This article focuses on practical tips of implementing Your Logic enabled for Salted CX Live Conversations.
Main Loop First
When implementing Your Logic focus on the main conversation events first — opening and closing of the conversations. Open and closing should be present in 100% conversations, plus there is risk of other events happening. Main loop is something that is executed with every customer message.
You should cover these high level parts of the conversation first:
- Opening. Greeting the customer and asking what they want. You need to handle opening for any conversation as you want to make a decision to
- Closing. Opportunity to confirm that the customer request is resolved, collect customer satisfaction and goodbyes. See more in Collect Customer Feedback section.
- Escalation detection. Detection that you want to escalate the request to an agent. This can be extremely broad in the first iterations (ask for help for all conversations after customer mentions any request) and very narrow at the end (do not even escalate when customer asks for an agent).
- Resolution detection. Detecting when the customer request is resolved and when it is time to wrap-up the conversation is essential to gather customer feedback and get confirmation that your bot and agents have really managed to resolve the conversation. The resolution detection can be very easy basic at the beginning (listening for an event that agent engagement is completed and marked as resolved when you expect all conversations to be escalated) to a detection based on the provided responses and the customer answers (when the conversation is handled by a bot).
There are additional scenarios you might want to cover when you let bot handle conversations themselves without escalating everything to an agent:
- Abuse detection. Some customers may be bad actors and try to exploit the chatbot by providing its instructions to favor them even when they would be against policy. You should detect those before you let bot perform actions independently.
- Detection of requests unrelated to your business. You do not want your bot to be used for general purpose requests such as comparing you with your competition, writing long homework assignments, etc. You shoudl detect this and politely respond to the customer that you are able to respond only requests related to your company.
- Smalltalk detection. You might want to tolerate certain level of smalltalk unrelated to your business. You should handle smalltalk according to your company preferences and not ask agents for help with it. Smalltalk may help to build a rapport with the customer so unlike requests unrelated to you business you might want to engage in it a little bit. With each response to a smalltalk you should drive the customer back to an actionable request.
- Detection of escalation. You might want to have escalation detection independent on specific scenario to ensure good customer experience. At the beginning it can be fairly simple based on customers giving thumbs down to the bot reply, customers telling their request is not resolved or giving you poor satisfaction score. Later you can use detection of customer explicit request for an agent, detection of poor customer sentiment and other content based detection.
Measure Key Metrics
Ensure your implementation is successful by measuring important metrics. All conversations that happen in Salted CX live conversations are automatically available for analytics. Ensure you define a helthy targets and keep metrics with those targets.
Example metrics you might want to watch:
Iterate on Handled Scenarios
When you have the main loop ready and you can confidently ask human agents for help start implementing individual scenarios you want to completely resolve without an agent involvement. Also remember that for any impactful action you can still escalate to an agents to let them perform the action itself.
Pick you scenarios based on complexity to resolve and volume of conversations that will not have to go to agents. There is no exact rule for this. You can use Salted dashboards and visualizations to find what are the most common reasons customers contact you and what are the most common outcomes to understand where to focus your energy.
With every implemented scenario measure how you are doing. You might do it by collecting customer feedback, checking feedback on your replies and measure other metrics. You might want initially to check manually at least sample of the conversations. Start with sample of traffic to decrease the blast radius if something goes wrong and move forward if you are happy with the metrics.
You should be able to split the conversation into these categories and keep the number of conversations in those categories within your targets:
- True resolution. Your Logic resolved the customer request. You have it confirmed from the customer or detected by other means (successful transaction) or reviewed by a person.
- No resolution. Your Logic provided an information to the customer but it is not a working solution. We have no response from the customer that it does not work. You should strive to have this low - eg. 5% of conversations.
- False resolution. Your Logic provided wrong information to the customer. You think the customer request is resolved and customer may think the request is resolved. However the false information may negatively impact you or the customer in the future. It is impossible to be perfect (even human agents will provide false information). However you should strive for very small percentage - eg. below 1% of conversations.
- Customer left. Your Logic needs information from the customer to continue but the customer has not provided this information and is not responding. Thus Your Logic cannot resolve the request but it knows it cannot resolve the request. This may happen when the customer resolved the issue in the meantime, it is a low priority for the customer, or they got frustrated and gave up. These are tricky and may require research into what happened and even reaching back to the customer.
- True escalation. Your Logic escalated a conversation that you intended to be escalated (the scenario is not yet supported, customer asking for escalation, etc.). These represent future scenarios you might consider handling better.
