Chatbots can create competitive advantage. Ignore them at your peril. Here, we apply a #SOSTAC ®  Plan for developing your own chatbot.  Written by PR Smith and Tom Sickert.

 

 

robot chat bots
  • Facebook
  • Twitter
  • LinkedIn

Image by Parker_West from Pixabay

SOSTAC(r) Planning Framework
  • Facebook
  • Twitter
  • LinkedIn

SOSTAC ® Planning Framework www.sostac.org

SITUATION ANALYSIS

Hyper-competition is here to stay and chatbots have a role to play. Data, AI, and chatbots in particular can create a new competitive advantage. Ignore these at your peril.

AI-driven chatbots can boost the CX (customer experience), strengthen existing customer relationships, reach new prospects, screen enquiries, identify best prospects, give them personalised answers/services instantaneously, convert to sales and thereafter be used to nurture stronger (potentially , lifetime, relationships).  Ignore chatbots at your peril.

Typical customer service staff feedback reveals: “We have all this excellent information on our public website – product description, prices, delivery times & costs, return policies… but still we get countless calls and emails about these.” It appears that many people are just not willing to work their way through websites, searching, scrolling and hoping to find a solution.

Chat Bots are creating a gap in the market for better service
  • Facebook
  • Twitter
  • LinkedIn

Should you have a Bot?

 

All customers have a ‘job to be done’ (Christensen et al 2016) when they visit a website (or an app). They want to find product information, check a price, read reviews, buy a product, be entertained, informed etc. Customers just want, access to the information or experience as quickly as possible (with or without a bot).

Chatbots or Humans - survey: what do people want?
  • Facebook
  • Twitter
  • LinkedIn

Chatbots or Humans – survey: what do people want?

 

Would anyone prefer to queue in a bank to withdraw money from a human or queue for an faceless automated cash withdrawal machine in the wall?

Bots present an opportunity to improve the relationship with both existing and potential new customers. This strengthens, what are arguably, your two greatest assets today: your brand, and your customer data in a new AI-supported world of chatbots. Let us apply this to a fictitious washing machine company.

OBJECTIVES

Be clear about why you want a chatbot? ‘Because everyone else has one’ is not a good enough answer. ‘Reduce costs’ is a popular answer but misses the real opportunity. The ultimate answer is to help customers to have a better CX (customer experience) and also identify your best potential lifetime customers. An AI-driven chatbot can instantaneously answer product questions, share advice, book appointments (for salespeople or technicians), take orders, trigger a follow-up onboarding series of messages in a personalised way 24/7/365.

All of these can, and should, be quantified objectives – in fact, SMART Objectives e.g. Boost CX Satisfaction Scores from 50% to 70 to 90% in years 1, 2 & 3 (or in Q1, Q2 and Q3?); Boost Net Promoter Scores (likelihood to recommend your service to a friend from 10 to 30 to 50); Reduce time taken for the visitor to purchase (reducing these times makes customers happy).

Cost-saving operational objectives are popular e.g. To reduce the number of calls/emails handled by 25% in the first 12 months.

More specific MVP (Minimum Viable Product) objectives can also be set. e.g.  the chatbot must help customers to:

  • Find serial numbers and product names for all units produced by the company;
  • Solving the top ten common problems – using images, links or text based on customer input;
  • Create a service ticket for ALL ENQUIRIES (which includes capturing the customer phone number) so the helpdesk can call and resolve the customer’s issue via phone.

Be very clear about why you want an AI-Driven ChatBot. Think about how chatbots might help your business even more in, say, 3 years from now?    

Is ignoring chatbot potential to save money (and boost CX) like either  burying your head in the ground

Sculpture of someone burying their head in the ground
  • Facebook
  • Twitter
  • LinkedIn

Burying your head in the ground doesn’t solve any problems nor exploit any opportunities

or like throwing money down the toilet as you pay for slower, less personalised, human customer service?

Money thrown down the toilet
  • Facebook
  • Twitter
  • LinkedIn

Wasting Money

STRATEGY

Stage 1: Build a pilot AI ChatBot for Brand X washing machine website  –  aimed at helping customers find what washing machine is best for them, in a personalised, friendly and reassuring way. The chatbot dialogue must be knowledgeable yet friendly (to match the brand personality). Helping the customer in an informal way, yet demonstrating common sense knowledge (without jargon).

The chatbot must at all times support the values and the purpose of the overall business.   Environmentalism is an important issue for our customers, so and useful green guidance and tips should be offered where relevant and whenever the customer expresses interest (or wherever interest is detected e.g. if a visitor watches any of our green content e.g. ‘3 Tops Tips to Save Energy‘ video).

Data collection is critical to the long-term success of the AI-Supported chatbot and the business overall. Customer preferences, interests, demographics and other data feed into each customer profile, which in turn helps to find correlations to further improve both specially tailored offers and new products in the future.

Stage 2: Roll out to all other white good product range (e.g. washing machines,  dishwashers and microwaves) within 18 months and increasing NPS scores from 30 to 50 (as listed in the objectives).

Chess set (represents strategy
  • Facebook
  • Twitter
  • LinkedIn

Strategy

TACTICS

There are several tactical choices available when developing a chatbot .

