Callback and virtual hold features are one of the fastest ways contact centres improve the customer experience, and one of the easiest ways to wreck a staffing plan. The dashboard looks great after go live: queue times fall, abandonment drops, and the wallboard turns green. Then, a few weeks later, a backlog of callbacks sits at the end of the day and nobody can explain why the roster that "matched the forecast" cannot clear the work.
The reason is simple. A callback does not remove a contact. It moves it. This post uses a simple Erlang C model to show what that means for forecasting, rostering and the service level metric itself.
Why callbacks confuse the forecast
Most teams forecast offered calls, then apply Erlang C to the offered volume. When virtual hold or queue callback goes live, a share of those callers press a key and leave the live queue. The ACD reports fewer handled calls in the live queue, fewer abandons and a shorter wait. If the forecast is rebuilt on that new live volume alone, the plan quietly drops the work that moved into the callback list.
The workload is still there. It is simply waiting for an agent to dial out. That is why every callback programme needs two forecasts: live inbound contacts and callback work items.
The test day
To keep the numbers honest, We modelled a single 12 hour day in 30 minute intervals.
| Assumption | Value |
|---|---|
| Total offered calls | 2,095 |
| Peak interval | 180 calls (around 10:30) |
| Average handle time | 360 seconds |
| Service level target | 80% in 20 seconds |
| Callback take up | 30% of callers when the interval has 110 or more calls |
| Callbacks created | about 445 |
Three planning approaches were compared for the same day.
- Real time only (no callback). Standard Erlang C on all offered calls.
- Callback aware. Live load is capped at 150 calls per interval. Anything above the cap is called back in the next interval with room.
- Naive volume cut. The 30% of callers who took the callback are simply removed from the forecast, and nobody plans to work them.
What the three plans cost
| Plan | Peak agents | Agent half hours rostered | What happens |
|---|---|---|---|
| Real time only | 42 | 525 | Meets 80/20 at every interval |
| Callback aware | 36 | 527 | Peak is 14% lower, total work is the same |
| Naive volume cut | 31 | 430 | Looks 18% cheaper, but the work does not disappear |
The callback aware plan is the real prize. It does not save hours. It flattens the peak. You need 6 fewer agents at the busiest time because the extra demand is served later, in the quieter intervals. That helps with peak shift coverage, break scheduling and overtime. It does not shrink the total workload.
The naive plan is the trap. It looks like an 18% saving on paper, but those 445 callbacks still need around 44 hours of agent time.
The hidden backlog
In the naive plan, agents work callbacks only when they have spare capacity (capped at 88% occupancy so people are not run flat out). The result for the day:
By closing time, about 22 hours of callback work is still waiting, which is roughly half of all the callback work created. Customers who were promised a call back "shortly" are now being called the next day, or not at all. Add failed dial attempts and voicemail (modelled at 40 extra seconds per callback on average) and the peak backlog climbs from 34.5 to 39.5 hours.
This is the classic callback failure pattern: a great morning, an overloaded afternoon, and a rising number of repeat contacts the following day.
Service level lies when callbacks are on
There is a second problem. Service level is calculated on calls that stay in the queue. When your longest waiters accept a callback, they leave the queue and stop counting against the metric. So the live queue SL improves, even while the customer actually waits longer for the issue to be solved.
A healthy callback report should therefore track these together:
| Metric | Why it matters |
|---|---|
| Live service level and ASA | Shows queue health |
| Callback offered and accepted rate | Shows how much work is being moved |
| Time to callback (median and 90th percentile) | The true customer wait |
| Callback success rate (connect on first attempt) | Failed dials add hidden workload |
| Callbacks open at end of day | The backlog that rolls into tomorrow |
| Blended wait (queue wait plus callback wait) | The metric that matches customer experience |
A small error compounds fast
Even without a backlog, a small mistake in the planned handle time hurts. In the model, the real time peak with 42 agents delivers 82% service level. If callbacks add just 5% extra work per contact (dial time, voicemail, redials) that is not planned for, the peak service level falls to 68%. At 10% extra it falls to 46%. Callback AHT must be forecast separately from inbound AHT, because the call has different phases.
A practical WFM playbook for callbacks
- Forecast two queues. Live inbound and callback work items, each with its own volume and AHT.
- Model take up by interval. Take up rises with wait time, so it is highest at the peak. Use a curve, not a flat percentage.
- Set a callback promise. For example, 90% of callbacks completed within 60 minutes. Staff to that promise.
- Create a callback skill or priority. Make sure callbacks are worked in the first spare capacity, not left to chance.
- Plan the dial out time. Reserve capacity in the intervals right after the peak, when the backlog is being cleared.
- Cap the callback window. Stop offering callbacks 60 to 90 minutes before close, otherwise you create work you cannot finish.
- Review failed attempts. Use retry rules and set the number of attempts, because every redial is workload.
- Report blended wait. Do not celebrate the live queue SL alone.
Key takeaways
- A callback moves workload. It does not remove it.
- The real benefit is a flatter peak (6 fewer agents at the busiest time in this model), not a smaller roster.
- Removing callback volume from the forecast creates a hidden backlog of about half the callback work by close of day.
- Live service level will look better than the customer experience. Measure time to callback and blended wait.
- Forecast callback AHT separately, including failed attempts.
To test your own numbers, use the Erlang C calculator for the live queue and the capacity planning calculator to size the callback work. If AI or automation is also changing your volumes, read why AI deflection breaks your staffing model and why a hiring freeze is not a staffing plan, because the same mistake applies: volume that moves is not volume that disappears.
The figures above come from a simplified model with the stated assumptions. Real centres should test callback take up, AHT and success rates against their own data.