- False escalation. Your Logic escalated a conversation that it should be able to handle. This is not a big issue if you have enough agents to handle these escalations but you should incrementally reduce their number.
Additionally to those, also watch additional metrics such as customer satisfaction, engagement time, volume of refunds in transactions, etc.
Conversation Opening
Even when you escalate to resolve all the customer requests use the opening of the conversation to greet the customer and ask them what they want. This saves some time to human agents as they can start responding to the customer right away.
Even simple chatbot not focused on achieving specific tasks can do the following:
- Greet the customer. Be polite and start and fulfill basic customer expectations.
- Manage expectations. Provide information that this is a bot and it may not be able to help with everything and tell customers what might be the expected response times for both the bot and wait time for a free human agent.
- Collect the customer request. Try to shorten the followup human agent engagement by having the customer request in front of their eyes when they open the conversation.
Additionally you can also use buttons to Give Customer Quick Choices to resolving the most common issues.
Help of Human Agents
Your Logic can easily ask human agents to help with the conversation using action Needs Help. It is completely up to Your Logic to decide in which cases it will ask humans for help. In case Your Logic is not available, does not respond in time, or responds in an unsupported format Salted CX will ask agents for help automatically.
Consider these factors when asking for help:
- Be aware of number of available agents. You should rollout Your Logic incrementally to ensure that there is enough human agents as a fallback in case Your Logic does not work as expected, customers ask questions you haven’t though of and similar scenarios. Keep in mind that high percentage of conversations handled automatically with Your Logic naturally leads to fewer agents available over time as they would sit idle most of the time. So the more conversations you handle automatically the more pressure there is not to ask for help as your agents would get overwhelmed.
- Mind the business hours. This is related to the previous point but it deserves a separate point. Some companies may have business hours in which there are less agents or no agents outside of business hours. Using bots that operate 24/7 can increase the customer expectations. So when a human help is needed make sure to manage the customer expectations and tell them they may need to wait.
- Manage handover to agent. Your Logic can perform multiple actions at once. So when you are asking agents for help you can smoothen the transition. You should let the customer know that they may need to wait for a human agent. Optimally you can try to respond to customer request the best you can (even when customer asks for human, or you have lower confidence in your answer) with the disclaimer to the customer. If the conversation is longer you can also give some notes to the agent by creating a note Save Note. The note can summarize the previous conversations, explain the agent why help is needed, etc.
Take Back Control from Human Agents
You Logic stays in the conversation even after it asked for help of a human agents. You can take control back if there is an event that enables you to do so — for example customer clarifies their question, the human agent leaves the conversation and tells the bot can finish it.
Taking back control of the conversation helps to decrease the load on the agents. Your Logic can continue or wrap up the conversation without the agent involvement. If Your Logic needs help it can still ask for help again and agents will be asked for help.
Invite External Agents
The cost of resolving the customer requests grows with the time and resources you have to invest into resolving the request. The simplified view looks like this:
Your Logic can use Invite External Agent to ask people who are not even users in Salted CX to join the conversation and help to resolve the customer requests. When you invite and external agent they will get narrow permission to handle the specific conversation.
Remember that external agents may have different service level policy, may not be that responsive as your agents. It is necessary to set expectations with the external agents so they are aware of the customer expectations. Make sure you also manage customer expectations when involving the external agents.
The following example shows how the Your Logic can respond in case when it is not sure how to answer. In total these are 3 actions. First managing the customer expectations especially due to possible wait time and enable the customer to talk to us in the meantime. Second Your Logic saved a note visible only to agents that sums up the current conversation so was and tells what is expected from the agent. ThirdYour Logic sends the actual invite to the person.
Give Customer Quick Choices
Your Logic can use Ask Question and Ask Ad-Hoc Question to give the customer a set of possible options to move the conversation forward without having to write. This is a great way to make the customer experience better and more efficient.
Using questions with answers has some advantages:
- Some customers prefer clicking through menus rather than chatting with a bot. Sometimes it is easier and faster for customers to pick one of the offered option than describing the request.
- Easy to process the answer. You do not have to use AI to analyze the customer response. You know exactly what the customers clicked. This enables to respond faster to customer clicks, saves money and avoids level of uncertainty that is always associated with analyzing a free customer answer.
The usefulness of menus in chat depend on their clarity and options available to the customer. You should watch how often customer use the menus and try to increase their usage over time. You can use questions that are managed by users directly in Salted CX or generate custom menus on the fly.
