- Chapter I: Introduction
- Chapter II: What is a Contact Center Conversation Intelligence Platform?
- Chapter III: AI-Driven Voice Analytics: The Heart of Conversation Intelligence
- Chapter IV: Generative AI-based Automated Quality Management (Auto QA)
- Chapter V Automated Data Redaction
- Chapter VI AI-Powered Agent Assist
- Chapter VII Additional Considerations
- Chapter VIII Conclusion
1. Speech-to-Text Transcription
As described below, speech-to-text transcription is the foundation on which Voice Analytics stands. Therefore, the first quality marker — one of the most important — is how good the speech recognition and Natural Language Processing (NLP) engine is. While there are amazing NLP engines available on the market, their high cost is a prohibitive factor for many organizations. Therefore, most vendors have opted to create their own.
Ask questions that allow you to gauge the quality of the transcription. For example, inquire about the current accuracy rate and how well it handles accents. Ask if there are any limitations you should know about, e.g., the number of minutes transcribed per month. Is there a product roadmap that maps out planned improvements over the next 12-24 months? Keep in mind, however, that there will be a point where the quality of the NLP is good enough for what you need, so don't get too carried away. Otherwise, you will have to make a trade-off between accuracy and affordability.
Image: This screenshot shows the Voice Analytics tab within the call details in MiaRec. The contact center manager can immediately see that this call involved went very well both on the agent and the customer side and that it was about a purchase.
2. Automatic Keyword Extraction & Keyword Spotting
While having transcripts is fantastic (reading transcripts can be done twice as fast as listening to audio), being able to extract keywords from those transcripts automatically is where it becomes interesting for contact centers due to the large number of calls they deal with every day. Look for a solution that allows you to easily create and enhance your own list of keywords and key phrases.
For example, keywords and phrases like "Ahem," "Let me check," "I have to look," "Let me ask my colleague," and others show agent insecurity. You can add competitor names (and their product names), your product and campaign names (to measure the effectiveness of a new product launch or marketing campaign), and so much more. The sky is the limit here, so easy administration is key!
Another nice feature is the ability to color code keywords. For example, anything related to shipping problems could be marked orange, refund/product returns/replacement requests could be marked red, and positive feedback or referrals could be marked in green. Color coding allows you to easily spot how a call progressed (bad to good, good to bad) as you look at the highlighted keywords in the transcript.
3. Topic Analysis & Categorization
Now that the keywords are automatically extracted, it is time to aggregate them into groups and categorize calls by topics, which could be anything from a reason for the call (e.g., refund request) to the next steps (e.g., flagged for follow-up). The so-called Topic Analysis categorizes conversations into relevant topics.
By doing this, you can analyze call volume and topic trends over time. (For example, you are getting twice as many calls about shipping issues this week, but why?) You can also search for all calls related to a certain topic.
Choose a solution that allows you to customize topic categories for in-depth analysis. Highly customizable topics allow you to define the topics that matter most to your business and align with your organization's specific needs. Watch out! Some vendors will offer "pre-canned topics solutions" that may have built-in topics that are not customizable and have limited utility for your company. Instead, look for a highly flexible solution that will adapt to your needs.
Secondly, choose a solution that allows you to set up and manage topics in an easy, yet controlled fashion. Topics should be seen as living, breathing things that should be constantly improved upon and fine-tuned.
Finally, explore how topics are used in a reporting context.
- Can you visually see a trend over time?
- Within the context of a single call, can you drill down — even to a granular level — to obtain further details?
- Are the topics highlighted? Do they show the associated keywords in the transcript?
- Are they color-coded so it is easy to detect which segment of the call they correlate with visually?
Image: This screenshot shows the number of calls and the call minutes analyzed based on specific topics in MiaRec's topic overview screen. As you can see, the most commonly analyzed topic was Resolution Indicators with 150 calls and 301 minutes.
4. Sentiment Analysis
Customer and agent sentiment analysis is extremely useful in evaluating the performance of a single agent, an entire team, or your overall call center as it lets you gauge how both the customer and the agent feel (positive, negative, or neutral) about the conversation they just had.
Image: This screenshot shows the sentiment analysis screen. Here, you can get an instant feel for what is going on in your customer base and how it changes over time. While there aren't many very positive or negative calls, the majority of those calls were negative or neutral.
The most common approaches are lexicon/rule-based (for basic sentiment categorization), pitch/tonality-type assessments, and machine learning (ML) based sentiment analysis.
