
How to Forecast Live Chat Staffing
Quick answer: Work it out hour by hour, not for the whole day. For each hour, take the chats that reach a person, the average handle time, and how many chats an agent really handles at once. Run Erlang C with an adjustment for those parallel chats to find the smallest team that answers your target share in time, such as 80 percent within 60 seconds. Then add cover for breaks and absence, check agents are not too busy, turn the hourly numbers into shifts, and compare the plan with real waits every week.
Step 1: Forecast chats per hour that reach a person
Export two to four recent weeks of conversations from your help desk and count them by day of week and hour of day. Three rules keep the forecast honest:
- Count only chats that reach people. If a bot answers first, the conversations it closes on its own never need an agent. Use chats with a human handoff, and track the bot's share separately as your deflection rate.
- Plan the busy hours, not the daily average. A day of 160 chats might hold 30 in the 11:00 hour and 5 at 17:00. Averaging hides the hour that breaks your target.
- Adjust for what you know is coming. A sale, a product launch or a broadcast campaign raises volume. Assembled warns that a forecast off by 25 percent can leave a team several people short or over, so keep a note of what you expected against what arrived.
Step 2: Measure handle time and real concurrency
Handle time is the agent time per chat from first reply to wrap-up, including notes and tagging; see average handle time. Chats usually take longer than calls because customers pause between messages. Measure it from your own reports, not from another company's figure.
Concurrency is how many chats an agent works at the same moment. Your help desk probably has a chat limit per agent; the useful number is how many they actually handle on average at busy times, which is usually lower. Call Centre Helper notes that handle time tends to rise as concurrency rises, so check both together: a limit of three that pushes each chat from 10 to 16 minutes may help less than it seems. Choosing and testing that limit is covered in setting live chat concurrency limits. If agents handle very different topics, measure each group separately, as routing chats to the right agent splits them.
Step 3: Set the answer target
Erlang C needs two numbers: the wait you aim to stay under, and the share of chats that should meet it, for example 80 percent within 60 seconds. Choose a target your team can keep at the busy hour. The steps for picking it, with priorities and business hours, are in setting live chat SLA targets, and the term is defined in service level agreement (SLA).
Step 4: Run Erlang C with a concurrency adjustment
Erlang C was built for phone calls, where one agent handles one contact. There is no agreed correction for chat, and practitioners use one of three shortcuts:
| Method | How it works | Source |
|---|---|---|
| Shorter handle time | Divide handle time by a concurrency factor, then run Erlang C as normal | Call Centre Helper, Erlang C for Chat? |
| Divide the result | Run Erlang C as if chats were calls, then divide agents by the average concurrency | Call Centre Helper, Chat Operations Capacity Planning |
| More servers | Multiply agents by chats per agent | Our response time SLA calculator |
For the first method, Call Centre Helper suggests factors below the chat count, because agents are not always at their limit: 1.7 for two chats, 2.5 for three and 2.9 for four. Increase the number of agents one at a time until the service level meets your goal. Any of the methods gives a starting estimate, not an exact answer, and comparing two of them is a quick sanity check. Call Centre Helper names simulation as the more accurate alternative, at the cost of more setup, and recommends the Erlang estimate as a reasonable starting point.
Step 5: Add shrinkage and check occupancy
Shrinkage is paid time when agents are not available for chat: breaks, meetings, training, holidays and sickness. Call Centre Helper divides the agents needed by one minus shrinkage, so 6 agents with 30 percent shrinkage becomes 6 ÷ 0.7 = 8.6, rounded up to 9; Assembled and Call Centre Helper both use 30 percent as an example, not a rule. Use your own records. For a very small team, it is often more useful to plan breaks and meetings outside the busy hours than to divide every hour by the same factor.
Occupancy is how busy agents are: workload divided by agents. Call Centre Helper uses an 85 percent ceiling in its example, because agents kept busier than that tire and handle times rise. Small teams often sit well below it because each whole agent adds a lot of spare capacity. That slack is what lets a small team keep a fast promise.
Step 6: Build shifts and check against real waits
Line up the hourly numbers for the week and draw shifts that cover each hour's need, with breaks placed in quiet hours. Then compare the plan with what happened:
- Every week, compare predicted and actual. Look at the share of chats answered within the target by hour. A gap at the same hour each week points to a wrong forecast or wrong handle time for that hour.