Collect Customer Feedback
There are at least two essential questions you should always at the end of the conversation:
- Is the customer request resolved? When you have high confidence that you have resolved the customer request try to gather explicit confirmation. Use Ask Question to get answer to this. If customer answers No you can take is a signal to ask human agents for help.
- How is customer happy with the conversation. After you know that a customer request was resolved ask the customer to give you feedback as soon as possible. Use Ask Question to collect the customer satisfaction feedback. Make sure you
Gathering customer feedback is a great way for quality assurance that you crowdsource to your customers and thus costs you very little time and money. Customer feedback is also available in analytics to help you track how you are doing and identify problematic conversations.
The following flow shows the example conversation wrap up.
Great Good Bad
Act on Live Customer Feedback
Salted CX enables to give customers Thumbs Up/Thumbs Down to the replies they receive from Your Logic, your agents or external agents. Your Logic receives notification about this feedback and can respond accordingly.
Thumbs Up it is a great signal that should increase your confidence you are addressing the customer request and may be a good indicator to ask whether the request was resolved.
Thumbs Down is a signal that you may need to take action. The exact action depends on who received the negative feedback. If it is bot, you might consider it a signal to escalate to an agent. If an external agent receives the feedback you might want to ask your agents for help.
All the feedback received from the customers is also available in analytics and you can use it to find out problematic conversations.
Reply Structure
Recommendations on how to formulate replies sent by Your Logic are very similar to recommendations given to human agent. While using AI to answer the customer requests may cause extra caution you should built-in the caution into the replies in a way that does not feel robotic.
When replying to a customer you should always consider the following:
- Start with direct and short answer. Customers should see what they look for first, not dig it out from the middle of a lengthly response.
- Confirm you are answering the actual customer question. The more complex the question is the you should confirm you are answering what the customer is asking for. Ideally merge the confirmation into the explanation of the answer.
- Provide an explanation to the customer. First this helps you to confirm you are actually answering the question. Secondly it helps to manage customer satisfaction in case they might not like the answer. The explanation should always try to communicate you empathize with the customer and try to be fair.
- Answer potential followup questions. If the original request often leads to follow-ups answer them righ with the reply. You can save the customer some time, increase the customer satisfaction and reduce chance of the customer contacting you in the future.
- Highlight important information. If the explanation is longer, use bold to highlight the most important sections for the customer.
- Offer the next best action the customer. If there is a decision you want from a customer or some action they should take, tell the customers. You can give them menu options to perform these actions or provide useful hyperlinks.
- Educate the customer about self-service. To make customers more independent consider offering them a self-service actions they can take. This decreases their dependency on your customer care and decreases your future load. Make sure you do not sound too pushy and you are acting in the customers’ own interests.
Consider the following exchange example:
The example reply to a customer request is technically correct (contains the correct information) but does not focus on readability. Rather than starting with the response it starts with lengthy confirmation the bot understands the question. While this statement tries to play it safe in case the AI did not manage to extract the customer request intent correctly it makes the reply hard to read, overly wordy and not authentic.
The actual answer “Yes” is answered using a long phrase from a policy (not customer) perspective. It uses factual data and company policy information instead of communicating the impact on the policy and dates on the customer. We also do not offer any next best action for the customer.
Alternate reply to the same customer question:
To refund this trip open the trip page and click the Refund button or just ask me to refund.
Refund to Credits
The example above contains the same information. However it takes into consideration the guidelines. Is starts with direct short answer. It then combines answering potential followup question while confirming it is answering the customer question. It then highlights the information that may be important the customer — the deadline for refunding.
We also use the opportunity to educate the customer how they can refund the trip themselves directly in our application. This is useful to reduce future load on the customer care and also make customers aware of the features of our application. We do not force the user to use the application as they can still ask about the refund in the application. We offer them quick button for the step we would prefer the customer to take.
Unsupported Features
Salted CX may send a various types of events to Your Logic. In the future there might be additional events sent to Your Logic. So you should make it ready to handle unknown event types.
It is not necessary to act on every single one event by taking an action. Even for those events you should respond to Salted CX with No Action. This tells Salted CX that it can continue sending you next events in the conversation.
When you encounter an unknown event you have two choices:
- Take no action. We recommend to respond with no action by default.
- Ask for help. This increases the load on agents. We do NOT recommend using this for unknown events.