners who must justify an automation or decide which of several ideas to build first. By the end you will have a one-page business case with inputs you can defend, a payback figure, a sensitivity check, and a go or no-go rule.

The example numbers in this article are illustrative inputs for the arithmetic. They aren't benchmarks. Replace every one with your own measurements.

Step 1: Measure the current process

Skip estimates and time the real task. For two weeks, or for a sample of 20 to 30 instances, record:

• Volume: how many times the task happens per month.
• Handling time: minutes per instance, from start to finish, including finding the information and any follow-up.
• Who does it: the roles involved, because their hourly cost differs.
• Exceptions: the share of instances that need judgment, rework or escalation.
• Error cost: what a mistake costs in rework or complaints, if you can quantify it.

Use a loaded hourly cost for each role (salary plus employer costs and overhead), taken from your finance team. A raw salary understates the cost.

Pick tasks with the right shape

Automation pays back fastest when the task is frequent, takes noticeable time, follows repeatable rules or can be handled by a model with checkable output, and has low cost when it goes wrong. A task that happens twice a month rarely repays a build unless each instance is long. A task with high-stakes errors needs human review, which cuts the saving.

Step 2: Estimate the value

The core formula:

Annual hours saved  = volume per month x 12 x minutes saved per task / 60
Annual value        = annual hours saved x loaded hourly cost

Minutes saved per task is not the old handling time. It is the old time minus what remains:

Minutes saved = old minutes - remaining minutes
Remaining minutes = review time + (exception rate x exception handling time)

For example, suppose a team handles 600 inbound requests a month, each takes 6 minutes to read, categorise and write a first summary, and the loaded cost of the role is 30 per hour (in your currency). After automation, a person reviews each result in 1 minute, and 10% of requests are exceptions that still take 6 minutes.

Remaining minutes = 1 + (0.10 x 6) = 1.6
Minutes saved     = 6 - 1.6 = 4.4
Annual hours      = 600 x 12 x 4.4 / 60 = 528
Annual value      = 528 x 30 = 15,840

Notice that reviewing and exceptions cut the saving from 6 minutes to 4.4. Many business cases fail because they assume the whole task disappears.

Non-time benefits

Faster response, fewer errors, and consistency can have real value, but include them only if you can measure them. If you list a benefit you can't measure, mark it as upside outside the payback calculation so it doesn't decide the case.

Step 3: Estimate the costs

There are two kinds of cost: one-off build costs and recurring running costs.

One-off costs

• Build time: hours to design, build, test and document, times the loaded rate of whoever builds it.
• Pilot and testing: time to run the process on real samples and measure accuracy.
• Integration work: connectors, permissions, approvals from security or compliance.
• Training and change management: time for the team to learn the new flow.

Recurring costs

A planning rule many teams use is to set aside a fixed share of the build effort per year for maintenance. That share is your judgment, not a standard, so state it as an assumption.

Step 4: Plug in real prices

Prices change, so take them from the vendor pages on the day you build the case. These are the figures published on the pages listed in the references at the time of writing.

Power Automate

The Power Automate pricing page lists, among others:

• Premium: $15.00 per user per month, paid yearly.
• Process: $150.00 per bot per month, paid yearly, for unattended desktop flows.
• Hosted Process: $215.00 per bot per month, paid yearly, with a Microsoft-managed virtual machine on Azure.

The page notes that prices are for marketing purposes and may differ by currency, country and region, and that the actual price is shown at checkout. Premium connectors need a suitable licence; see Power Automate licensing and premium connectors. AI Builder prompt actions also consume AI Builder credits or Copilot Credits, based on tokens and model, as covered in Power Automate Run a prompt action.

n8n

The n8n pricing page shows Cloud plans billed annually: Starter at 20 euros a month with 2.5K workflow executions and Pro at 50 euros a month with 10K executions, and it defines an execution as a single run of the entire workflow regardless of step count or data volume. The self-hosted community edition is available on GitHub, so its cost is the hosting and your own maintenance time. See n8n self-host with Docker.

