Speech Analytics Call Sampling as a Cost Management Tactic

Speech analytics has proven itself as a powerful tool for improving contact center performance. Large enterprises everywhere use it and the growth of the speech analytics industry is a powerful signal that speech analytics delivers results.

One drawback, however, has been the level of investment required. An enterprise system requires ‘six figures’ investment. Scarce talent to run the system is hard to retain. For this reason, speech analytics has been relatively less affordable for mid-market businesses and smaller call centers that don’t have the economies of scale that much larger firms do. This has locked out many smaller companies from keeping up with their larger competitors in the areas of customer satisfaction and sales effectiveness.

Two recent developments promise to change that: ‘Speech Analytics as a Service’ and ‘call sampling.’ In this post, we review the second of these developments.

What is Call Sampling

When companies license their own dedicated instance of a speech analytics platform, they routinely process all their calls. They get a 100% view of the customer phone contacts. This is the ideal end state. The data generated by the system are a full reflection of the state of the call center. For use cases like compliance, no call is missed. This assures that non-conforming calls do not fall through the cracks.

When delivered as a SAaaS targeted at the mid-market, speech analytics services from companies like Zenylitics are usually structured on a cost per hour basis. Minimums can be as low as 500 recorded call hours per month. Even with the ability to process low call volumes, that cost may be prohibitive for smaller companies. Luckily, managed service providers like Zenylitics have developed sampling strategies to bring sophisticated speech analytics processing within budget for companies that otherwise fear they cannot otherwise afford the cost.

The approach is conceptually simple. Instead of processing 100% of all recorded calls, a sample is selected. This reduces costs while still providing important insights into the customer’s business — assuming it is done in a statistically meaningful manner.

Sampling Strategies

Sampling can be a powerful way to contain costs. For it to provide meaningful results, however, it is important to choose an approach that ensures that the conclusions generated are meaningful.

Sometimes, the sample can be trimmed by the customer at the source. For example, omitting calls less under a certain duration threshold can eliminate calls that have minimal value. Unless the interest is specifically about what happens in calls under thirty seconds, this is one easy way to reduce the volume. In Zenylitics experience, a rule of thumb across industries are that 10% of calls last 30 seconds or less.

Also, perhaps a customer wants to ‘walk before they run.’ For example, they may have a sales group and a support group. It would be a plausible approach to start with the highest value group and prove the value there. Perhaps an increase in close rate from the sales group would demonstrate value clearly. So the program would focus only on those calls and then move on to support calls once the sales program is established.

In an ideal world, the distribution of a sampled population would reflect the same distribution of the underlying population a customer wants to analyze. For example, the percentage of calls answered by a given agent in the sample would be similar in proportion to the full sample. Another example: the time-of-day distribution of calls in the sampled population would reflect that of all calls.

Of course, it is unlikely all traits will exactly match up. Therefore, it is important to determine the most important measures and to quantify the confidence with which the sampled population must reflect the whole call population.

Calculating Confidence in a Speech Analytics Sample

We can calculate the confidence in which a sampled population reflects its parent group. Confidence is expressed as the ‘plus’ or ‘minus’ percentage at a given standard deviation. For example, a sample confidence of +-4% for a 75% compliance with a Right Party Contact (RPC) behavior would mean that 95% of the time the full population’s RPC performance for that metric is between 71% – 79% based on the analyzed sample.

The exact metric for RPC can only be obtained by processing the entire call population. But when that is not feasible, one can see that this sampled measure is way ahead of how many call centers review calls today. Many call centers review 3% or less of their calls.

Not only are 3% or less of calls analyzed, they often are chosen in an unscientific manner. For example, calls in many shops are selected from just the working hours that the person reviewing the calls happens to keep.

In certain industries, only specific call types are reviewed. For example, in an insurance call center, perhaps only the calls resulting in a sale may be reviewed for compliance. Sometimes only ‘squeaky wheel’ calls exhibiting an escalation are reviewed. And sometimes the calls are even self-selected by the agent. Agents often have a stake in the data and therefore potentially will exhibit strong selection bias. Such biased samples can lead to erroneous conclusions.

The above example may be well-intentioned attempts to get a meaningful sample. Unfortunately, even more common is that calls are just ‘randomly’ chosen with no attempt at generating a balanced sample. Too often, call auditing is just a job to get done. The listening to a few calls for each agent per month is a job to check off the list. This undercuts the investment in call review and makes the resulting data suspect.

Well-planned sampling then, can deliver a huge advance over many shops’ status quo. When properly done, samples in the range of 50% – 75% of the call population can yield high confidence rates for which the generated statistics vary less than 5% from the underlying call population.

Applications for Sampling in Speech Analytics as a Service

In general, the more homogenous a population of calls, the more appropriate it is to use sampling. Here are some of the best uses:

Call Categorization

Often an important first step in deploying speech analytics is just to characterize the distribution of call types. This important information can reveal sources of customer dissatisfaction that inadvertently generate unneeded calls. By identifying call intent and resolving customer concerns before they generate a call, companies can reduce call traffic and avoid customer dissatisfaction by removing the problem before an unneeded contact is created.

