Intro

Here’s Part II of the Watson Assistant blog series on Building a Virtual Assistant Your Customers Want to Talk To. In this three part series we show how Watson Assistant’s web chat UI and service desk integrations help you build a virtual assistant that your customers actually want to talk to resolve their problems.

In part one we discussed how to build customer trust and drive engagement. In part two we demonstrate how to ensure your customers don’t hit frustrating dead ends.

Virtual Assistant Dead Ends

Most people have little faith that virtual assistants will resolve their problems. Why? Because virtual assistants that they’ve tried in the past lead them to dead ends. Maybe you can relate. Hitting a dead end looks like this:

Or this:

Or this:

If you’ve had one of these experiences, best case scenario you got annoyed before picking up the phone to call and wait for a live agent on phone. Worst case scenario you took a screen shot and posted it to social media where the New York Times used it as an example in a column on what’s wrong with customers service in the digital age.

Okay, maybe that’s a dramatic worst case — but chances are you lost faith in the virtual assistant and did not return the next time you had a problem to resolve.

No More Dead Ends

The problem is that virtual assistants, like humans, can’t understand everything, but unlike humans, we expect them to. We train virtual assistants with data, but we don’t teach them to seek clarification, provide alternatives, or get help when they are having a hard time resolving a customer’s problem — until now.

With Watson Assistant’s web chat user interface and service desk integrations, you can ensure your customers never hit a dead end by teaching your assistant to behave more like a human would when they’re having a hard time helping your customers.

Clarify When Confused

When people aren’t sure what someone else is asking them, they ask the other person to clarify their meaning before responding. Watson Assistant’s disambiguation feature teaches your assistant to do the same.

With disambiguation your assistant acknowledges when it is confused between similar meanings and will prompt your customers to clarify. This way your assistant does not risk responding incorrectly with a dead end.

Provide Alternatives When Wrong

A common curtesy in customer service is to provide alternative options when the initial option does not seem to be what the customers is looking for. Watson Assistant suggestions feature now fills this role.

If your assistant provides the wrong response, suggestions provides your customers with alternative paths forward. This way, even when your assistant responds incorrectly, your customers can find the response they’re looking for.

Suggestions are always there for when your customers to access when they are not getting the response they want. As well, your assistant will nudge your customers to consider suggestions when it detects that it’s not providing the right response.

Seek Help When All Else Fails

Even the best customer service representatives can’t solve every customer problem on their own. Sometimes it’s necessary to get help. When it’s clear the customer is losing patience, good customer service representatives will find someone else who has the expertise to help.

With Watson Assistant if your assistant can’t find what your customers need by clarifying with disambiguation or by providing alternative options with suggestions, it will connect them to a live agent who can. Your assistant will detect when it doesn’t seem to be getting your customers what they want, and before your customers give up, it will connect them to a live agent. Once again — dead end avoided.

We have all kinds of plans for improving the Watson Assistant web chat user interface and service desk integrations to ensure your customers are delighted to chat with your virtual assistant. Follow our release notes to follow the latest and greatest Watson Assistant features and read our documentation to understand how to configure your assistant to use suggestions and disambiguation.

Thanks for reading and stay tuned for part three of Building a Virtual Assistant your Customers Want to Talk To!

Building a Virtual Assistant that Your Customers Want to Talk to (Part 2) was originally published in IBM watsonx Assistant on Medium, where people are continuing the conversation by highlighting and responding to this story.

Most companies have a record of every end-customer inquiry within their service desk tool(s), yet when they train an assistant (aka chatbot) to automate answers to certain questions, they’re often just guessing at what customer issues to handle. This is super painful and draws out the time it takes to get a compelling assistant in front of customers.

And that’s exactly why we’re launching Intent Recommendations! This feature lets you get up and running with new topics (aka intents) in your assistant really quickly — all based on the raw logs from your existing customer service operation. Watson combs through all of your users’ messages, groups them according to topic, and then prioritizes the highest impact clusters (so that you know where to start building).

In addition faster training time, Intent Recommendations will also help you create a much more effective assistant in the long run. This is because real-world requests lead to a much more accurate system than do employee manufactured training sentences or sample training data.

Here’s How it Works…

Step 1: Import a File

Gather your customer requests into a CSV and then import the CSV into Watson Assistant.

