The following is an approach I’d like to share which I’ve used to translate Customer enquiries into content that can be used by chatbots to more readily and efficiently help other Customers.
This approach is likely documented somewhere out there on the Internet but at the time it evolved organically from data analysis and from the need to create domain-specific responses that would answer Customers’ questions.
Note that it’s manual and labour intensive but if it’s the first time this analysis is being done, the process will provide the opportunity to get familiar with the data before moving to more autonomous methods.
The outputs can also be immensely valuable well beyond the intended purpose and this depth of analysis from my experience is not generally done.
Overview
- Getting the data
- High-level analysis
- Selecting what to analyse
- Lean Six Sigma – Translating Voice of Customer
- Context of Needs
- Pivot table love
- Creating Intents, Training phrases and Responses
1. Getting the data
Customer verbatims are gold… they’re exactly what someone said before it was interpreted, prioritised, synthesised, distilled, framed and packaged up into a Powerpoint presentation!
Themes, insights, priorities… these are great for macro initiatives, but interactions happen at a more personal, one-on-one level and you’re going to provide much more value in an interaction the more relevant you are.
When asking for data, you don’t know what you don’t know… unless you own the data or have access to it you’re not going to be able to see what’s there and you could end up missing out on some really valuable information, so having the chance to explore the whole dataset, getting a view of all the data values, or simply getting small samples cut is important.
Excel… it’s the usual go-to in looking at and working with data but unfortunately, on MAC you can get quickly bogged down with larger files, and given the data can be quite sensitive, uploading and using a cloud service is often not an option either so creating and using a local database can speed things up considerably.
Key learnings
- Get a view of the whole dataset
- Get a small initial cut of data to explore and work with
- Work with a local database
2. High-level analysis
We’re not going to be able to analyse every verbatim, nor will every verbatim necessarily be useful for the purpose, so the next step is to work through how to best go about deciding on a sample to work with.
This will generally be pre-determined by the project or available data values, but often some simple sorting and filtering will quickly provide a good macro view on key areas such as channels, types of enquiries and categories.
Often you’ll get classic long-tail graphs which will be helpful in defining what areas to include in your scope or investigate further.
One approach I like to use in analysis is to think in perspectives
One way to do this is to simply list every stakeholder you have and explore what they would be interested in knowing from the data.
From a data value perspective you can also play one lot of data off against the other, e.g., phone vs digital vs retail channels, desktop vs mobile, etc.
I’ll be looking to cover analysis methods in more detail in another article.
Key learnings
- Don’t expect the data to be perfect
- Data values and labels only go so far in describing the underlying data
- Don’t stop at data values and labels!
3. Selecting what to analyse
Selecting your sample to work with is going to depend on a number of factors, your project requirements, insights from high-level analysis, etc.
If the intended purpose of this activity is to provide content for a chatbot then it makes sense we focus on enquiries through the digital channel, i.e., the channel we intend to launch the chatbot, but what of phone enquiries, should we not also look at those so we can more readily help through the digital channel?
If the intended audience for the chatbot is to help Customers then it makes sense to focus on enquiries, however, there’s just as valuable insight in complaints and why shouldn’t we try and help someone who has a complaint as it could be viewed they have an unresolved enquiry in need of more immediate attention?
The key thing though is that whilst the process is qualitative in nature, it would be ideal to have quantitative measures, so the number of data values you decide upon will go some way in determining the sample size if you want to reach statistical significance in your outputs.
Key learnings
- Chatbots are better suited to handling simple enquiries
- Bigger categories are more likely to contain complex enquiries
- Some channels may include particular skews worth investigating further
4. Lean Six Sigma (LSS) – Translating Voice of Customer
There is a great method in the Lean Six Sigma toolkit for capturing and translating Customer verbatims into requirements (bottom of page) – Six Sigma Institute
Let’s use our own example:
| Verbatim | “I want to know my account balance” |
| Need | Customer wants to know their account balance |
| Requirement | Provide a link to log in |
Here a Customer verbatim is simply translated into a Need and then into a Requirement for the project.
