This short guide gives an overview of the logical steps on how to make a chatbot. Like elsewhere in life, nothing is set in stone in customer service either. This means that these steps can be taken in a slightly different order or some can be iterated multiple times before actually making a chatbot.
In this article, we’ll take a look at
How to set up chatbot goals
How to define automation cases
How to get training data for AI
How to define solutions flows
How to launch a chabot
Next steps after the launch
How to make a chatbot? How long will it take to make a chatbot? – Mostly, it depends on your team and objectives. Technically, it is possible to make a chatbot in a couple of weeks. Business-wise, it might take longer to get aligned with your team. We have seen that even in large and complex organizations, it is possible to launch a chatbot in a couple of months.
In most cases, before actually proceeding to launch the chatbot, we advise you to find a proper partner with experience and a high level of competence in AI to guide you. We have seen cases when a chatbot does not bring any business value due to a wrong understanding of the process necessary for chatbot implementation. For this case, let’s assume you found a partner, so let’s proceed.
1. Set automation goals
Chatbot automation is a good fit for companies whose customer service gets a considerable amount of repetitive contacts which can be solved following similar solution patterns. The goal of automation might be to reduce costs, meet promised SLAs, detect contacts requiring urgent attention, or to free representatives, so they can deal with contacts that actually require a human touch.
Action points for you
Define the main goal and side goals for an AI chatbot
Define the metrics which support the goal
Assign chatbot owner who coordinates necessary activities and ensures goals are achieved
2. Define customer service cases and solution flows
The next step is to talk to your customer service staff and skim through all relevant data. This could be chats, emails or calls, and its goal is to understand the most popular contact reasons and how those contacts are solved. Of course, chat data will be the most pertinent, but communication in other channels will not be completely irrelevant. Just keep in mind that it might require much more work to go through that data. It’s good to find between 5-20 contact reasons even if some of these might not be ideal candidates. Note that all cases where the customer support agent does something in any interface (web or native), can be automated via integrations. These integrations come either in the form of an API (if these exist or can be built easily), or robotic process automation (RPA).
Next, try to sketch out a simplified solution for each contact reason. The granularity of these sketches should enable you to understand the following:
Can the contact be solved in a standardized way? If not, it’s not a good fit for full automation, but AI might still be beneficial in routing the customer to the correct agent.
Does the contact require agent intervention that cannot be automated? If such intervention is required, the contact is not a good fit for full automation. However, forwarding to the correct agent is still useful. For example, selling a complex product might be a case not worth automating.
Is the contact so generic that you would not know how to provide a solution? If yes, go more granular. For example, “complaint” might mean a lot of things, but “complaint about package arrival time” has a solution.
Do you need to ask clarifying questions to give a solution? If yes, categorize it as a solution requiring business logic. If not, categorize it as a static textual solution. For example, if before providing the solution, you need to clarify whether the person will pick up the package at the store or at the parcel machine, it is counted as business logic.
Are integrations with other systems necessary for providing a solution? If yes, categorize it as a solution requiring integrations. For example, if you need to check info about the customer from a CRM.
Action points for you
Define which customer service cases are good fits for automation
Define which customer service cases have static textual solutions
Define which customer service cases have business logic and no need for integrations
Define which customer service cases have business logic and integration needs
3. Provide examples, validate with data and AI trains itself
AI needs examples to differentiate between customer service cases. You can jump-start the AI training process by writing 10-20 examples per customer service case. Write the examples by thinking about how the customers would write these – this usually means giving up expert in-house vocabulary that a typical customer does not use.
AI does its magic based on manual examples and provides you with the most likely matches from your data. Your job is simply to approve or reject these suggestions to help the AI be even smarter. In most cases, it is enough to provide 30 examples from real data, but the need might grow in case of more complicated cases which are tricky to differentiate. The platform will guide you to label more data if it sees that it is likely to improve the KPIs. If you do not have any preexisting data, do not worry, you can add these later from incoming chats.
