Overview
Cuein AI is a Co-Pilot for Customer Experience Teams. CueIn's Large Language Model Co-Pilot optimizes customer experience to drive better answers, reduced escalations, fewer abandonments, and happier customers.
Cuein AI
- Unifies customer support channels silos across vendors
- Automates mining and analytics of conversational data
- Optimizes bots, human interactions, and knowledge base
Conversation Source
The conversation source page lets you connect your existing data source. Cuein has built-in connectors to various popular data sources where conversation transcripts are stored.
Once connected, you can monitor the status and timestamp when the last time data was ingested.
| Note The conversation source page lets you connect your existing data source. Cuein has built-in connectors to various popular data sources where conversation transcripts are stored. |
Trends
The Trends page lets you discover trending contact patterns from your customer support conversational data. It also shows the top business metrics to understand the health of CX operations.
Business KPIs
- Deflection Rate - the percentage of conversations that are self-served without needing human agent assistance
- Inferred CSAT - estimated CSAT score computed using AI on the fly by analyzing the entire sequence of the conversation
- Conversation Volume - total number of conversation sessions
- Abandonment Rate - the percentage of conversations where users abandon or bail before getting a response from the bot
- Average Handle Time (Bot) - time to resolve user inquiry by the bot
- Average Handle Time (Agent) - time to resolve user inquiry by a human agent
Patterns
These display top contact patterns automatically generated using Large Language Models from support conversational data. The patterns do not rely on internal tagging or taxonomy. The input is the raw transcript. Cuein’s AI models determine where in the conversation, users are expressing their questions or query. It can be in one place but often it's spread out as users add clarification to their initial question. Cuein takes all those into consideration and finds semantically similar questions to group together into patterns in an unsupervised approach.
| Note The approach is custom tuned for every unique dataset, and hence will surface queries unique to a business that otherwise might not bubble up. |
Each contact pattern has:
- Topic label
- Sample user phrasings that contributed to that contact pattern
- Conversation volume for the pattern over time
- Business KPIs segmented by that contact pattern
There is a pattern granularity slider on top. By default, coarse-grained patterns are displayed. By using the slider you can also look into the fine-grained patterns.
Coarse: Produces fewer, larger patterns for a broad overview of conversational data.
Fine: Produces more, smaller clusters for detailed analysis of conversational data.
The choice between coarse or fine depends on your specific use case and the level of detail required.
| Note We recommend starting with coarse patterns and exploring finer patterns as a more advanced step. |
Additionally, you can drill into a specific conversation by clicking on the user phrasing in each pattern. This is a great way to review sample conversations to get a good idea of the contact pattern.
Lastly, you can filter the patterns and KPIs by any dimension or segment using All filters on the top right.
Statistics
Use the statistics section to double-click on specific contact patterns. Let’s take an example to explain this better. Assume you are a food delivery service provider and need an answer to the following question:
For Food Delivery Estimate contact pattern, what is the most common follow-up query and what percentage of times does that happen? What fraction of the time the chatbot is confused or unable to understand the user query? How do I drill down into a few of those sample conversations where the chatbot got confused?
To get answers to these questions, we need to go deeper into a specific contact pattern by clicking on more statistics.
Statistics provides additional slicing and dicing by various segments and deeper information per contact pattern. These include:
- Deflection rate trend over time
- Conversation volume over time
- Top multi-intent flows for this contact pattern - This visualizes the flow of top follow-up queries. For conversations with multiple queries, this depicts the contact patterns following the primary contact pattern. In the screenshot below, Late Delivery is the most common follow-up query and it happens 24.4% of the time.
- Distribution of the contact pattern by emotion, inferred CSAT, turn count, duration, location, conversation data source, and confusion.
To answer the question of what fraction of the time the chatbot is confused, you can use the volume of conversation by confusion chart in the statistics section. All the charts are clickable and you can click on a chart to search for conversations meeting certain criteria. For example, to find conversations where the chatbot got confused, you can click on the confused = true slice of the chart and you are presented with the following search option:
| Note There are scenarios, where there might be no multi-intent flows. This is normal as there are too few to report. |
| Note Similar to the Patterns page, you can use All filters on the top right to filter the statistics by several dimensions - source, turn, duration, deflection, speaker, location and tenant. |
Search
Search is a powerful interface to perform both semantic as well as keyword-based searches across all your conversation support data across channels and sources.
Let’s take an example to explain this better. For the food delivery service provider, you need to find utterances (messages in a conversation), where users ask questions similar to food delivery issues, and where they ultimately abandon the conversation.
To get the answer to this question, you can click on the search icon in the navigation bar. You can just type in food delivery issues in the search bar at the top. Since we are interested in user utterances, we can open the All filters drawer on the top right and select speaker = user. Additionally, since we are interested in abandoned conversations, we select abandoned = true in the filters.
This brings up all the relevant user messages when the user abandoned the conversation.
As in the Patterns page, clicking on any of the utterances opens up the entire conversation which you can also review and audit.
| Note Filters selected are displayed under the search bar as well as with a badge on the All filters button showing how many filters are selected. You can easily remove a filter by hitting the x next to it. |
| Note You can also do an exact match search by putting the search query in double quotations such as “food delivery issues”. |
| Note If you want to export messages and conversations from this search interface, you can just select them using the checkbox next to it. When you do that there is an option at the bottom to export it in a CSV. |
Conversation Flow
The conversation flow visualizes the process map in the conversation workflow. A process map is a flow diagram displaying the different unique conversation paths (variants). Each variant has steps/turns. You can quickly identify the high ROI variants, and zoom in to see the details (more granular insights). You can filter this by similar dimensions as on the search and patterns page.
