Two hundred and thirteen roubles
against four thousand two hundred and twenty-eight

Why the figure cannot come from an industry report

The first figure comes from the high-volume line of a proxy infrastructure service: live chat, round the clock, routine questions about connection and payment. The second comes from support for SCADA-class industrial software, where a contact means investigating a technical case and reproducing a fault on a test rig, sometimes over several days, sometimes with engineering involved.

Hence the first conclusion. Not because industry reports are bad, but because the word "contact" stands for different units of work in different companies. Until you have said what exactly counts as one unit, any benchmark figure is somebody else's number about somebody else's work.

The mistake I made myself

At first my figure was not 213 roubles but 100. And I very nearly compared one hundred roubles with four thousand two hundred, producing a neat fortyfold spread.

A fortyfold spread does not exist. The hundred roubles covered only the agents on the line: their pay, taxes and the ticketing licence. The 4,228 roubles covered the whole department, including the manager's payroll. Once I added the manager and their tooling to the first calculation, the figure became 213 roubles and the spread turned into twentyfold. Still a lot, but now it is true.

That last point taught me something on its own. The calculation included one employee on a standard employment contract out of five, the rest on other arrangements. Social contributions for that single person cost the function about 8.7 roubles on every contact out of a hundred. The employment arrangement of one person in five moved the unit figure by almost nine per cent.

The formula is simple; the difficulty is in the denominator

Costs for a period divided by the number of contacts handled in the same period. That is all. The work begins after that, and almost all of it sits in two places.

One-off payments

A project bonus, extra pay for clearing a ticket backlog, an annual bonus. It is paid in May but earned over January to May. If you book it in May, May becomes more expensive than its neighbours and you go looking for what happened in May. It has to be spread across the period it was earned in, in proportion to something sensible: in my case, hours on the line.

A weighted average, not a simple one

Monthly unit figures must not be averaged. You add up the costs and the contacts, then divide one sum by the other. The gap between the two approaches widens the more the traffic swings.

What the figure shows and what it does not

Four months of the high-volume line, cost per contact for agents only: April 86 roubles on 1,780 contacts, May 106 on 1,551, June 103 on 2,178, July 104 on 2,689.

• April is the cheapest, not because the work was more efficient: there were fewer people, and two of them did not start at the beginning of the month
• May is both the most expensive and the emptiest. Costs hold, volume falls, and the unit figure rises on its own
• From June it plateaus at about 103 roubles even though traffic grew by half: the team grew with the traffic
• From August the figure will drop by roughly 6.5 roubles because the only employee on a standard contract left and the contributions went with them

So cost per contact is not an efficiency metric. It shows the structure of costs and the price of a unit of work. Efficiency only becomes visible when quality and speed stand next to it and all three move together.

The second measure that checks the first

I also calculate the cost of a minute of departmental work. It is useful in its own right: contacts vary in length, a minute is always a minute.

There is a trap here too, and I walked into it. Counting departmental hours from the shift timesheet is wrong: timesheets contain overlaps where shifts run into each other. In my case June came out seventeen hours long and July fifty-five. The line runs round the clock with one person always on it, so departmental hours are twenty-four for every calendar day, and no timesheet is needed.

The result was 9.9 roubles per minute for the department and 4.7 roubles for agents only. In industrial software support the same measure was 210 roubles. Contacts differ twentyfold, minutes twenty-one fold: two unit measures built from different denominators produced the same order of difference. That is the cross-check. Had I made an error somewhere, they would have diverged.

What it is for, beyond tidy tables

• The payback of automation. An AI agent handles a third of the traffic. While the effect is measured in "percentage of conversations closed", the conversation with finance does not work. In roubles per contact it becomes an ordinary investment discussion
• Where to fix things, in service or upstream. Diagnosis showed that a third of the traffic came from supplier quality. Fixing it upstream is cheaper than hiring people to handle the consequences
• Workforce planning. Once the cost of a minute is known, "do we need another agent" becomes arithmetic rather than an argument about workload

The sequence that works for me

And one last thing. A figure calculated once stays true for about three months. What is valuable is not the figure but the recalculation procedure, written down in a file and clear to whoever comes after you.