Philippines staffing blog ·

Set a safe concurrency limit for chat agents

How to choose how many chats one agent handles, with the evidence, guardrails, and review routine that keep quality from slipping.

Set a safe concurrency limit for chat agents help desk operations illustration

Direct answer

Operational guidance

Why concurrency needs a stated limit

Chat work looks efficient when one agent holds several conversations at once, because the agent stays busy while customers wait and the queue appears to move. That picture hides the pauses, the half-read messages, and the replies written before the previous answer was understood. A customer does not experience the average of those conversations, only the one thread where an answer arrived late or missed the point. A stated limit turns an invisible trade-off into a decision the desk can review.

Without a limit, concurrency rises whenever volume rises, because the tool allows it and the queue rewards speed. The fastest agents absorb more chats and the slowest absorb fewer, so the average hides a wide range of customer experiences. Supervisors then compare a calm transcript with a rushed one and conclude that the difference is effort rather than load. The limit should protect the quality of a single thread, not the look of the dashboard.

Measure the work between messages

Time each step of a chat: reading the message, checking the account, writing the reply, and waiting for the customer. The wait is the only part that can safely overlap with another conversation, and even then the agent must remember the earlier context. An agent handling three chats is not doing three times the work, because the typing and checking still happen one at a time. Measure the active minutes and the context switches, not the number of open tabs.

Count the interruptions that force a context switch, because they are the real cost of concurrency. A customer who sends four short messages creates four chances to lose the thread, and each re-read takes time that never appears in a volume report. Record how often an agent must scroll back before replying, and how often a reply answers the wrong question. Those moments are the evidence a limit should be built on.

Choose a limit you can defend

Set the limit from the work, not from a vendor default or a competitor's screenshot. Start with the number that lets an agent answer within the stated first-response time while keeping every reply specific to its own conversation. Test one step lower than the busiest agents claim they can handle, and watch the transcripts for a week before raising it. The limit should feel slightly conservative on a normal day, not heroic on a good one.

Write the limit into the tool rather than leaving it to habit, because an unwritten rule disappears during a busy shift. Name who may raise it, for how long, and what evidence is required, and make the override visible on the queue board. A supervisor overriding the limit during a spike should record the reason and the end time. The exception is then a documented decision instead of a quiet new normal.

Protect the quality checks

Sample chats that happened at the limit, not only the calm ones that are easy to review. Read the full thread and ask whether each reply made sense in its own conversation, and whether the customer had to repeat information. Look for copied greetings, answers that skip a question, and closures that arrive before the customer confirms. A quality score built on easy chats will not show the cost of concurrency.

Review the first reply and the last reply as a pair, because together they show the shape of the conversation. The first shows how quickly the agent engaged, and the last shows whether the thread stayed coherent and the customer left with a next step. Check that the ticket note matches what was actually said in chat, including any promise made in passing. When the note and the transcript disagree, the limit is probably too high for that request type.

Handle spikes without breaking the limit

A queue spike is a reason to change routing, not a reason to ignore the limit. Move simple requests to a self-service path, add a second shift, or tell customers when an agent will reply instead of leaving them in a silent queue. If the desk must exceed the limit, cap the duration and assign a supervisor to watch the transcripts during the exception. Short, supervised exceptions are safer than a silent permanent increase.

Give agents a way to flag a conversation that needs full attention, such as a billing dispute, a vulnerable customer, or a thread with conflicting instructions. The flag should pull the agent down to one chat until the case stabilises, and it should be quick to use without a written justification. Review the flags later to see which request types deserve their own limit. Agents usually know where the risk sits before the report does.

Review the limit after a month

Compare the month at the limit with the month before it, using first-response time, full-resolution time, reopened chats, and quality-sample failures together. A limit that improves speed while raising rework is not working, and a limit that slows the queue slightly but cuts repeated contacts may be the better trade. Look at the distribution, not only the average, because a few very long chats can hide behind a healthy mean. The review should describe what customers experienced.

Ask agents what changed in how they read and write, because their answers will describe the pauses the metrics cannot show. Adjust the limit one step at a time and keep the previous month as the comparison, so the effect of the change is visible. Document the reason for each adjustment and the evidence behind it. The goal is a limit the desk can explain to a client and defend with transcripts rather than opinion.

Document the decision

Write a one-page note that states the limit, the date it took effect, and the evidence used to choose it. Include the request types covered, the exception process, and the person who can change the number. Keep the note in the same place as the queue playbook so a new supervisor can find it during a shift. A limit with no written basis will be renegotiated every busy week.

Review the note with the client when the service scope changes, because a new product or a new channel can alter the work behind each chat. Ask the client whether the limit conflicts with any promise it has made to customers. Record the answer and the date, even when the answer is that nothing changes. The document should show that the trade-off was discussed, not only that a number was set.

Train the limit

Let new agents practise with two chats before they handle the full limit, and review the transcripts with them at the end of the session. The practice should include a customer who sends several short messages and one who changes the subject mid-thread. Ask the agent to narrate how they keep the conversations separate. The exercise builds the habit before the live queue adds pressure.

Calibrate supervisors on the same transcripts so they judge load consistently across shifts. A supervisor who raises the limit for one agent and not another creates an invisible performance rule. Review two or three chats together each month and compare the observations, including the moments where the agent had to choose which customer to answer first. Consistency in supervision matters as much as the number itself.

Authoritative sources

Related planning pages