- Lexicon- or rule-based sentiment analysis requires you to set up a catalog of keywords and phrases that the algorithm is looking for. However, it is limited to the words you entered, which makes it less reliable. In addition, you need a data engineer to use the expression syntax.
- Many vendors use simple indicators like voice pitch and volume to determine sentiment. But this can often lead to inaccurate results because people can be just as excited (and loud) about something positive as they can be about something negative.
- Machine Learning-based sentiment analysis is the most technologically advanced method of determining sentiment. It is also the most accurate, nuanced, and reliable.
We recommend, therefore, that you opt for a Machine Learning approach.
5. Trend Analysis & Reporting
Reporting capabilities are essential for Voice Analytics. Supervisors should be able to easily and quickly analyze call volume trends per topic over time, e.g., changes in the number of calls related to product returns, shipping problems, and upgrade opportunities from last month to this month. This includes the total number of calls and a trend diagram for each topic for the selected period (days, weeks, months, quarters, or years).
It is a good idea to play around with the reporting and dashboarding to see what you can report on. Can you report on the number of calls containing competitor names or how many times certain key phrases were used? Can you create automatic alerts and notifications should a certain threshold be exceeded? Can you set up recurring reports to be sent out automatically on a particular weekday?
Image: This screenshot gives you a glance into the usefulness of MiaRec reporting. Here you can see the call sentiment across locations or groups. You can get deeper insights in your topics and call volume as well.
"MiaRec allows us to simply and quickly record our agents, listen in real-time to live calls, and lightly integrate targeted search for call recordings from our CRM. We’ve been using it for years, and it has improved and grown over time to become a workhorse in our QA and compliance flows."
Paul McGurn
Telecommunications Architect - LogMeIn
Generative AI-based Auto QA aims to overcome these limitations by automating the process, allowing for a more comprehensive and efficient assessment of interactions.
- AI-Powered Analysis: Auto QA is based on Large Language Models that have been trained on thousands of calls. This extensive training allows generative AI to analyze and evaluate interactions with a high degree of accuracy.
- Comprehensive Coverage: Unlike manual evaluations that might cover a small fraction of interactions, the AI-powered solution offers visibility into 100% of interactions, ensuring a more thorough quality assurance process.
- Unparalleled Accuracy: Auto QA surpasses traditional methods that rely on predefined keywords and key phrases. Generative AI allows for a deeper understanding of customer interactions, resulting in higher accuracy levels.
- Context Understanding: The AI comprehends the intricacies of call conversations and captures the context within which they occur. This means it can provide meaningful insights even in conversations with varied phrasing that couldn't be predicted in advance.
- Zero Configuration: The system comes ready to use, eliminating the need for extensive setup or configuration. This user-friendliness extends to the point where there's no need for data analysts or AI engineers to operate it.
- Integration with Other Analytical Tools: Auto QA integrates seamlessly with Sentiment Analysis and Topical Analysis, enriching the depth of insights obtained from conversations.
If you're considering implementing an Automated Quality Assurance solution, it's essential to understand both its advantages and limitations. MiaRec's Auto QA offers unparalleled accuracy, context understanding, and comprehensive coverage. It's crucial to assess if these features align with your contact center's specific needs. Always consider your unique requirements, budget, and existing infrastructure before deciding.
Image: The screenshot above shows an automatically performed call evaluation by the generative AI powered Auto QA feature based on the contact center's call evaluation criteria.
If you do decide that a generative AI-based Quality Assurance solution is right for you, we recommend that you look for a Conversation Intelligence platform that is flexible and will allow you to combine both manual and automated approaches for the best results possible.
We needed a platform that went beyond basic interaction recording. We wanted a single solution that would provide usable customer service insight. We needed analytics tools to measure performance and customer sentiment. Most of all, the data had to be organized, easily accessed, and easy to understand without having to become experts in analytics."
Jared Jenevein
IT Analyst - Companions & Homemakers
Data redaction is paramount in ensuring the protection of sensitive customer information and maintaining trust. As contact centers often handle vast amounts of personal and confidential data, including financial details, health records, and other Personally Identifiable Information (PII), it's essential to mask or remove this data to prevent potential misuse or breaches.
Implementing data redaction not only safeguards customers' privacy but also helps contact centers comply with strict data protection regulations, thereby avoiding legal repercussions, and upholding their reputation in the industry.
