How Long Will a Caller Wait? Hold Time Statistics

What does the research actually say about caller patience?

The honest answer is that there is no magic number. No credible study gives a single figure like "callers hang up after 90 seconds" that applies across all businesses, industries and call types. What the research does show is more useful: caller patience is shaped by context, by what callers can see and hear while they wait, and by operational choices the business makes.

A 2015 study of 1,485 customers across two call centres used actual waiting times from system databases and self-reported patience to develop a correction formula¹. The researchers developed a correction formula based on polynomial regression estimates that aligned self-reported maximum impatience with actual behaviour. The implication for small businesses is that self-reported maximum impatience can be corrected using actual waiting time data, and observed behaviour is the more reliable guide.

20 statistics on caller patience and abandonment

Call abandonment and operational data

1. A 2026 study analysed 1,037,363 call records from a large academic healthcare system across four operational periods².

2. Machine learning models trained only on nonpersonal, operational data could predict call abandonment, with the best configuration achieving an AUC of 0.767 on an unseen holdout set².

3. Adding operational metrics improved model performance by 0.03 to 0.13 on the area under the curve compared with models using only temporal and skill features².

4. Operational features in the best-performing models included in-queue time, occupancy, handle time, after-call work time and the number of active agents².

5. The best configuration was a CatBoost model trained on one operational period with no imbalance correction, achieving a PR-AUC of 0.068, expected calibration error of 0.006 and Brier score of 0.027².

6. Models trained solely on preintervention data generalised poorly to postintervention periods when restricted to temporal and skill features².

7. The study spanned four operational periods marked by workflow changes and skill consolidation, covering January to December 2023².

8. Three imbalance-handling strategies were tested: none, class weighting and synthetic minority oversampling technique².

Perceived progress and reneging

9. A 2024 study using lab and online experiments demonstrated a "relative progress effect" in queues³.

10. Participants who observed fast service times early in a wait, followed by slower service times, perceived faster progress and were less likely to renege, even when total wait time was identical to a queue progressing at a steady rate³.

11. In one of the studies, participants experiencing relative progress waited for a longer queue or longer duration than those in a benchmark queue without the effect³.

12. A simulation analysis showed that individual-level effects increasing patience for some reduced the overall renege rate in the queue³.

13. The relative progress effect was created when a series of fast service times were observed early in a wait, generating a sense of relative progress for the observer³.

14. Queue practices that modulate the perception of relative progress can reduce renege behaviour³.

Scheduling and patience distributions

15. A 2023 scheduling study modelled multi-class impatient customers in parallel server queueing systems, where each class represents service and patience time distributions as well as cost characteristics⁴.

16. For nonexponential patience distributions, processing customers of the same class under first-come, first-served is suboptimal⁴.

17. The researchers proposed a multi-class time-in-queue policy that prioritises both across customer classes and within each class using a simple rule⁴.

18. Most of the gains could be achieved by deviating from within-class first-come, first-served for at most one customer class⁴.

19. For systems with exponential patience times, the proposed policy reduces to a simple priority-based policy, which is asymptotically optimal for Markovian systems with an optimality gap that does not grow with system scale⁴.

Time-dependent queues

20. A 2021 study developed a new stationary backlog-carryover approach for time-dependent queues with generally distributed abandonments, where arrival rates vary over time⁵.

21. The method allows for the analysis of underloaded and overloaded systems with time-dependent arrival rates and generally distributed time to abandon⁵.

22. A discrete-event simulation benchmark confirmed the method approximates queue performance well across different parameter configurations, including an empirical distribution from real-world call centre data⁵.

23. The study presented a limit result indicating convergence of the method for stationary parameters⁵.

24. The approach was effective for instances with small numbers of time-dependent servers and gamma-distributed abandonment times with different coefficients of variation⁵.

Self-reported patience vs actual behaviour

25. A 2015 study of 1,485 customers across two call centres used actual waiting times from system databases and self-reported patience to develop a correction formula¹.

26. The researchers developed a correction formula based on polynomial regression estimates that aligned self-reported maximum impatience with actual behaviour¹.

27. Survival analysis provided valuable information which could easily be utilised by service providers¹.

28. The correction of self-reported maximum impatience can be applied to self-reported statements of maximum patience¹.

29. The study aimed to measure maximum patience, the point when the customer first becomes dissatisfied with waiting time, rather than maximum impatience, the point when they renege¹.

30. The results empower service providers in optimising utilisation of human resources while maintaining customer satisfaction¹.

What drives call abandonment?

A 2026 study of 1,037,363 call records from a large academic healthcare system tested whether machine learning models could predict call abandonment using only anonymised operational data². The answer was yes, with meaningful accuracy.

The models included operational metrics such as in-queue time, occupancy, handle time, after-call work time and active agents². Adding these operational metrics improved model performance by 0.03 to 0.13 on the area under the curve compared with models using only time and skill features. The best configuration achieved an AUC of 0.767 on the unseen holdout set T3b².

What this means for a small business is that abandonment is not random. It tracks measurable, controllable factors. When wait times stretch, when staff are stretched thin, when calls sit in queue, more callers leave. Models trained solely on preintervention data generalised poorly to postintervention periods when restricted to temporal and skill features². In plain terms: when you change how you handle calls, old patterns of caller behaviour stop predicting new ones.