Conversation Flow Taxonomy
The following table provides definitions of all the steps that can be displayed in the conversation flow:
|
Category |
Description |
|
ABANDONED |
Users abandon or leave the conversation before a resolution is provided by the bot. |
|
AGENT_TRANSFER |
Bot initiates a transfer to a human agent. Or a human agent initiates transfer to another human agent. E.g. Let's see if any agents are currently available to assist you... |
|
AGENT_UNAVAILABLE |
The bot states that no agents are available to transfer the conversation. E.g. Unfortunately, no one on our Customer Service team is available at this time. |
|
CONFUSION |
The bot is confused about the last user utterance. E.g. sorry, I did not understand that. |
|
DATA_COLLECTION |
Bot/Agent collects data or information (also called a slot) to be able to process the user question and provide a response E.g. Do you have the order number there for that or can I take your email address? My email is [EMAIL_ADDRESS] |
|
ESCALATION |
User requests to speak to a live human agent or a supervisor/manager E.g. agent please Chat with Customer Service |
|
INTENT |
The user asks the question or shares the reason for their contact. E.g. When should I expect my food order to arrive? |
|
NO |
The user answers with a ‘no’ to a question. E.g. No No, thank you |
|
RESOLVED? |
Agent/Bot following up after providing a response/resolution to see if that resolved the user issue. E.g. Daniel did that answer your question? |
|
RESPONSE |
Agent/Bot provides a response to the user’s question. E.g. I checked the delivery route and the driver assignments. Your driver might have been assigned to another delivery before you. Sometimes, drivers are managing multiple orders on their route. Please be informed that your food still should arrive within the estimated time. |
|
YES |
The user answers with a ‘yes’ to a question. E.g. Yes Yes, that's correct. |
Conversation Flow Controls
There are a few ways you can navigate conversation flows.
Flow Selection
Flow selection can be used to change the primary intent of the conversation flow. You can use this to investigate a different conversation flow. Flow selector is accessed by clicking on the Flow name in the Conversation Flow header page:
Once clicked, the Flow Filter panel opens up. Here you can select a different Intent in the drop-down. Also, you can change the Flow Type.
Flow Type
Most Common Flow - For a specific user intent, the top dialog paths they go through including follow-up intents. The starting intent is the primary intent and is represented with a shaded UX. In the example below, Food delivery estimate is the primary intent whereas Driver going wrong way is a follow-up intent.
Most Common Intent Flow - Same idea as Most Common Flow but restricted to a single intent. This helps you to diagnose a single intent more closely without getting lost in the noise of follow-up intents.
Multi-Intent Flow - Here you can specifically diagnose a certain multi-intent dialog path. E.g. the paths where Food delivery estimate intent is followed specifically by Late delivery intent.
Filters
As in the Trends and Search page, you can filter the conversation flow by any dimension or segment using the All filters on the top right.
Zoom Control
As in the Trends and Search page, you can filter the conversation flow by any dimension or segment using the All filters on the top right.
By default, the top 2 dialog paths or flows are displayed. To review more variations in the flow, you can use the zoom control at the bottom left. The following is the mapping of how many variations are displayed at each zoom level:
|
Zoom |
Number of variants |
|
1x |
2 |
|
2x |
6 |
|
3x |
10 |
Search Conversation Flows
This extends the search capability to conversation flows. Let's say you want to answer the following question:
Find conversation flows, where users ask questions on food delivery issues
To get an answer to this question, click on the search icon in the navigation bar on the Conversation Flow page. The flows are automatically filtered and will now only show the dialog paths where food delivery issues are mentioned
Bot Insights
The insights panel is displayed by clicking on any conversation flow.
Bot insights are actionable insights for the human-bot span of a conversation flow. It provides further details on that flow including the following:
Unhandled Utterances - These are user questions or intents which the bot did not handle. Common scenarios include the bot not understanding what the user meant (E.g. bot is confused or asks the user to rephrase) or the conversation being unresolved and escalated to a human agent.
Unhandled utterances also include 3 representative sample utterances resulting in this breaking point or friction.
| Note There is a feature in the near-term roadmap that will let you export representative sample utterances which can then be used directly to train the intent classification model for your chatbot. This will lower the effort needed to improve an existing intent as well as bootstrap implementing a new intent. |
Successful Utterances - These are user questions or intents which the bot successfully handled - no confusion detected and no agent escalation. Successful utterances also include 3 representative sample utterances resulting in this successful behavior.
Slots - These are data or information collected from the user by the bot which is essential for the chatbot's operation and fulfilling user requests.
Agent Insights
These are actionable insights for the human-agent span of a conversation flow. It provides further details on that flow including the following:
Agent Actions - These are the responses offered by the human agent to resolve the user query aggregated and summarized across a large number of conversations. Cuein looks at hundreds of agent responses that resolve the user query, groups the ones which are similar enough, and uses generative AI to generate the best comprehensive answer summarized as an insight.
| Note There is a feature in the near-term roadmap that will let you export generated knowledge articles based on hundreds of similar user questions and agent answers. This will help evolve your knowledge base and enhance your bot response. |
Agent Slots - These are data or information collected from the user by the human agent which is essential for the agent to resolve user queries and fulfill user requests.
Intent Utterances - These are user questions or intents expressed to the human agent. It includes 3 representative sample utterances to provide a clear understanding of user questions.
Sample Transcripts
This section displays a maximum of 10 sample raw transcripts that conform to a specific conversation flow path. This is a great way to see some examples to drill down and narrate a story around the conversation flow.
| Note There can be scenarios, where there will be less than 10 transcripts matching a specific conversation flow highlighted. |