- Watch the leavers. Erlang C assumes nobody gives up. If many visitors leave the chat queue before an answer, read managing a live chat queue and track your abandonment rate.
- Use the bot as a lever. If the numbers say you need a person you cannot hire, a bot that answers more questions lowers volume, and one that collects details before the handoff can shorten handle time. The chatbot handoff design guide covers what to collect. Tools differ in how they report handoffs; compare inboxes in our Intercom review and Zendesk.
Check the arithmetic
An invented example, not a measurement. A shop's busiest hour, 11:00 on Mondays, brings 30 chats that reach a person. Handle time is 12 minutes, agents take up to two chats at once, and the goal is 80 percent answered within 60 seconds.
| Step | Calculation | Result |
|---|---|---|
| Effective handle time | 12 min = 720 s; 720 ÷ 1.7 | 423.5 s, about 7.1 min |
| Workload | 30 × 423.5 ÷ 3,600 | 3.53 Erlangs |
| Service level with 5 agents | Erlang C | 68.6%, misses 80% |
| Service level with 6 agents | Erlang C | 87.1%, meets 80% |
| Occupancy | 3.53 ÷ 6 | 58.8% |
| Scheduled with 30% shrinkage | 6 ÷ 0.7 = 8.6 | 9 people rostered |
Cross-check with the divide-the-result method: as if chats were calls, 30 chats of 12 minutes is 6 Erlangs; 8 agents give 69.8% and 9 give 84.7%, and 9 ÷ 1.7 = 5.3, rounded up to 6. Both methods agree on 6 at the moment. In the shop's quiet 17:00 hour, 10 chats give 1.18 Erlangs: 2 agents reach 61.2% and 3 reach 89.6%, so 3 are needed. All results were computed with the standard Erlang C formula; the shop and its numbers are made up to show the method.
Common mistakes
- Planning on the daily average instead of the busiest hour.
- Counting chats the bot closed as work for agents.
- Using the chat limit as concurrency when agents actually handle fewer at once.
- Forgetting shrinkage, so the rota covers the hour on paper but not after breaks.
- Treating the result as exact. Erlang C is a starting estimate; real waits are the test.
FAQ
How many live chat agents do I need? It depends on chats per hour reaching people, handle time, concurrency and your target. In our invented example, 30 chats per hour at 12 minutes each, two at a time, needs 6 agents on chat to answer 80% within 60 seconds, and 9 rostered with 30% shrinkage.
How do I calculate staffing when agents handle several chats at once? Adjust Erlang C for concurrency: shorten handle time with a concurrency factor, or run it as for calls and divide the agents by average concurrency. Both are estimates to check against real waits.
Can I just divide phone staffing by the number of chats per agent? Call Centre Helper says it is not that simple, because agents are not always at their maximum and handle time rises with concurrency. Use a factor below the chat limit.
What shrinkage should I use for a chat team? Your own: add up breaks, meetings, training and absence as a share of paid hours. Published examples use 30 percent, but that is an example, not a standard.
Does Erlang C work for email or tickets? Not well. Assembled notes it suits real-time channels, not channels where replies take hours or days.
How often should I redo the forecast? Check it weekly against real waits, and redo it when volume, handle time or the bot's deflection changes.
Sources
- Call Centre Helper, Erlang C Formula – Made Simple With an Easy Worked Example — callcentrehelper.com/erlang-c-formula-example-121281.htm, read 16 November 2026. Source of the Erlang C steps, the shrinkage division and the 85% occupancy ceiling.
- Call Centre Helper, Erlang C for Chat? — callcentrehelper.com/erlang-c-for-chat-206984.htm, read 16 November 2026. Source of the effective handle time method, the concurrency factor table and the note that handle time rises with concurrency.
- Call Centre Helper, Chat Operations Capacity Planning — callcentrehelper.com/chat-operations-capacity-planning-207067.htm, read 16 November 2026. Source of the divide-the-result method.
- Assembled, How do you use Erlang C to plan staffing? — assembled.com/university/how-do-i-use-erlang-c-to-plan-staffing, read 16 November 2026. Source of the 30% shrinkage example, the 25% forecast warning and the email caveat.
- Chatbotscape evaluation methodology. /methodology (continuously updated).