Some call centers already categorize call types by tagging them with agent-generated ‘dispositions.’ Dispositions are labels the agent assigns to signify how a call was concluded. Sometimes, the dispositions are too general or not applied consistently. Speech analytics can shed new light on emerging trends in call traffic by processing the call contents and generate data-driven dispositions that are more accurate than relying on the agent’s subjective tag.

Understanding what drives a customer to pick up the phone and contact a business can be a major driver for positive change.

• To determine granular call intent. For example, is the customer looking for the status of loan funding as opposed to status of loan application.
• To provide insight into marketing initiatives. For example, does proactive account status notifications lead to reduction in call volume.
• To notice emerging trends in the marketplace. Are customers requesting discounts or looking to cancel more frequently.

Measuring Customer Satisfaction

When the measure of customer satisfaction is clear, sampling can often generate a useful customer satisfaction metric. For example, a powerful tactic for measuring CX is to ask the customer at the end of a call if the call solved the problem they were calling about. Thus, a ‘yes’ or ‘not yes’ disposition is generated. With a clear disposition value, sampling holds the potential to generate meaningful data.

Comparing Groups

If there are multiple call centers or agent groupings, sampling may be useful in characterizing performance between or among the groups. For example, you should be able to compare groups for script adherence, active selling language, and objection handling. By doing this, one may detect the need for additional training or detect when a contractor’s performance is lagging. Also, sampling calls from a new contractor vs. other more experienced call centers may quantify the time and additional support needed to effectively onboard new providers.

Comparing the Impact of Policy Changes

When a company grows rapidly, it needs to learn how customers experience it and how that experience evolves. This can be done by establishing a customer satisfaction baseline. The baseline may be generated in anticipation of a product launch, a process change, price change, or other policy events that affect customers.

Once the benchmark is generated, the company can introduce its change and compares the impact with the baseline. Because change creates unexpected side effects, the next step is then to analyze the two measures for a statistically meaningful difference. By using speech analytics to measure changes in customer experience, companies can respond quickly and thereby avoid unnecessary churn.

False Comparisons

While sampling can be useful as a cost-containment tactic, you must be sure the resulting data is statistically significant. This requires that the differences are greater than the confidence range. For example, a difference of 15% would be significant for a measure with a confidence of +-5%. If the difference in the measure was less than 5%, it would not. It can be tempting to assign meaning to insignificant differences when in fact differences less than the confidence range are technically not meaningful.

Limitations of Speech Analytics Sampling

There are scenarios where sampling must be used with caution. In these scenarios, you should favor larger samples or process all calls to get actionable data.

Compliance Monitoring

Where speech analytics is used to ensure compliance with consumer protection regulations like FDCPA/UDAAP and HIPAA, sampling may not be appropriate. The danger is too great that you miss important out-of-compliance calls. While you may generate useful KPI’s that show that most calls are conducted appropriately, it may not be acceptable to miss any calls that require mitigation.

Individual Agent Performance Assessment

Using speech analytics as a coaching tool is a classic use for improving call center effectiveness. This remains true even when the data is generated with a small sample. Sampling is likely to capture calls that reveal the strengths and weaknesses of individual agents. You can use this for coaching. Speech analytics will detect common errors agents make on every call. Each call has its own ‘truth’ and can be a resource for coaching

However, when the sample size is small, use caution in drawing larger conclusions about individual agent performance. You may not be able to make meaningful inferences without processing a bigger sample of call activity from any single agent.

Hybrid Sampling – The Sampling ‘Canary in the Coal Mine’

One important strategy is to blend sampling with the processing of an entire population of calls.

For example, a debt collections client established a policy for data gathering requirements on every inbound customer contact. All agent training and support documents reflected this policy. Company leadership was under the impression that the policy was being followed. During the analysis of the initial sample of calls, speech analytics processing showed there was a gap between expectation and reality.

Because of the gap, the customer processed all outbound calls for several months to fully examine the calls for compliance. The analysis revealed a pattern in which agents were skipping the verification policy on accounts where the customer had already been contacted. The policy was only followed with first contact consumers. This was a legacy behavior based on a prior policy that no longer applied. The client addressed this behavior with additional training. This process improvement ensured that subsequent outbound contact efforts were focused on up to date and stable information.

Sampling can be the ‘canary in the coal mine’ to detect areas that require further analysis. Then, by processing concentrated samples of a suspect call type, the client receives a complete analysis enabling them to fix the root cause of the situation.

The opposite tactic is also possible. For example, for a first-time user of speech analytics for compliance, a customer may process all calls to generate a complete picture. Many companies tell us, “We don’t know what we don’t know.” Thus, we start by processing 100% of calls to bring the unknown fully to light. Then, after refining their compliance program and building confidence that agents are following that program, customers can transition to a sampling program that monitors compliance.

Conclusion

Speech Analytics as a Service is a powerful approach to obtaining the benefits that previously have only been available to larger firms. Delivered as a Service, a managed speech analytics service frees companies from hiring specialized staff and making the large capital investment that enterprise software systems historically require.

Bought by the call hour processed, speech analytics may be a bigger investment than some shops feel they can afford until the program has truly proven its worth. In these cases, the use of sampling can provide many of speech analytics’ benefits. By combining sampling strategies with an informed approach to interpreting its results, SAaaS speech analytics can be a game-changing resource for improving call center effectiveness.