Getting started with the new Watson Assistant part II: refine your assistant

We’ve made refining your assistant easier and faster than ever. Now learn how to multiply its abilities with additional actions and advanced features.

By | 11 minute read | December 20, 2021

Before we go into all the ways you’ll be able expand on your first assistant, take a moment to congratulate yourself! You did it, and what’s more, you finished in record time. Who would have thought that a virtual assistant capable of helping customers pay their bills or reach a live agent could be launched in under 30 minutes?

Let’s recap what you learned in the first part of our getting started guide:

The new Watson Assistant revolves around the concept of actions, which simplifies many of the complexities in matching customer intents with the way your chatbot responds. To define it once more, an action is “a problem or task that your customer wants to resolve.”

For the first example bot, we focused on the customer action “I want to pay my cable bill.” But actions can cover any number of user problems, such as help desk answers (like what to do if they lose their phone), to resolving errors in invoicing, or correcting orders. There is no end to the possibilities.

In Part I, we introduced you to your assistant’s first home page. This is an interactive task list we created as your personal launching pad to make sure you get a valid, functioning MVP off the ground as quickly as possible.

Here, you:

  1. Named your assistant
  2. Learned more about Watson Assistant
  3. Created your first action
  4. Added your first example phrase
  5. Created the steps in your conversations
  6. Added at least 5 examples to help train your bot

You also learned how to preview and share your first assistant so that you can immediately see it in a real-world context, as well as get your team experimenting with it and contributing to its success.

So now that your rocket is off and the smoke has cleared at the launch site, let’s take a closer look at the payload your assistant is carrying with it right out of the box.

Welcome to the new Watson Assistant dashboard

In your new assistant’s dashboard, we take you directly from the launch pad to mission control. While your introductory journey was aimed at learning the basics and ease of execution, this new home is equipped with all the features you’ll need to evolve your assistant over its lifetime.

The most important starting tasks here are:

  1. Refine your assistant
  2. Test your assistant
  3. Deploy your assistant

As well as all the tools you need to:

  1. Add actions at will
  2. Preview your assistant
  3. Access the draft environment
  4. Publish your assistant
  5. Access the live environment
  6. Analyze performance
  7. Achieve Watson Assistant mastery

Refining your assistant

In this article we’ll focus on the task list below:

Step 1: Customize your assistant’s greeting (2 min)

The first step in refining your assistant is a simple personalization that allows you to customize how your assistant greets customers. The preloaded greeting is “Welcome, how can I assist you?” But you can transform this message in several ways to make sure customers take note.

Here’s some advice:

1) Give your assistant a name that reflects your business

2) Choose a greeting that describes your assistant’s purpose

3) Settle on a greeting that satisfies any number of customer scenarios. For our purposes:

Hi, I’m Telly — Need help with your service today?

To customize your greeting, you’ll need to visit your actions menu. Click Set by assistant, then on Greet customer, and finally into the first Conversation steps box, which will prompt you to edit away.

Once you’ve given your assistant that slice of charm that your customers will love at first sight, you can exit and go back to the new home screen.

Step 2: Lights, camera… more actions! (3 min)

Your assistant now has a name and a purpose. It can satisfy a common customer problem (or action) based on three possible requests, and two solutions (see Part I). Now what else do you want it to do?

In our example case, data shows that your customers not only want to pay their bills when they contact you, but also want information on bundling two or more services together. Bingo! We’ve got your next action ready to go.

Go to your actions menu once more, but this time stay on the Created by you tab. Click the big blue New action button in the top right corner and type the name of what you want to resolve from the customers point of view. Click Save.


Since your customers can frame their request any number of ways, you’ll want to flesh out the kinds of language your assistant can expect from them.

To do this, click the arrow in the corner of the Customer starts with box and add five alternate phrases to help train your assistant’s NLP model. This is a good start, but you can always add more.

Just like your first action, let’s define a couple of responses, since we know customers will fall into one of two categories here. Click on Conversation steps number 1, type in the assistant’s response (something like “I can help you with that! Do you want to:”), then select Edit response and type in two options: Compare pricing and Bundle now.

In the second and third steps, you’ll want to define conditions and assistant responses for both above options, which will again direct your customer either to the information they seek, or directly to an agent, who will make the changes to their account with voice confirmation and consent.


Nice work! Your assistant is now able to resolve two separate customer needs, proving its value to your business.

Step 3: Build out subactions (2 min)

Some actions will be necessarily dependent on other actions. These are called subactions. The best way to think of subactions is as a common need your customers may have at diverse points in their conversations and across conversational topics.

To make use of subactions, just create a separate action in your actions menu in the same way you have always done. The difference comes when you call them into action.

Pro tip: Subactions differ from topic changes. In a subaction, your customer is required to provide information that will be used later in the conversation. Topic changes are now handled automatically through NLP in the new Watson Assistant.