Model API costs

LLM usage is billed per million tokens, with separate input and output rates. Anthropic's pricing page, for example, lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and Claude Haiku 5.5 at $0.10 and $0.50 per million for prompts up to 100,000 tokens. It also lists a Batch API discount of 50 percent on input and output tokens for asynchronous work. Other providers publish their own rates. For a method to estimate tokens, see Claude API pricing and token cost estimates.

Worked example: the 600 requests a month each send about 800 input tokens and receive about 150 output tokens.

Input tokens per month  = 600 x 800 = 480,000
Output tokens per month = 600 x 150 =  90,000
Using $2 and $10 per million tokens:
Input cost  = 0.48 x $2  = $0.96
Output cost = 0.09 x $10 = $0.90
Model cost  = about $1.86 per month

At this volume the model cost is negligible. The platform licence, build time, review and maintenance are what decide the case. At much higher volumes or long documents the token cost becomes material, which is why you calculate it rather than assume.

Step 5: Compute payback and ROI

Continue the example with illustrative costs: a build of 40 hours at a loaded rate of 50 per hour, plus 10 hours of pilot testing, so one-off cost is 2,500. Running costs: a Premium licence for one owner at $15 a month (180 a year), model cost about 22 a year, and 20 hours a year of maintenance and monitoring at 50 per hour (1,000).

One-off cost        = (40 + 10) x 50 = 2,500
Annual running cost = 180 + 22 + 1,000 = 1,202
Annual net benefit  = 15,840 - 1,202 = 14,638
Monthly net benefit = 14,638 / 12 = about 1,220
Payback             = 2,500 / 1,220 = about 2.0 months
First-year ROI      = (14,638 - 2,500) / 2,500 = about 4.9, or 490%

Those results come entirely from the input numbers. If the real saving were 1.5 minutes per task instead of 4.4, annual hours would be 180 and annual value 5,400. Net benefit would be 4,198 and payback about 7 months. That sensitivity is the point of the exercise.

Formulas to keep

Net annual benefit = annual value - annual running cost
Payback (months)   = one-off cost / (net annual benefit / 12)
First-year ROI     = (net annual benefit - one-off cost) / one-off cost

Step 6: Stress-test the case

Run the numbers again under pessimistic assumptions:

• Half the saving. Assume minutes saved is 50% of the estimate.
• Double the exceptions. Assume the exception rate doubles.
• Double the usage cost. If volume or tokens double, does the cost still hold?
• Maintenance doubles. Connectors and models change; assume more fixing.
• Adoption is partial. If only 70% of instances go through the automation, scale the value.

If the case still pays back within your threshold under the pessimistic version, it is robust. If it only works at the optimistic setting, run a pilot first.

Step 7: Pilot before you commit

A pilot replaces estimates with measurements. Run the automation, or a manual version of it, on 50 to 100 real items and record:

• Accuracy of the output against what a person would produce.
• Review time per item.
• Exception rate.
• Failures and their causes.

Because models can produce incorrect or fabricated output, plan human review for anything customer-facing or irreversible, and count that review time in the case. Microsoft's guidance on prompts recommends human oversight for this reason.

Decision rules

A simple scorecard helps compare ideas:

Common mistakes

• Using the full manual time as the saving instead of old time minus remaining time.
• Forgetting maintenance, monitoring and review time.
• Pricing licences for one person when several users or bots are needed.
• Treating a vendor's marketing price as your price; check the currency, region and billing terms.
• Counting savings that don't convert to anything, such as five minutes freed across many people without a plan for the time. Convert only time that you will actually redeploy or that avoids a hire or overtime.
• Skipping the pilot.

Checklist

• Measure volume, handling time, exception rate and who does the work.
• Calculate minutes saved as old time minus review and exception time.
• List one-off and recurring costs, including usage charges and maintenance.
• Take prices from official pages and record the date.
• Compute payback and first-year ROI.
• Stress-test with half the saving and double the costs.
• Pilot on real samples, then decide.

References

• Power Automate pricing (Microsoft)
• Licensing and AI Builder credits (Microsoft Learn)
• Human review for automation with a prompt (Microsoft Learn)
• n8n pricing
• Claude API pricing (Anthropic)

This guide was researched and drafted with AI assistance and checked against the vendor documentation listed in its References. Read the maintained version and related guides on EpiFive.