Automating email handling: the numbers for a distributor
Take a typical B2B distributor with a five-person support team and around 900 emails a month: orders, status questions, returns, and complaints. Someone has to read every email, classify it, check the data in the ERP, and type out a reply. Below we break that process down into numbers and show what changes once an agent is in place. A caveat up front: this is a model calculation built on typical values, not a report from a specific implementation — the point is that you can check it and drop in your own data.
- This is a model calculation, not a description of a specific client — you can substitute your own numbers.
- At 900 cases of 14 minutes each, manual handling costs roughly PLN 9,450 a month.
- An agent taking over 77% of cases leaves the team with 48 h a month instead of 210 h.
Starting point: 900 emails a month
In an audit we always measure the same three values: the number of cases, the average time per case, and the hourly rate. Here: 900 cases a month, 14 minutes per case, PLN 45/h. That is 210 hours and PLN 9,450 a month — for work that mostly consists of copying data between the inbox and the ERP. For the pilot you pick the most frequent case type: quote requests and B2B returns.
How the agent works inside the inbox
The agent works on the same inbox and the same ERP the team uses today — no migration. One case goes through four steps:
Reading the request. The agent reads the email with its attachments and extracts the data: customer, order numbers, subject of the case. 0.6 s.
Classification. The case gets a type — a B2B return, say — and a priority. Unusual cases go straight to a person. 0.9 s.
Data from ERP and CRM. The agent pulls the order status, commercial terms, and customer history through the existing API. 1.4 s.
Reply or escalation. Standard case: the reply goes out to the customer and the ERP gets an entry. Any doubt: the agent drafts a reply and waits for approval. 1.1 s.
Average handling time per case before and after implementation, across all 900 cases a month, escalations to a person included.
What changes in the numbers
Assume the agent handles 77% of cases on its own. That leaves the team 207 cases a month, or 48 hours instead of 210 — 162 hours go back into substantive work. The cost of handling one case drops from PLN 10.50 to PLN 2.41, and the monthly cost of the process from PLN 9,450 to PLN 2,173. Over a year that is roughly PLN 87,000 on a single process.
What the numbers don't show matters just as much: customers get an answer within seconds in the evening and at weekends too, and the support team deals with exceptions and conversations rather than copying order numbers. How much of this you would see depends on your share of unusual cases — which is why we start the audit with your numbers, not with this model.
Where this calculation most often breaks down
Two places where the model drifts from reality. First: the share of unusual cases. If half the emails need a commercial decision, 77% is unrealistic and 40–50% is the better estimate. Second: data quality in the ERP — if order statuses are sometimes out of date, the agent inherits that error, and it has to be fixed before automating, which stretches the pilot. In any case, the hardest work in an implementation like this is not on the AI side but on the rules: what the agent does alone, and what always needs approval.
Frequently asked questions
Does the agent reply to customers entirely on its own?
For standard cases, yes — within the rules agreed with the team. When it is uncertain, it drafts a reply and waits for a person to approve it, and the full decision history is logged.
What happens to unusual emails?
They reach the team with the context already assembled: classification, ERP data, and a proposed reply. The person decides, the agent learns — and the exception list shrinks week by week.
Did the ERP or the inbox have to be replaced?
No. The agent connects to existing systems through APIs, the database, or exports — including systems without a modern interface. The team works in the same tools as before.