An AI Reply Is Not a Resolved Support Ticket

A reply is an event. A resolution is an outcome evaluated over a period of time. Measuring the first while budgeting for the second can make a support workflow look inexpensive without explaining how much work remains.

Open the support-resolution example. Its default inputs reproduce this article’s base case. Replace the assumed rates with measurements from your own ticket cohort and follow-up window.

Follow one cohort all the way through

Start with 1,000 incoming tickets. Assume two AI attempts per ticket at $0.01 per attempt. The total AI bill in this simplified model is $20.

Initially, 80% of the tickets appear resolved by AI, giving 800 initial resolutions. Assume 25% of those initial resolutions reopen during the selected follow-up window. That is 200 reopened tickets, not 25% of all incoming tickets.

The durable AI count is therefore 800 minus 200, or 600. The other 400 tickets consist of 200 that were never resolved by AI plus 200 that reopened. In this model, all 400 escalate to a person exactly once.

Human time belongs in the same cost model

Assume each escalation takes six minutes at a labor rate of $30 per hour. Each escalation then costs $3, and 400 escalations cost $1,200. If the human path resolves every escalation, the cohort ends with 1,000 resolved tickets.

Dividing the combined $1,220 by those 1,000 resolved tickets produces $1.22 per resolution. Quoting only the $0.01 per AI attempt would answer a much narrower question.

const tickets = 1000;
const durableAi = tickets * 0.80 * (1 - 0.25);
const escalations = tickets - durableAi;
const aiCost = tickets * 2 * 0.01;
const humanCost = escalations * 6 / 60 * 30;
const resolved = durableAi + escalations;
console.log((aiCost + humanCost) / resolved); // 1.22

The assumed labor rate should reflect the cost definition you use consistently. Adding a loaded rate in one alternative and an hourly wage in another makes the comparison asymmetric.

An unresolved ticket still incurred cost

Now change the human resolution rate from 100% to 80%, holding the escalation count and time constant. Humans resolve 320 of the 400 escalations. Combined with the 600 durable AI resolutions, that produces 920 resolved tickets.

Spending remains $1,220 because all 400 escalations consumed the assumed time. Cost per resolved ticket rises to about $1.326. The 80 unresolved cases are not removed from the cost numerator, and they should not be quietly counted as successes.

This scenario still stops after one escalation. Repeated contacts, transfers, or later work can add costs outside the model. Extend the measurement window or include those costs explicitly instead of assuming they are free.

Compare against a real alternative

To assess whether automation helps, measure an alternative workflow against the same incoming cohort or a defensibly comparable cohort. Include resolution quality, delay, reopens, and human effort. A higher resolution rate is not sufficient when an answer resolves the wrong issue or causes another problem.

The model does not assign a monetary value to customer satisfaction or response speed. Those can affect the decision, but they need explicit measures rather than being inferred from a low token bill. Likewise, no claimed savings are produced here against an invented all-human baseline.

Use the accepted-task calculator for a more general outcome-cost comparison when your escalation logic does not match this simplified model.

Sources and method

This is an original illustrative cohort calculation, not customer support research. Durable AI resolutions equal initial resolutions multiplied by one minus the reopened share. All remaining tickets escalate once; human resolution applies only to that escalated set. Both base and incomplete-human-resolution cases are covered by regression tests.

• AI Support Resolution Cost calculator
• Cost per Accepted Task
• AI Plan Stress Test