There are three primary automated data redaction methods:
- Numeric Data Redaction focuses on masking or altering numerical values, like credit card or social security numbers. Specific digits are replaced with placeholders to safeguard sensitive details. While this approach is mainly direct and straightforward for numerical data, it comes with a number of disadvantages. For example, it can lead to over-redaction, mistakenly masking non-PII data like customer account numbers, which diminishes the data's analytical value. It's also ineffective for non-numeric entities such as names or brand names.
- Pattern-Matching Data Redaction uses predefined patterns and manually set rules to identify and mask sensitive data. Automated tools process call transcripts, searching for sequences matching these patterns, and triggering redaction when a match is found. This can be tailored to specific organizational needs and is effective with proper configuration, but it requires significant upfront manual configuration. In addition, pattern-matching data redaction is notorious for over- and under-redaction, so consistent and extensive auditing is necessary.
- Machine Learning-Based Data Redaction is the most technically advanced method as it utilizes Machine Learning (ML), a subset of Artificial Intelligence (AI). Instead of relying on manual rules, ML models improve with experience, becoming more accurate as they analyze more data. They can discern context, differentiate between similar data types, and recognize new terms. Named Entity Recognition (NER), an ML-based Natural Language Processing technique, is particularly effective in extracting a wide range of Personally Identifiable Information (PII) from unstructured texts.
ML/NER-based data redaction is not only automatic — which means it is much more scalable as it requires no manual labor — but is also highly accurate as it has a deep contextual understanding and can therefore identify sensitive data much more reliably. It also adapts and improves over time, making it the most effective auto data redaction method available today. It is worth noting that while the most common data redaction scenarios (e.g., redacting payment card details for PCI-DSS compliance) are covered right out of the box, you might need to train your own custom Large Language Model for other use cases, e.g., recognizing and redacting patient information.
Image: The above screenshot shows MiaRec's ML/NER-based data redaction feature called Auto Redaction.
In conclusion, while each method has its strengths, it's crucial to choose the one that best aligns with an organization's specific needs and the data being handled.
We needed a platform that went beyond basic interaction recording. We wanted a single solution that would provide usable customer service insight. We needed analytics tools to measure performance and customer sentiment. Most of all, the data had to be organized, easily accessed, and easy to understand without having to become experts in analytics."
Jared Jenevein
IT Analyst - Companions & Homemakers
Since the rise of generative AI, Conversation Intelligence solution providers are trying to figure out how to utilize this new technology to benefit contact centers. Most vendors offer AI-powered Call Summary features, but over the next months and years, this will be where the most innovation happens. We will update this section regularly to reflect the latest developments.
AI Call Summary
In a fast-paced contact center environment, there are multiple opportunities in a conversation for an agent to miss a piece of information, capture it improperly, or be distracted with note-taking to not fully understand and solve the reason for the call.
Call summaries are crucial. Whether it is auditing compliance for government or insurance entities, agent performance review, trend analysis, product research, etc., it needs to have accurate and complete information.
This is where generative AI comes in. It can be used to automatically generate call summaries, removing the human error factor. Consequently, agents can focus on the caller instead of taking notes, while the AI summarizes the pertinent information from 100% of calls.
When considering a solution that offers call summaries, we recommend that you pay close attention to the accuracy and flexibility of the call summaries. Keep in mind that certain solutions may offer predefined call summary formats, which provide predefined information and may have limited room for customization. Instead, look for a solution that allows highly customizable call summaries, where you can highlight details important to your organization and applicable to different call scenarios. Finally, look for a solution that offers advanced functionality such as Call Categorization and Key Facts.
Image: The screenshot shows MiaRec's Auto Call Summary feature. Using generative AI, MiaRec extracts not only summarizes the call automatically within seconds, by extracts key facts and even makes suggestions for areas of improvements in the call notes.
With the AI doing the heavy lifting of call summaries, the agent spends significantly more time talking to customers instead of note-taking and admin tasks. This increases agent efficiency and reduces the amount of labor needed, which leads to cost savings and an improved customer experience.
We needed a platform that went beyond basic interaction recording. We wanted a single solution that would provide usable customer service insight. We needed analytics tools to measure performance and customer sentiment. Most of all, the data had to be organized, easily accessed, and easy to understand without having to become experts in analytics."
Jared Jenevein
IT Analyst - Companions & Homemakers
We needed a platform that went beyond basic interaction recording. We wanted a single solution that would provide usable customer service insight. We needed analytics tools to measure performance and customer sentiment. Most of all, the data had to be organized, easily accessed, and easy to understand without having to become experts in analytics."
Jared Jenevein
IT Analyst - Companions & Homemakers