The conversation about AI in support is usually held at two extremes: “it will replace everyone” and “it is a toy”. Both are useless for making a decision. Below is the calculation we run before the brief, so that a client can say “no thanks” before paying anything.
What we are counting
First line means the enquiries whose answer already exists in your own material: where is my order, how do I return an item, what is included in the plan, how do I recover access. That is not consulting and not selling. It is exactly this part that gets automated.
Take an online school: 3 000 enquiries a month, three first-line operators, and an average hourly cost per employee including taxes and overheads of 800 roubles.
The cost side
The rollout. The work splits roughly like this: half the time on the knowledge base, a quarter on integrations and escalation rules, a quarter on quality evaluation and rework. For the volume described that is 3–6 weeks and a range of 300–500 thousand roubles.
Half the timeline on the knowledge base is not an exaggeration. The material almost always contradicts itself: the help centre gives one return window, the contract gives another, and the operator has a third one in their head. The agent surfaces those contradictions immediately and mercilessly, and people have to resolve them.
Running it. The cost of a conversation is made up of model calls and search over the base. At realistic volumes that is 5–15 roubles per conversation depending on length and model. For 3 000 enquiries a month, 15 to 45 thousand roubles.
Maintenance. A knowledge base goes stale: plans, terms and processes change. Without updates the quality of the answers drops within a quarter. That is a few hours a month—yours or ours.
The saving side
Not “making operators redundant”. The realistic scenario is taking load off them. The agent closes 60–70% of first-line enquiries; the operators handle the remaining 30–40% and the hard cases they previously only touched in snatches.
Three operators × 800 roubles × 160 hours = 384 thousand roubles a month. Taking away 65% of the routine load frees up roughly 2 full-time posts—about 250 thousand roubles a month in working time that moves over to retention and to difficult enquiries.
Saving minus running cost is on the order of 210 thousand roubles a month. Payback on the rollout is 2–3 months.
Where the calculation breaks down
Too few enquiries. Below 500 enquiries a month the saving does not cover maintenance. It is cheaper to hire a person or write a good help centre.
No material. If the knowledge base exists only in employees’ heads, the project starts by creating it—that is separate work, and it has to be budgeted into the timeline and the price.
The enquiries are not routine. If every second one requires looking into the CRM, doing a calculation and making a decision, that is not first line, and it is automated differently: through actions with human review, not through answers.
The cost of an error is high. Medicine, finance, legal advice. Where a wrong answer creates liability, the agent prompts the operator and stops there. It never becomes a channel of its own.
What to measure after launch
- Share of enquiries closed without an operator. The main figure.
- Share of escalations and their reasons. Rising means the knowledge base is falling behind reality.
- Accuracy on a labelled set. 100–300 questions with reference answers, replayed on every change. Without that, “it got better” is a feeling.
- Cost per conversation. It creeps up as the context gets longer.
- The user’s rating of the answer. It catches what the labelled set does not see.
The short version
The right question is how much one enquiry costs now and how much it will cost afterwards. If the difference is smaller than the cost of maintenance, there is nothing to roll out. We say that at the brief.