By Tina20 Nov,2026
The cheapest AI message is not necessarily the cheapest resolved customer issue. A low usage price can still produce an expensive workflow if agents must check every answer, repair mistakes and reconstruct failed transfers.Calculate cost at two levels: the AI-specific cost of independently resolved work and the total service cost of the whole workload. Use the same issue definition, period and quality standard when comparing with the manual baseline.
1. Gather the inputs before calculating
| Input | What to include |
|---|---|
| Workload | Distinct issues or another consistently defined unit |
| Completed outcomes | Resolutions under a stated observation rule |
| Software | Subscription, seats and relevant add-ons |
| Usage | AI units, overages and applicable channel or provider charges |
| Human handling | Residual cases and assisted-case work |
| Review | Quality checks and exception supervision |
| Maintenance | Knowledge updates and relevant administration |
| Setup | Training, configuration and migration, allocated explicitly |
Use current supplier quotations for actual prices. The figures below are invented arithmetic examples, not Duoke prices, labour benchmarks or a forecast for any country.
Assume 500 service issues are resolved in the month under the same quality rule. Without AI, each requires six active handling minutes. At an illustrative loaded labour cost of $12 per hour, the baseline is 50 hours, or $600.
With AI, suppose 220 issues are independently resolved and 280 still need human handling. The remaining cases take five active minutes each. Add four hours of quality review and two hours of knowledge maintenance. The illustrative software and usage costs are shown below.
| Monthly cost | Calculation | Amount |
|---|---|---|
| Subscription | Assumed | $120 |
| AI usage | Assumed | $60 |
| Channel or provider cost | Assumed | $20 |
| Remaining human handling | 280 × 5 / 60 × $12 | $280 |
| Quality review | 4 × $12 | $48 |
| Knowledge maintenance | 2 × $12 | $24 |
| Total operating cost | Sum | $552 |
The all-in cost per resolved issue is $552 / 500 = $1.104, approximately $1.10. The manual baseline is $600 / 500 = $1.20. Operating savings are $48, or 8%, before any setup cost.
This is a modest result despite 44% of issues being independently resolved. That is the point of including the remaining work: automation coverage alone does not determine the financial outcome.
In the example, software, AI and channel costs total $200. Dividing by 220 independent resolutions gives about $0.91 per AI-resolved issue. That excludes the $352 spent on human handling, review and maintenance.
It is legitimate to report the narrower number if labelled clearly. It is misleading to present it as the cost of the entire service operation. Likewise, dividing $200 by every bot message would describe a different unit, not resolved issues.
If some of the 220 cases reopen within the observation window, revise the independent-resolution count and include the follow-up effort. Do not keep the lower denominator-based cost by treating failed closures as permanent successes.
Suppose initial setup takes ten hours at the same hypothetical $12 hourly cost: $120. Charging all of it to the first month raises that month's total to $672, or $1.344 per resolved issue. Allocating it over several months produces a different monthly view; state the allocation rather than hiding it.
Then change the assumptions that matter most. What if fewer issues are independently resolved? What if human cases take longer because they are the difficult ones? What if review remains intensive during the first month?
| Assumption to vary | Why the result may change |
|---|---|
| Independent resolution volume | Spreads fixed costs across more or fewer completed issues |
| Residual handling time | Remaining cases may be disproportionately complex |
| Review effort | Early-stage supervision can be substantial |
| Usage growth | Campaign volume may trigger overages |
| Reopened work | Adds human effort and reduces durable independent completion |
Do not assume the non-automated cases retain the old average handling time. Removing easy questions can leave a harder human workload.
Time released is not automatically payroll removed. A small team may use the capacity to answer older enquiries, improve coverage or handle a campaign without hiring additional temporary help.
Report those benefits in their own terms. “Agents recovered ten hours for unresolved cases” can be useful evidence without claiming the business immediately saved the full wage equivalent in cash.
Treat revenue uplift separately and cautiously. An order placed after a chatbot conversation may not be incremental. Avoid funding a purchase decision with an untested assumption that every assisted sale was caused by AI.
Ask each vendor to price the same workload and define its billable unit. Clarify whether a resolution, conversation, message or credit event creates the charge and which subscriptions or channel costs remain separate.
For Duoke, verify the current qualifying independent-resolution credit rules and the Lite subscription offered to your account. A resolution-based AI rule does not mean every software or channel charge disappears when a case needs a person.
Use pilot measurements to replace assumptions. A cost model becomes decision-ready when its workload, quality standard and residual effort are observable, not when its spreadsheet contains a particularly attractive percentage.

Email:[email protected]
Address:6/F MANULIFE PLACE, 348 KWUN TONG ROAD, KOWLOON, Hong Kong, 999077