For a deeper look at what abandonment rate means for a small firm, see our guide on call abandonment rate.

Does perceived progress matter more than actual wait time?

One of the more striking findings in the queueing literature is that how a wait feels can matter as much as how long it is. A 2024 study using lab and online experiments demonstrated a "relative progress effect" in queues³. When participants observed a series of fast service times early in their wait, followed by slower service times, they perceived faster progress and were less likely to renege, even when the total wait time was identical to a queue progressing at a steady rate.

The effect was not trivial. In one of the studies, participants experiencing relative progress waited for a longer queue or longer duration than those in a benchmark queue without the effect³. A simulation analysis showed that individual-level effects increasing patience for some reduced the overall renege rate in the queue³.

For a small business, the practical translation is this: the early part of a caller's wait sets the tone. If a caller hears a prompt response, an acknowledgement, or sees progress, they are more likely to hold. Silence and stasis do the opposite. This aligns with what we know about why customers hang up instead of leaving voicemail.

Do different callers have different patience levels?

Yes, and this has operational consequences. A 2023 study on scheduling in queueing systems modelled customers as having different patience time distributions and cost characteristics⁴. The researchers showed that within-class first-come, first-served is suboptimal for nonexponential patience.

Their proposed policy prioritises customers both across classes and within each class, using updated estimates of residual patience as time elapses⁴. Most of the gains could be achieved by deviating from within-class first-come, first-served for at most one customer class⁴. In practical terms, a small business that can identify which callers are most time-sensitive, for example an existing customer with an urgent issue versus a general enquiry, can route or prioritise accordingly.

How do time-dependent queues behave?

Call volumes are rarely steady. A 2021 study developed a new approximation method for time-dependent queues with generally distributed abandonment times⁵. The researchers modelled systems where arrival rates vary over time and customers may leave if their wait exceeds their personal patience.

A discrete-event simulation benchmark confirmed the method approximates queue performance well across different parameter configurations, including an empirical distribution from real-world call centre data⁵. The relevance for small businesses is that peak periods behave differently from quiet periods. A staffing level that handles Tuesday morning fine may fall apart on Friday afternoon, and abandonment will spike accordingly.

What can a small business actually do with this?

The research points to a few concrete principles, even if it does not hand you a single hold time target.

Shorten the wait, especially at the start

The relative progress effect suggests the first moments of a wait matter disproportionately³. Answer quickly, acknowledge the caller, and give them a reason to stay on the line.

Watch your operational metrics

Abandonment can be predicted from operational factors such as wait time and staffing levels². If you can measure how long callers wait and how many give up, you can see the problem before it becomes a pattern.

Don't rely on what customers say they'll tolerate

Self-reported maximum impatience can be corrected using survival analysis of actual call centre data¹. Track what callers do, not what they claim they would do in a survey.

Consider how calls are routed

Not all callers are equally patient, and treating them as if they are leaves performance on the table⁴. A simple triage step, such as asking whether a call is urgent, can help.

Plan for peaks

Time-dependent demand means your worst hours, not your average hours, define the caller experience⁵. Staff or route calls accordingly.

For a practical guide to handling calls outside normal hours, see our comparison of voicemail, AI and answering services. If you want scripts for common situations, our small business phone script toolkit is a good next step.

What the research does not tell you

None of the studies cited here gives a universal hold time threshold. The healthcare study involved a large academic system with over a million call records, which is a very different context from a two-person salon or a solo tradesperson². The experimental queueing studies used lab and online participants³. The scheduling research is mathematical modelling validated by simulation^4,5.

So if you were hoping for "callers will wait exactly 2 minutes and 14 seconds," the literature does not support that. What it supports is a more useful conclusion: caller patience is measurable, malleable and partly under your control.

How to use this data this week

Start by measuring one thing: how long callers wait before they are answered or abandon. Most phone systems, including basic mobile and VoIP setups, log missed calls with timestamps. If you see a pattern of abandonment during specific hours, that is your peak problem⁵.

Then test one change. Answer faster during your busiest hour. Add a short acknowledgement message if you cannot answer immediately. The relative progress effect suggests that even a brief, early response can reduce abandonment³.

Finally, review your routing. If every call goes to the same queue regardless of urgency, you are treating impatient and patient callers identically, which the research suggests is suboptimal⁴. A simple "press 1 for urgent, press 2 for everything else" or a receptionist who triages can make a measurable difference.

For businesses that cannot staff a phone line during peak or off hours, an AI receptionist that answers 24/7 and books appointments into your calendar is one option to consider. It is not a substitute for understanding your call data, but it removes the silence that drives abandonment.

Bottom line

The evidence on caller patience is moderate in quality and consistent in direction. Abandonment can be predicted from operational factors such as wait time and staffing levels², perceived progress reduces reneging³, patience varies across callers⁴, and time-dependent demand shapes the experience⁵. What is missing is a universal threshold, because none exists. The businesses that handle calls well are not the ones that memorise a magic number. They are the ones that measure, shorten and signal progress.

How to cite this page

Unleashed Newsroom. "How Long Will a Caller Wait? Hold Time Statistics." Unleashed, 1 October 2026, unleashed.vision/news/how-long-will-caller-wait-on-hold.