Skip to content

Measuring AI ROI honestly: baselines and full cost

How to calculate AI ROI: set a baseline first, count realised benefits rather than hours saved, and include full running cost. Includes a worked example.

6 min readBy AI Guru AcademyFacts last checked

In short

To calculate AI ROI, measure the current process before launch, count only the benefits that show up in budgets or revenue, and set them against the full cost of running the system. Hours saved are a leading indicator, not cash. A result you can defend in front of a finance team needs all three: a baseline, a realisation rate and a total cost.

Why do you need a baseline before you launch anything?

Without a baseline, any improvement claim is a guess. Measure the process as it runs today: volume per month, time per task, error or rework rate, and cost per completed item. Record how you measured it, so the same method can be repeated after launch.

Pick two or three measures that matter to the business, not a long list. If the pilot team also changes how it works for other reasons, note that too, because it muddies the comparison.

Which benefits count, and which are only possible?

Tangible benefits have a number that finance can verify; intangible benefits matter but need careful wording. Keep them in separate columns so a soft benefit never inflates a hard one.

Which benefits count, and which are only possible?
TypeExamplesHow to treat it
TangibleFewer contractor hours bought, lower rework cost, more orders handled by the same teamCount in the ROI figure once evidenced
IntangibleFaster onboarding, better staff morale, more consistent toneReport separately, track with a survey or proxy
Avoided costHiring you would otherwise have neededCount only if the hire was approved and would have happened

Is time saved the same as money saved?

No. Time saved becomes money only when the freed time is turned into something the budget can see, such as fewer paid hours, avoided hiring or extra billable work. Otherwise it is capacity, which is worth having but is not a saving.

Apply a realisation rate: the share of the theoretical saving you expect to actually capture. Set it from evidence, such as how similar earlier process changes performed, and review it after launch. A low rate is not a failure. It tells you where the next step lies, perhaps redeploying the capacity or changing the target headcount plan.

What does the full cost of an AI system include?

Total cost of ownership is more than the model bill. It covers the running cost of the AWS services plus the people and processes around them.

On the AWS side, the pricing structure depends on the service. Amazon Bedrock on-demand use is priced per token, with input and output tokens priced separately, and rates depend on the model and provider. Bedrock Provisioned Throughput is billed per model unit per hour, and the commitment options available vary by model. Amazon SageMaker AI real-time endpoints are charged for usage of the instance type you choose, while serverless inference is charged for the compute used to process requests and for the data processed. These structures behave differently as usage grows, so model each one against your expected volume.

Beyond the service charges, add:

  • Build and integration work, including the data preparation that retrieval or fine-tuning depends on
  • Evaluation, testing and the human review that keeps quality acceptable
  • Monitoring, support and ongoing change
  • Training and change management for the people using the tool
  • Governance and security effort

How do you estimate and then track the cost?

Use the AWS Pricing Calculator to model estimated costs for a new workload under different usage scenarios, and treat the output as an estimate that excludes taxes and reflects only the inputs you entered. After launch, use AWS Cost Explorer to view and analyse actual costs and usage over time, and filter by usage type. If you apply cost allocation tags, note that they must be activated before they appear in Cost Explorer.

The gap between estimate and actual is itself a useful measure. Review it monthly in the first quarter.

What leading indicators should you watch before the numbers arrive?

Financial benefits lag. Leading indicators tell you early whether the benefit is on its way.

  • Adoption: the share of eligible staff using the tool each week
  • Task coverage: the share of the target process that goes through it
  • Quality: acceptance rate of outputs without edits, and rework rate
  • Cycle time against the baseline
  • Cost per completed task, using actual usage

If adoption is low at week six, no realisation rate will rescue the business case.

Worked example

This is an invented illustration, not a real organisation or result.

A support team handles 4,000 tickets a month. The baseline shows 12 minutes of agent time per ticket, which is 800 hours a month. An AI drafting assistant is expected to cut handling time by a third on the 60 percent of tickets it can draft for.

Worked example
StepCalculationResult
Theoretical saving800 hours x 60% x one third160 hours a month
Realisation rateThe team expects to turn half of this into reduced overtime and avoided temporary cover80 hours a month
Benefit in money80 hours x the team's loaded hourly costBenefit B
Full monthly costService usage + support + review time + amortised buildCost C
ROI(B minus C) divided by CReport as a range

Report the figure with the assumptions beside it, and show what happens if the realisation rate is lower. The other 80 hours are capacity, and you can report them as such.

Key takeaways

  • Measure the current process before launch, or no improvement can be proven.
  • Separate tangible from intangible benefits and count only evidenced ones in the ROI figure.
  • Time saved is capacity until it changes a budget line; apply a realisation rate.
  • Include build, review, support, training and governance in cost, not only the service bill.
  • Use the AWS Pricing Calculator to estimate and AWS Cost Explorer to track actual spend.
  • Watch adoption, quality and cost per task as leading indicators.

FAQ

How do you calculate ROI for an AI project?

Subtract total cost from realised benefit and divide by total cost. Use benefits that finance can verify and costs that include people and process, not just the cloud bill.

What is a realisation rate?

It is the share of a theoretical saving that you expect to capture in practice. Hours saved only turn into money when the freed time reduces spend or adds revenue.

Can the AWS Pricing Calculator give me the ROI?

No. It estimates the cost side from the usage you enter, and the estimate excludes taxes. Benefits and the wider cost of ownership come from your own baseline and plan.

How soon should ROI be measured?

Track leading indicators from the first weeks and review financial results after usage has settled. Set the review dates before launch.

Sources

Last checked 2026-11-16.

AI Guru Academy is an independent training provider and is not affiliated with or endorsed by Amazon Web Services. AWS service names are trademarks of Amazon.com, Inc. or its affiliates.

Next step: take the free AIB-C01 diagnostic at /diagnostic, then see the course at /courses/aib-c01-exam-prep.

Measuring AI ROI honestly: baselines and full cost | AI Guru Academy