By providing examples manually and from real data you will form the training set for AI. While validating with real data, you might discover that these customer service cases are not as popular as you would have expected or that customers are approaching you completely differently or that representatives are solving these in alternate ways. If this happens during validation, you might need to go back to step 2 with such cases.
Action points for you
Provide 10-20 manual examples using the chatbot platform
Provide 30+ examples from real data using the chatbot platform
Validate customer service cases and solutions with real data
4. Finalize solution flows
The solution flows defined in step 2 need to be realized and refined in the tool. It is definitely the easiest to start with those which have static textual solutions, followed by business logic, and do not need integrations. The most complex to execute are those solution flows that have both business logic and integration needs.
Every solution flow should be assigned to an owner, the person who should make sure that the solution flow is correct and up to date – at the beginning, it is sufficient if that person is the chatbot owner, but over time ownership should transfer to the correct person. This is necessary because it is not realistic that the chatbot owner is aware of everything that is going on in the business. For example, it might happen that the package sending policy is changed overnight. In that case, the correct person to be responsible for making changes in solution flows is the person who is responsible for the package sending policy. Additionally, the true business owner also has the power to fix any processes that may reduce or get rid of a customer service case in general.
Once a solution flow has been implemented, it is a good idea to test the solution flows from the end-customer perspective. You and your colleagues can play through different flows and think whether the problem actually gets solved. The domain experts should also review the solution flow from the business logic perspective – just to make sure that the implemented logic is indeed correct and up to date. Finally, if possible, test the solution flow on real customers. With popular cases, this should not be an issue.
Do not be discouraged if you cannot perfectly automate every combination right away. For example, this might be because of some missing integration possibility. It might be acceptable to hand those over to an agent.
With most cases, you will see room for improvement after the solution flow has been deployed. Usually it takes 3-4 iterations to get it right.
Action points for you
Design solution flows
Test solution flows
5. Launch the bot
By going through steps 1-4, you should be ready for the launch. It is advised to launch as soon as these steps are done because you only get real feedback for improvement from real customer interactions. To reduce the risk of potentially causing confusion among your customers, you can consider deploying the chatbot on only a part of the website or for only a particular customer segment. On the other hand, make sure that there are enough customer interactions for getting a representative picture.
From a technical point of view, the launch will be relatively easy. A snippet of code needs to be inserted on the website or app, whereas integrating with the most popular CRM and chat systems is usually relatively painless.
If agents have been doing live chatting before, then going over to an AI chatbot is almost seamless. If that was not the case, it will take a few hours of training. Of course, the first days might be a bit slower, but the process will speed up fast.
Action points for you
Decide if you wish to launch fully or partially
Perform technical works necessary for the launch
Train personnel to use the system
6. Scale-up, manage business ownership, and improve iteratively
Now that you launched the chatbot, you will start getting real data and feedback. This will help you improve existing solution flows and get insights into which solution flows should be implemented next. Titan chatbot software will recommend which customer service cases should be focused on next to get the most business value, but you will always have the option to act differently should you disagree.
If you did not manage to do this in step 4, it is definitely time to get more business support behind the project once results start pouring in. This means assigning business owners to solution flows and developing necessary integration endpoints or RPAs enabling further automation.
Over time you will see more traffic in the chat, greater automation rates in implemented customer service cases, and you should work towards implementing more customer service cases. Do not forget to celebrate these accomplishments with your team.
Action points for you
Iteratively improving solution flows
Implementing new customer service cases
Assigning business ownership if it was not done before
In summary
We have prepared an illustration with all the key points for understanding how to make a chatbot
Do you want to implement a chatbot?
Do you want to implement a chatbot?
Kristjan Jansons
Co-founder, CEO
Kristjan has helped various companies from Japan, Saudi Arabia, and Switzerland to implement the latest AI solutions. Having a strong technological background and understanding of the business processes helps him to understand specific business needs and offer necessary AI solutions for matching these goals.