How to Calculate the ROI of a Support Agent
You don’t need a financial model on forty tabs. Three numbers any head of support has give an answer accurate enough for a decision.
A CFO does not need a presentation about “customer experience transformation”. They need an answer to two questions: how much this saves per month and when the investment pays back. Good news — for a first-line support agent this math is done on a napkin with accuracy sufficient for a decision. Bad news — most calculations we see in other vendors’ proposals are inflated 2–3x because they ignore what the agent does not cover. Below is the model we give our clients, with all the correction factors — and with where it lies.
Three numbers you already have
N comes from the helpdesk. T is the share of questions an operator answers from a template or looks up in the knowledge base: “where is my order”, “how do I return it”, “which documents are needed”. Usually 55–80%. C is the support payroll divided by the number of requests; typically 120–350 RUB for chats and 400–900 for phone.
The math
An example with real orders of magnitude. An online retailer, 12,000 requests per month in chats and messengers, 65% routine, 180 rubles per request (14 operators including taxes, vacations and a supervisor). Gross savings: 12,000 × 0.65 × 0.8 × 180 = 1,123,200 RUB per month.
Now subtract what other proposals stay silent about
The gross number is not savings. To get an honest result you must subtract four items, and each of them matters in real projects.
1. Inference cost. One support request is on average 4–6 dialog turns, each with a RAG context of 3–5 thousand tokens. On Claude Sonnet or GPT-4o-class models that is 12–25 rubles per request at current rates; on Qwen 72B in your own perimeter — 3–6 rubles plus GPU amortization. For our example: 12,000 × 0.65 × 0.8 × 18 ≈ 112,000 RUB per month on an external API.
2. Maintenance. The agent is not a box. Pricing changes, regulations change, new question types appear, quality degrades after the provider updates the model. A realistic norm is 0.25–0.5 FTE of an engineer or a support contract: 50,000–90,000 RUB per month depending on the volume of changes.
3. Escalations got more expensive. This is counterintuitive and almost nobody counts it. The agent took the simple questions — the operators are left with the hard ones. The average handling time of the remaining requests grows by 30–60%, which means the cost of one “human” request grows too. Part of the gross savings is eaten by the fact that the remaining 35% of requests now cost 250 instead of 180 rubles.
4. Launch costs. Turnkey agent development with integration into the helpdesk and CRM — 400,000–600,000 RUB by our price list; plus 20–40 hours of your employees’ time for labeling references and acceptance. These are capital costs; they go into the payback period, not into the monthly economics.
What the model doesn’t count — in your favor
There are effects the calculation doesn’t include because they are hard to measure before launch, but they are real and usually outweigh the model’s conservatism.
- Nights and weekends. According to our retail deployments, 18–25% of requests come outside working hours. Previously some of these clients went to a competitor without waiting for an answer. That is lost revenue, not a cost, and it does not get into the model.
- First-response time drops from minutes to seconds. It affects CSAT more than the quality of the answer itself — clients forgive an imperfect answer but not the wait.
- Scaling without hiring. A seasonal peak 2.5x bigger does not require hiring and training temporary operators a month before Black Friday. The agent scales with tokens.
- Operator turnover on routine questions is the highest. Less routine — people stay longer, hiring and training get cheaper.
When a support agent won’t pay off
The model is also useful because it shows when the venture is bad. Three typical scenarios.
Too few requests. With N below 1,500–2,000 a month, gross savings don’t cover maintenance and tokens. Here the agent makes sense only as a product feature (24/7), not as savings.
Too few routine ones. B2B support of complex software with T around 20–30% — the agent will close few requests, and expensive escalations will eat the rest. Better to start with an internal assistant for the operators themselves: it speeds up search over the knowledge base without talking to the client.
No knowledge base. If the answers live in the heads of three senior operators, the first two months of the project go into extracting them. Useful work, but it must be baked into the timeline and budget — and it is not about AI.
Plug in your N, T and C. If the net savings after deductions pay back the launch faster than six months — let’s run a pilot. If not — we will say so at the audit, because a failed project costs us more than a deal that does not happen.