This method is critical in the process as identifying and addressing Needs is at the very heart of a Human-Centred Design approach.
Key learnings
- Use the best method for the task at hand, not because it’s the next thing to do in the methodology or process you subscribe to
- Lean Six Sigma (LSS) provides a great, Human-Centred Design orientated method for capturing and translating the Voice of Customer
5. Context of Needs
Customer Needs can be as simple as in the above example but often are not which is why they’re likely reaching out for help in the first place!
Let’s add to the verbatim:
| Verbatim | “I want to know my account balance because I can’t log in“ |
| Need | Customer wants to know their account balance |
| Requirement | Provide a link to log in |
The Need is still valid but without context, i.e. “because I can’t log in”, simply providing a link to log in is not going to be helpful…
So we need to include Context:
| Verbatim | “I want to know my account balance because I can’t log in” |
| Need | Customer wants to know their account balance |
| Context | Customer can’t log in to their account |
| Requirement | Provide a link to login issues |
Context provides the chatbot with a conversational structure so its responses are relevant, a means to more accurately determine and provide an answer the Customer is looking for and do so in a fashion that is more human-like.
In exploring Customer enquiries sometimes the Need is clear and doesn’t require Context, but often it does so it’s this next level of analysis that may be needed depending on the data.
Key learnings
- Need/Context pairs can assist in more accurately answering an enquiry
- Needs and Contexts can sometimes be answered in isolation
- A structured approach in Excel will set you up for the next step
6. Pivot table love
I’m just going to come out and say it…
I love Excel and pivot tables are awesome!
Some simple sorting and filtering can quickly take a mass of raw data and reveal valuable insights. Pivot tables take this one step further and provide a multi-dimensional view from which even more insights can be gained.
To illustrate this multi-dimensional aspect, the following is a dummy pivot table that displays a list of Needs against related Contexts:
The numbers represent the total number of times a Context arose for a particular Need and quartile lines were used to prioritise work.
This view provides the means to answer Customers’ enquiries more specifically and instead of writing a Requirement we simply write the Response for each Need/Context pairing:
| Need 1 | + | Context A | = | Response 1A |
| Need 1 | + | Context B | = | Response 1B |
| Need 1 | + | Context C | = | Response 1C |
Whilst this method quickly creates a large volume of permutations, not all Needs or Contexts will require this treatment, for example, there may be a general Response for all of Need 2 and its Contexts:
| Need 2 | + | All Contexts | = | Response 2 |
Key learnings
- Pivot tables provide a great dimensional view which is perfect for simple, two-dimensional Need/Context pairings
- This view can then be quickly reused to present prioritisations, visualisations and requirements traceability
7. Creating Intents, Training phrases and Responses
The final step is to take each Response from above and create Training Phrases the chatbot can use to deliver the correct Response.
I found it easier to write the Response first and work backwards to create the Training Phrases even though in some platforms, e.g. Dialogflow, these precede Responses in the Intent creation process:
In creating Training Phrases, I first created a model to standardise how I went about writing these so if these needed to be changed in the future, i.e., in the case of troubleshooting overlapping Intents, the model could first be adjusted and they could then all be addressed in a consistent fashion.
Components of this model included:
- Key extracts from actual Customer enquiries
- Keyword variations, and
- First-person singular pronoun variations
Regarding this last point, focusing on First-person singular pronouns I felt was a great example of the data providing direction. I originally had a different model in mind but found the majority of enquiries were questions that began with “I”, “My”, and to a lesser extent “Why”.
Key learnings
- Write the Response first and work backwards
- Training phrases can include actual extracts from verbatims
- Prepare and work to a model to guide the work
Credits: Feature photo by Adli Wahid on Unsplash
1 Comment
good read Sasha! thanks for sharing