A voice agent can talk to one customer at a time. A chat agent can, in principle, talk to several, because a chat conversation has gaps: the customer is typing, looking for an order number, or simply away from the screen. Concurrency is the number of chats an agent handles simultaneously. It is the main reason chat support can be cheaper per contact than voice, and also the main reason chat support can feel slow and robotic to customers when it is pushed too far.

Where the spare capacity comes from

Take a typical e-commerce chat about a delayed order that lasts 10 minutes from start to finish. Watch the agent closely and perhaps 6 of those minutes involve real work: reading messages, checking the order system, typing responses. The other 4 minutes are spent waiting for the customer. With one chat at a time, those 4 minutes are idle.

Concurrency fills the gaps. While customer A looks for her order ID, the agent answers customer B. Done well, two chats overlap neatly and the agent’s idle time disappears.

The stretch effect

The catch is that chats do not wait politely for each other. When two customers reply at the same moment, one waits. With more chats open, those collisions become more frequent, and each customer waits longer for each reply. Every chat takes longer from start to finish. On top of that, switching between conversations has a mental cost: re-reading the history, remembering which customer has which problem, and occasionally pasting the wrong order number into the wrong window.

An illustrative model shows the effect on throughput (chats completed per hour) when handle time per chat stretches as concurrency rises:

ConcurrencyHandle time per chatChats per hour
110 min6.0
214 min8.6
320 min9.0
430 min8.0

Throughput is concurrency × 60 ÷ handle time. Moving from one to two chats gives a large gain, roughly 43% more chats per hour. Moving from two to three adds only a little more. Moving to four makes output fall, because each chat has stretched so much. Meanwhile, the customer experience has deteriorated at every step: at concurrency four, a 10-minute conversation has become a 30-minute one.

The specific numbers will differ by process, but the shape is common: a steep early gain, a plateau, and then decline. The goal is to find the plateau’s start for each process, not to push to the maximum the chat tool allows.

What decides the right number

  • Complexity of the issue. Order status chats can run at higher concurrency than banking disputes or technical troubleshooting, where the agent must concentrate on one system at a time.
  • System speed. If the order lookup takes 40 seconds to load, the agent is effectively blocked during that time and other chats wait.
  • Customer pace. Customers on mobile phones often reply more slowly, which creates more natural gaps. Customers on desktop during work hours may expect near-instant replies.
  • Agent experience. New agents should start at concurrency one or two. Experienced agents with strong typing and good canned responses can manage more.
  • Sensitivity. Complaints, bereavement cases or fraud reports deserve full attention. Many routing set-ups mark such chats as “single” so the agent receives no others until it closes.

Most chat platforms allow a maximum concurrency per agent or per skill. Better configurations also allow dynamic limits, for example reducing the cap when an agent’s response time exceeds a threshold.

Metrics that make concurrency visible

Voice metrics transplanted directly into chat can mislead. Chat AHT always looks long compared with voice, and it rises with concurrency, so an AHT target used without context can punish the agents carrying the most load. Useful chat measures include:

  • First response time: how long the customer waits for the first human reply after being connected.
  • Average response time: the typical wait between a customer message and the agent’s reply during the chat. This is the metric that degrades first as concurrency goes up.
  • Chats per productive hour: throughput, which captures the real benefit of concurrency.
  • Average concurrency: the actual number of simultaneous chats over time, not the configured maximum.
  • Customer abandonment mid-chat and CSAT: customers who give up because replies are slow are a strong sign of overload.

Watching throughput and average response time together tells you whether more concurrency is helping. If throughput is flat and response time climbs, the cap is too high.

Planning with concurrency

For staffing, chat is often planned using workload hours divided by effective concurrency, with care taken to use the handle time that applies at that concurrency, not the single-chat figure. A common planning mistake is to measure handle time when agents ran at concurrency one, then assume a concurrency of three, which overstates capacity because the stretch effect has been ignored. Erlang-based tools need adjustment for chat for the same reason.

Another pitfall is the “conversation that never ends”. With asynchronous messaging such as WhatsApp business channels, customers may come back hours later. Those threads are better handled as asynchronous queues with response-time SLAs than as live concurrent chats, otherwise they block an agent’s slot indefinitely.

The human side

High concurrency is tiring in a way that dashboards miss. Agents juggling four chats for eight hours make more errors late in the shift, and the errors are the embarrassing kind: wrong names, wrong orders, a reply meant for one customer sent to another. Customers notice when an agent clearly is not reading their messages. Good canned responses and knowledge base shortcuts help, but the most reliable protection is a sensible cap and a team leader who watches response times in real time.

Written by the RP Strativa Team. The handle times and throughput figures in the model are illustrative.

Leave a Reply

Trending