Cost model
What AI customer support really costs an ecommerce team
The subscription is visible. Knowledge cleanup, implementation, quality review and retained human coverage are where weak business cases usually hide.
The six-line cost model
- Helpdesk base: the inbox or platform subscription.
- AI usage: resolved conversations, AI conversations or another usage unit.
- Overages and channels: tickets above plan, voice, SMS and premium integrations.
- Setup: implementation partner, migration and workflow configuration.
- Internal implementation: policy cleanup, testing and the time of support, operations and legal owners.
- Ongoing control: weekly review, knowledge updates, failed-action diagnosis and escalation handling.
A useful model compares the new stack with the current one. If a new platform replaces €250 of existing software, subtract that amount. If the old helpdesk stays, do not.
Price the work that disappears
Start with monthly ticket count and average handle time. Apply an expected full-resolution rate, not the share of tickets that receive any AI reply. Then subtract the human time that remains for review or escalation.
For 2,000 tickets, seven minutes each and a 40% full-resolution rate, gross released time is 93 hours. If automated contacts still consume half a minute of human work, the net is about 87 hours. At a loaded labour cost of €32 an hour, that is roughly €2,770 of monthly capacity value.
Capacity value is not the same as cash saved. It can mean faster replies, absorbed growth, less overtime or time reassigned to retention work. Call it payroll saving only if the staffing plan actually changes.
Include the ramp
Suppose the new software adds €500 a month after replacing an older plan. Setup costs €1,500 and the team spends 40 internal hours at €32, making total implementation cost €2,780. Spread that across a six-month evaluation window and the ramp charge is about €463 a month.
Against €2,770 of monthly capacity value, the model shows roughly €1,807 of monthly net benefit during the six-month ramp. That is a planning case, not a promise. The result collapses quickly if the resolution rate is overstated or the queue is mostly complex exceptions.
Vendor pricing does not use one unit
Gorgias prices its helpdesk by billable ticket volume and adds an outcome-based AI Agent charge. Its September 2026 explanation says most AI plans charge around $0.90 for an interaction the agent fully resolves. Tidio describes Lyro pricing around $0.50 per conversation and offers 50 trial conversations. The units and bundle rules differ, so a headline per-conversation comparison is unsafe without matching what each vendor counts.
Ask each vendor for a written quote using the same month of anonymised volumes. Include normal season, peak season, channel mix and expected AI usage. Check annual discounts and overage rates separately.
Three scenarios, not one answer
| Scenario | Full-resolution assumption | Use |
|---|---|---|
| Conservative | 20% | Would the project still be worth owning if only the cleanest repetitive work automates? |
| Base | 40% | A planning case to validate in a controlled pilot. |
| Upside | 60% | A capacity view, not the number used to approve the budget. |
These percentages are scenario inputs, not AislePilot benchmarks. We will not publish an industry rate until the dataset passes the methodology thresholds.
Four ways an ROI case lies
- It calls assisted replies fully resolved tickets.
- It ignores the time humans spend fixing the knowledge base and reviewing failures.
- It values every released hour as cash reduction.
- It counts revenue lift and cost saving together without a controlled measurement plan.
The calculator keeps revenue lift off by default for this reason. Cost and capacity are measurable first. Conversion claims require a separate experiment.