In this article
  1. Choose the decision and the accounting view
  2. List the costs that sit outside the subscription
  3. A worked example with deliberately fictional numbers
  4. Now look at the cash, which tells a different story
  5. Check whether the assumptions survive an ordinary month
  6. Give released capacity an owner
  7. Buy evidence before buying scale
  8. Questions people ask

An AI demo replaces a twenty-minute task with a five-minute task. Someone multiplies fifteen minutes by every employee, adds an hourly rate and announces a large saving. Finance asks when the saving will appear in the bank account. The room develops a sudden interest in the next slide.

Time can be valuable without being a cash saving. An automation can free capacity, improve quality, reduce paid overtime or enable additional sales. Each benefit has a different route to value. A credible AI automation ROI calculation names that route and includes the costs needed to make it happen.

Choose the decision and the accounting view

Are you deciding whether to start a pilot, deploy a tested workflow or expand an existing system? Each decision needs the future costs and benefits relevant to that choice. Money already spent can inform lessons and reporting, but it should not be treated as a reason to keep spending.

State the period, currency and perspective. A twelve-month cash case answers what happens to cash over a year. A resource case values employee time and other capacity, even when salaries do not change. You can show both, clearly labelled, without adding them together.

HM Treasury's Green Book describes appraisal of costs, benefits and risks across options. It is public-sector guidance, not a private-company ROI calculator. I borrow the useful discipline of comparing alternatives and making assumptions visible, including the option of changing the process without AI.

List the costs that sit outside the subscription

Include setup, integration, data preparation, training, evaluation and the time employees spend practising. For ongoing operation, include licences, model usage, infrastructure, monitoring, review, corrections and support. Ask who updates the workflow when source material, business rules or a provider changes.

Internal time still consumes a resource even when it does not generate an extra invoice. Put it in the resource view at an agreed valuation. Keep paid contractor invoices and internal time separate so the same integration work does not appear twice.

Plan for exceptions. If the automation handles ordinary cases but sends difficult ones to a person, measure the difficult work too. If a wrong result reaches a customer, someone must investigate and repair it. Describe serious risks separately when a speculative expected-loss number would create false precision.

A worked example with deliberately fictional numbers

Imagine a fictional operations team processing 1,000 routine requests each month. Its baseline is twelve active minutes per request, or 200 hours. With automation, people spend three minutes reviewing each request, totalling 50 hours. Assume 100 requests need another eight minutes of correction, totalling 13.33 hours, plus eight monthly hours of administration and monitoring.

The full new workload is therefore 71.33 hours. Released capacity is 200 minus 71.33, or 128.67 hours per month. At a fictional loaded labour valuation of €30 per hour, that capacity has an estimated resource value of €3,860 per month, rounded. This already deducts review, corrections and administration. We must not subtract those hours again as separate ongoing labour costs.

Assume setup requires a €10,000 external integration invoice and 100 internal hours valued at €30, or €3,000. Ongoing non-labour costs are €400 monthly for licences, usage and infrastructure. The first-year resource cost is €13,000 plus €4,800, or €17,800.

Twelve months of released capacity is valued at €46,320. The estimated net resource benefit is €28,520. Using this defined view, illustrative ROI is €28,520 divided by €17,800, approximately 160%. This is an assumption-based example, not a forecast or a claim about my client work.

Now look at the cash, which tells a different story

Suppose salaries remain unchanged and the company sells no extra work. Cash outflow is €10,000 for integration plus €4,800 in ongoing charges. There is no direct cash benefit in this scenario. The resource case may be positive while the first-year cash position is negative by €14,800.

If the company can genuinely eliminate €2,000 of paid overtime each month, a separate cash scenario has a €24,000 benefit and €14,800 cost. Net cash benefit is €9,200 and the illustrative cash ROI is approximately 62%. Verify that the same service quality and workload can be maintained before accepting that assumption.

Do not add the full €46,320 capacity valuation to the €24,000 overtime saving. The overtime reduction uses some of the released capacity. Count that portion once. Similarly, if capacity enables extra sales, use attributable contribution after incremental costs rather than booking both the full time valuation and the resulting revenue.

Check whether the assumptions survive an ordinary month

Our resource estimate assumed the workflow runs at the stated volume for twelve months. Deployment delays, lower adoption and changing demand will alter the result. If only half the requests use automation and setup costs stay the same, the investment looks different. Show that scenario before somebody else discovers it.

Vary volume, review time, exception rates and ongoing costs. Identify the assumption with the biggest influence on the decision and test it first. In this example, a review that takes six minutes rather than three removes 50 hours of monthly capacity. A small change in the licence price would matter much less.

For larger or longer investments, have finance decide whether to use discounting and the company's normal appraisal method. Do not substitute a quick spreadsheet for the standards your business already relies on.

Give released capacity an owner

“Employees will do more valuable work” is a possibility, not a plan. Specify the backlog, service improvement or avoided overtime that will absorb the capacity. Agree who reallocates the work, whether demand exists and how the benefit will be measured.

A practical management phrase is: “We expect capacity, not immediate salary savings. Operations will use it to reduce the overdue request queue. We will review queue size, quality and paid overtime after the trial.” That sentence is less exciting than a large saving claim and far more useful for running a company.

The NIST AI Risk Management Framework supports considering risks across AI use and evaluation. In financial planning, I translate that into funded evaluation, clear escalation and a realistic allowance for maintaining the workflow.

Buy evidence before buying scale

A small pilot should test the assumptions that could overturn the investment decision. Establish the baseline, record complete task time and review quality. Agree what result justifies the next step. The goal is a decision your team can defend, including a decision to stop.

I can help connect your automation idea to the actual workflow, implementation scope and benefit assumptions. Bring me the process and the rough numbers, or review my AI product and delivery background. If the demo already exists but the value does not, start with why AI pilots fail to reach daily work.

Questions people ask

How do you calculate AI automation ROI?

Choose a period and accounting view, then estimate the benefits minus all relevant costs and divide that net benefit by costs. Include setup, integration, training, review, corrections and ongoing operation. State assumptions about volume and adoption. Separate cash savings from the value of released capacity, and avoid counting the same time saving again as revenue or avoided overtime.

Are hours saved by AI the same as cash savings?

No. When salaries stay unchanged, time saved usually represents capacity rather than immediate cash. Cash benefits require a credible mechanism, such as reduced paid overtime or avoided external spending. Capacity can still improve service or enable useful work. Give it an owner and a specific destination, then measure whether that benefit actually occurs after the workflow is introduced.

What hidden costs should an AI business case include?

Include data preparation, integration, evaluation, employee practice, output review, exception handling, monitoring, support and future updates. Count internal time in a resource view even if it creates no extra invoice. Keep it separate from cash costs and avoid double-counting work already deducted from the estimated time saving. Test uncertain assumptions before committing to a wider rollout.

Something here sounds like your workplace?

Let’s have a real conversation ↗