Note: The CSV should only contain requests from the customer (no agent responses) and each request should be in its own cell in a single column. Once your CSV is ready, find the “Get Recommendations” button on the Watson Assistant intents page and import the CSV from your device.

Step 2: Add Recommended Intents

After Watson is done grouping the customer requests, it will display the four most recommended topics on the intents page. You can view all recommended topics ranked by most recommended by clicking “view all recommendations”.

Choose a topic and review the requests inside of it before adding it to your assistant’s training. You can remove requests you don’t think are a good fit, and when you’re happy with all the requests inside of a topic, you can add them as a new intent or add them to an existing intent.

Step 3 Add Recommended User Example

After you’ve added a topic, you can continue training your assistant on that topic with Watson’s help. Simply click “Show recommendations” within the intent page and Watson will continue recommending similar requests to add to the intent — this helps to further strengthen how well your assistant can cover this specific topic. The more examples, the better!

Getting Started

Intent Recommendations are available now on Watson Assistant’s Plus and Premium plans! To get started on plus or premium, visit the Watson Assistant pricing page. To learn more about how Watson Assistant’s Intent Recommendations work, see our product documentation.

Train Your Assistant 10X Faster was originally published in IBM watsonx Assistant on Medium, where people are continuing the conversation by highlighting and responding to this story.

So, you built a virtual assistant, released it to users, and you’re excited to watch as its conversation logs grow. Now you have all sorts of real user conversations that you can use to improve your next version of the assistant. However, you face a common problem: How can I use this large number of conversations to improve my virtual assistant efficiently?

[TLDR: IBM is providing two Jupyter notebooks to help you continuously improve your virtual assistant. Learn more here.]

When you first release an assistant, low usage volumes make it possible to inspect every user conversation to find problems and identify areas for improvement. But the manual inspection approach does not scale as conversation volumes grow. The Watson Assistant team built two Jupyter notebooks that can help you maximize the training power of your conversations in less time and with fewer resources. These Jupyter notebooks contain pre-built Python code that someone with minimal coding experience can run to generate helpful analytics from your data. The results are displayed in visualizations that make the analysis easier to comprehend.

The Watson Assistant Measure Notebook focuses on the overall performance of the virtual assistant. The Watson Assistant Effectiveness Notebook identifies problem areas and patterns that highlight areas of your assistant that need to be improved.

Watson Assistant Measure Notebook

The Watson Assistant Measure Notebook gives you an understanding of how well your assistant is performing. It helps you decide whether to focus on improving your assistant’s coverage or its effectiveness. Coverage measures the portion of total user queries that your assistant attempts to answer. Effectiveness measures how well your assistant handles the queries it answers.

Analysis of effectiveness and coverage should complement each other. If your assistant is answering questions incorrectly, you can improve it by focusing on improving effectiveness. If your assistant is not answering enough questions, you can improve it by focusing on expanding coverage. The Watson Assistant Measure Notebook provides the flexibility to customize how you define coverage and effectiveness for your assistant.

Watson Assistant Effectiveness Notebook

The Watson Assistant Effectiveness Notebook helps you understand exactly what areas to focus on to improve your assistant’s effectiveness. It summarizes the intents that perform the worst, the pairs of intents that your assistant has trouble distinguishing between, and the dialog nodes and entities that commonly cause ineffective conversations. These summaries show you where to focus your improvement efforts, and what you need to do to get better results.

Get Started

These notebooks have the potential to drastically reduce the time and maximize the impact of your improvement effort. We encourage you to use these notebooks throughout the improvement lifecycle. As you have more and more conversations to sift through, the notebooks make it easier to optimize what you can learn from them.

To get started, you can run both notebooks for free in Watson Studio, or locally by using stand-alone Python code. To run them in Watson Studio, follow these instructions. To run them locally, get the notebooks from this IBM GitHub Repository. You can learn more about how to improve your assistant by reading the continuous improvement best practices.

Resources

Technical documentation on Measure and Effectiveness Notebooks

Stand alone Measure and Effectiveness Notebooks

Watson Assistant Measure Notebook in Watson Studio

Watson Assistant Effectiveness Notebook

Watson Assistant Continuous Improvement Best Practices

Continuously Improve Your Watson Assistant with Jupyter Notebooks was originally published in IBM watsonx Assistant on Medium, where people are continuing the conversation by highlighting and responding to this story.