Why most AI ROI numbers are useless

Two kinds of number circulate in this market, and neither helps you.

The first is the vendor case study: "Company X reduced handling time by 62%." It may be true. It tells you nothing about your business, your volumes, your staff costs or your process.

The second is the productivity survey: "AI saves the average knowledge worker 4.1 hours a week." Averages across industries and job types are not a business case; they are a headline.

The only number that matters is the one produced by measuring your own process before and after. This guide is how to produce it.

The four-step calculation

Step 1 — Pick one process. Just one. The one wasting the most time or losing the most revenue. ROI calculated across "the business" is not calculable.

Step 2 — Count what it costs today.

For time cost:

hours per week × number of staff involved × loaded hourly cost

Loaded hourly cost is salary plus overhead — employer contributions, benefits, space, equipment. Roughly 1.3× base salary is a workable planning figure. Use your actual number if you have it.

For revenue cost:

missed or slow-handled opportunities per week × conversion rate × average value

Include both where both apply. A slow after-hours response costs time and revenue.

Step 3 — Estimate the automated version.

What share of this work can the system handle end to end? Be conservative. Sixty to seventy per cent of a costly, well-documented process is a strong result, and it is a more defensible planning assumption than ninety.

Then add what remains: the exceptions still handled by humans, the review time, and the ongoing ownership cost — the hour or two a week someone spends maintaining it. This is the line most people forget, and it is the one that turns an optimistic case into an honest one.

Step 4 — Compare, honestly.

Before cost vs (after cost + running cost + amortised implementation cost)

Amortise the implementation over twelve months. If the after figure is not clearly better than the before figure, do not proceed. If it is, you have found a project that pays for itself, and you know roughly when.

If your ROI case only works at ninety per cent automation, it is not a business case — it is a hope.

A worked example

Illustrative only — run it with your own numbers.

A clinic misses around eight enquiries a day outside opening hours. Historically, roughly 40% of comparable enquiries convert, at an average first-visit value of SGD 120.

Revenue at risk: 8 × 0.4 × SGD 120 = SGD 384 per day, roughly SGD 8,000 per month.

Additional time cost: two staff spend about five hours a week between them handling the enquiries that do get answered, at a loaded SGD 38 per hour — about SGD 800 per month.

Total current cost: roughly SGD 8,800 per month.

Assume the automated version recovers 60% of the missed enquiries and removes 70% of the handling time:

Recovered revenue: SGD 8,000 × 0.6 = SGD 4,800.

Time saved: SGD 800 × 0.7 = SGD 560.

Gross monthly benefit: SGD 5,360.

Against that: a monthly running cost of SGD 400, an ownership cost of two hours a week at SGD 38 (about SGD 330), and an implementation of SGD 12,000 amortised over twelve months (SGD 1,000).

Total monthly cost: SGD 1,730. Net monthly benefit: SGD 3,630. Payback: under four months.

Now notice what makes this credible: every input is a number the clinic could actually measure, and the assumptions (60%, 70%) are conservative and stated. That is a defensible case. "AI reduces admin by 62%" is not.

Establishing the baseline properly

Your baseline is the whole business case, so measure it honestly.

Measure for two weeks minimum. One week is noise.

Include the chasing and the correcting. The visible task is rarely the whole cost.

Count the misses, not just the completions. Missed work is invisible in most systems and is often the largest number.

Use loaded costs, not salary. Salary understates by roughly a third.

Write it down and circulate it. A baseline agreed before the project starts cannot be renegotiated afterwards.

Proving it afterwards

Six weeks after go-live, measure the same things the same way. Then be disciplined about three traps.

Attribution. Did the number move because of the automation, or because it was a busy month? Where possible, compare like periods, or run a control — one location automated, one not.

Displacement. Did the work disappear, or move? A support system that resolves faster but generates more repeat contacts has moved the cost, not removed it. Always measure the counter-metric.

Hidden new costs. Ownership time, exception handling, integration maintenance. Count them. A project that ignores them will report a return it is not delivering.

The metrics worth tracking, by use case

Customer-facing automation: first-response time; resolution rate without human involvement; repeat-contact rate within seven days; conversion of enquiries received outside staffed hours; customer satisfaction split between automated and human interactions.

Finance automation: touch time per item; coding error rate; items chased; working days to close.

Sales automation: time from lead received to first substantive response; qualification turnaround; percentage of leads contacted within an hour; conversion by response-time band.

Internal knowledge: questions answered without interrupting a senior; time to find an answer, sampled; proportion of questions with no documented answer, month on month.

The soft benefits — count them, but separately

Shorter month-end closes, staff who stop doing the worst part of their week, faster response times that win work you never see, a complete audit trail when a query arrives. These are real and often larger than the measured savings.

Keep them in a separate section of your case. Mixing unquantified benefits into a financial calculation is how ROI models lose credibility, and losing credibility on project one makes project two much harder to approve.

The summary

One process. Two weeks of honest baseline measurement. A conservative automation assumption. Full costs on the other side, including ownership. Compare, and proceed only if the gap is clear.

Then measure again at six weeks, watch the counter-metrics, and report the result honestly — including if it did not work. A business that has one credibly measured result can approve the next project in an afternoon. A business with a folder of vendor case studies cannot.

Start with a free AI Readiness Assessment

Normally valued at SGD 1,500, currently free. You receive an opportunity report, a prioritised roadmap and honest ROI estimates for your own processes — with no obligation. Book yours at aigentify.tech/assessment.

AIgentify — Singapore-based AI implementation specialists. We design, build and support AI agents, workflow automations and intelligent business applications with measurable ROI. Live in weeks, not months. Your data stays yours.

This article is general information, not advice. Grant schemes, platform rules and regulatory requirements change; confirm current details with the relevant authority or provider before relying on them. Any figures shown are illustrative unless a source is stated.