How to calculate the real ROI of AI phone automation

ROI models for call automation usually compare an hourly wage against a per-minute rate, conclude that the numbers look good, and omit the two variables that actually determine the outcome.

Mistake one: costing the wrong thing

The comparator is not salary. It is fully loaded cost per handled call: salary plus employer contributions, recruitment, training, management overhead, workspace, systems, and the idle time between calls. In most teams the fully loaded figure is substantially higher than the headline wage, and utilisation is well below one hundred percent.

Mistake two: ignoring missed calls

Unanswered calls do not cost you a wage — they cost you the margin on the business that went elsewhere. For appointment-driven businesses this is usually the largest single number in the model, and it is the one nobody tracks because a missed call leaves no record.

Estimate it directly:

• Calls received outside answering capacity per month
• Percentage that represent genuine new business
• Conversion rate for calls you do answer
• Average gross margin per converted customer

Multiply those four and you have the recovery opportunity. It frequently exceeds the labour saving.

Mistake three: assuming full automation

Assume an escalation rate. A realistic first-year deployment resolves the routine majority and hands the rest to a person. Model the automation share honestly — sixty to eighty percent of a well-chosen intent is a credible target, one hundred percent is not.

A model that survives scrutiny

Cost side: subscription plus expected usage, plus the internal time to configure and review. Include roughly half a day a week of ownership in year one.

Benefit side: labour hours redeployed (not necessarily removed), recovered missed-call revenue, reduced no-shows where reminders are added, and the value of faster response on new enquiries.

Risk side: escalation rate higher than planned, and a knowledge base that needs more work than expected. Both are usually discovered in the first fortnight.

Validate on one intent

The fastest way to a trustworthy number is a narrow pilot. Route one call type, measure resolution and escalation for a month, then extrapolate from real data instead of assumptions. A model built on your own transcripts is worth more than any vendor benchmark.