In our example scenario, the customer may decide to browse plans/offers or have already made up their mind based on promotional materials they have received. If they choose Compare pricing, they will be directed to your pricing/promotions landing page. If they want to Bundle now, they will connect with an agent to finish the job.

To ensure this transition runs as smoothly and quickly as possible, we will rely on a subaction that prompts the customer to enter their account number. After creating the specific action Provide customer account number, you can easily tie it into this and other conversations.

Open the first step in your Add service action and select Go to another action under the And then menu.

This will open a pop-up menu from which you can select the subaction of your choice from a list of all the actions you have created. Click Apply.

The new conversation with the subaction should look like the following:

Congratulations! Your second action is now able to identify your customer and direct them to two ways of reaching their goal, based on where they are in the flow.

Step 4: What to do when no action matches (2 min)

Your assistant is looking great and is ready to support your customers with multiple tasks. But what happens when customers use language your assistant isn’t trained on yet, or they want to do something that is still outside its scope? This is where Retry comes in, and it is very simple to set up.

Head back to the Actions menu and make sure you have Set by assistant selected to show those that are preloaded with your assistant. Then click No action matches as seen below.

Once inside, you can customize your assistant’s response just like any other action. A good strategy to start out with is to state clearly what your assistant can and can’t do. Don’t be shy, your customers will thank you for saving them time!

When your assistant hasn’t been trained on a specific subject, it will automatically offer a list of actions to choose from. But you should still plan for additional requests, so you can get your customers back to what the assistant can do.

To fit this purpose, you can type any number of customer requests into the Additional training examples field. In this case let’s write I want to go paperless:

Then head over to Conversation steps and click the first box to edit the preloaded response as follows:

And finally, here’s what this example conversation would look like. Remember, you can add all the unique scenarios you want to prompt this response.

Step 5: Designating a fallback action to save the day (2 min)

If the assistant can’t offer the customer their desired solution, the customer may rely on your assistant’s built-in fallback action. The most basic is for when the customer asks outright for live agent support. You can find fallback actions in the Set by assistant actions menu.

Your assistant comes preinstalled with three situations. In addition to Agent requested, there are also entries for Step validation failed for when the assistant cannot understand the customer’s request, and No action matches if you decide to forgo retry and skip directly to live agent support in your flow.

You can customize these responses using conditions like any other action, or you can provide different responses based on agent availability. There is even a field to provide case background directly to the agent once you set up your service desk integration.

Here’s how a simple fallback looks in the assistant preview:

Step 6: Basic troubleshooting with step validation and debugging (4 min)

Your assistant has come a long way, so we recommend that you take a moment to perform step validation, to ensure it is firing correctly and functioning as it should. Think of this as the first in a series of short audits you’ll perform as you get closer to sharing your MVP.

Very little can go wrong if you follow the above steps. But if your assistant is asking for clarification on what your customer wants to do, or calling on the fallback action when it shouldn’t, this is the time for any troubleshooting.

For more complex actions and conversations, your new assistant is also equipped with a debug mode. Just click the ladybug icon at the top left of the Preview panel to turn it on.

This will provide a confidence score for each of the assistant’s possible responses. You can navigate directly to the step inside each action that is triggering it to make adjustments on the fly.

In this example, you can see that the assistant has chosen the retry option, which is correct unless you want your bot handing off to agents all the time. However, if this were not the case, you could use the confidence score to correctly analyze customer intent and implement the fallback action we created a few minutes ago.

The confidence score is low here, but your assistant has been able to connect “real person” and “live agent” entirely on its own without any input on our part. Wow!

Step 7: Asking clarifying questions (2 min)

Before closing the chapter on refining your assistant, let’s look at one last detail you may want to fine-tune. As mentioned, your assistant will offer default choices to your customers, if it hasn’t been trained on an action and you haven’t singled one out for the fallback. See below:

You can make sure this option is turned on by clicking Set up your assistant to ask clarifying questions on the Refine your assistant checklist (the last mandatory item), and make sure the toggle is switched to on. Later, you can also customize the language in the Assistant says and Label for a fallback choice fields if necessary. For our example scenario, we are ready to go.

Optional steps: add depth to what your assistant can do

The remaining two items on the task list are labelled optional. Feel free to experiment with Create additional actions to flesh out your assistant. We certainly did! Some examples for Telly, National Telco’s virtual assistant, include “I want to add a landline” and “Is my promo code valid?”

Finally, shop your new assistant for feedback! Make sure to run through its amazing abilities with your colleagues: not only to share the love, but to get the extra synergy and insight that will make your point-of-contact a shooting star, whether you use this example assistant or you build your own.


You’ve worked through two important steps toward getting your first assistant live. You spent under 30 minutes setting up your basic assistant on Getting Started Part I, and another 30 minutes expanding on it in Getting Started Part II. You now have a trusted ally that can handle multiple customer requests, and you’ve streamlined important aspects of your customer service flow.

What’s next? We’ve got two more guides for you in our Getting Started series. The next installment is dedicated to testing and deploying your assistant. Read on!