  1. Out-Of-The-Box Solutions, provided by vendors that have predefined models and features and functionalities, that can be customized based on your requirements. These solutions as well are powered (depending on your budget) by high performance AI solutions and features and functionalities. The provider of the chatbot solution can help you to assess your needs and find the optimal solution. There are now many chatbot companies including: Ada, AWS, Botsify, Chatfuel, Hubspot, Liveperson, Mobile Monkey, Microsoft’s Bot Framework and Cognitive Services

 

  1. In-house Solution Created by ‘Citizen Developers’. These require NO code or LOW code experience and can be created by anybody who is able to create email rules in Microsoft Outlook. Yes, it is (mostly) that simple – these drag & drop (communication flow) solutions are offered by Microsoft and AMAZON alike. It gives you not too many options to customize and apply specific functionalities, but it sure is enough to for professional use. NB Citizen Development – no technical and programming skills needed.

 

  1. Developed From Scratch – by coders and other IT professionals, in collaboration with your subject matter experts. These chatbots are powered either by custom AI with sophisticated algorithms and enhanced features and functionalities.

 

Sometimes out-of-the-box AI models are used as the basic framework e.g. Amazon’s LEX and Microsoft’s  Bot Framework and Cognitive Services solutions (object and image recognition, speech and sound recognition and reasoning). These can be tweaked later (either the interface or the code itself).

 

ACTIONS

In the end, the chatbot is just a little icon on your website.  However, there is still much work to be done. Miss these detailed ‘Actions’ and the AI Chatbot project will fail. Depending on which option you take will determine the details of the actions required.  Remember a chat project is never really finished. It can and should be continually improving via small tweaks and/or more data helping the chatbot to become more user helpful. So now the detailed work (actions) – we create the topics/answer, the question and 5-10 iterations of each question that can trigger the relevant answer.

Conversation Flows (marketing language)  / Decision Trees   

This In-house solution was created by ‘Citizen Developers’ (Tactical Option no.2).

Decision Trees - conversation flows - from Microsoft Virtual Agents
  • Facebook
  • Twitter
  • LinkedIn

Decision trees- conversation flows – from Microsoft Virtual Agents

You do not need any programme/development skills.   Conditions = scenarios – in this case, a condition can be whether you have a front loader washing machine or a top loader washing machine (see below).

Sample questions:

  • When is the special sales weekend for your washing machines?
  • When does my warranty expire?
  • My washing machine is leaking, what should I do
  • My washing turned pink, what can I do?
  • I washed all my knives and forks in my washing machine by mistake – what should I do?

A script to help chatbot converse with a human customer/visitor

Chatbot message Human entry
Hello, my name is <name>, I am here to help you. What can I do for you?
My washing machine leaks water.
I’m sorry to hear that. Let me see how I can maybe help you to fix the problem.

What brand is your washing machine from (just click on the one applicable>?

<option 1> <option 2> … <option n>

Clicks on applicable option
Thank you – I see you have a <option selected by customer> washing machine. What type is it (just click on the applicable option)?

<front loader> <top loader>

Clicks on applicable option
Perfect – now let me know the specific model

<displays model list as drop down>

Selects model
When does the leakage appear – when you start of the washing program or at the end?
 

 

Answers

 

Here is an overview if things you can check for yourself. <links to information sources on YouTube, Company website…>.

 

If this does not help you – I’d be happy to connect you directly with one of our agents or arrange for a technician to pass by your house.

 

Or would you like to see some self service options <links to information>

 

or would you prefer me to book an engineer for you now? <links to  engineers calendar>

 

 

Reviews self-service options and/or decides to be connected to a human and/or schedules an appointment.

 

 

Optimised Resource Planning (also called RSO Resource Scheduling Optimization) is where AI can help customers to book the most appropriate technician (based on the particular problem description plus the items the customer has already checked) plus the customer’s and technician’s time and availability. The booking data is obviously also made accessible for the engineer.

Meanwhile, last but not least, is deployment. You have the chatbot developed and ready to go but how do you deploy it can determine its ultimate success or failure.  The key missing piece in the actions section of many plans is ‘internal marketing’ which comprise: communication, motivation and training.

Make sure people know about the chatbot development early on. Bring them with you. Communicate to them and motivate them about how this will help the business to survive in a hyper-competitive world. Remove fears of redundancy because of chatbots. Get people behind the idea. Perhaps consider redeploying staff into new jobs if the chatbot proves to be very successful – many of which will require training. Managing the chatbots may well require training and certainly going forward maintenance, coding, data analytics and reporting are just some of the jobs required.

NB It is critical that one person takes ownership of the chatbot from the very start, which leads us nicely into the final stage of SOSTAC(r) PLanning – Control – ‘how do we know we are getting there?’.

CONTROL

How do you measure success? Measure the KPIs you wrote in the Objectives section. NPS scores etc.

During the chatbot development stages, measure the MVP you have set . Then when you start testing/training the chatbot – watch closely, make fast changes and repeat.

Stay very close.  Watch your KPI Objectives. After that, you can start comparing last month’s KPIs to next month’s projections.  Don’t overcomplicate things.

Check the data you are collecting. Can it be used to give you insights on customer needs, what they like/don’t like?  What they need more help with? What helps you to identify your ideal customers? You will get a lot of data and insights that will allow you to dig deeper and drive     continual improvements.

Agree Your MVP (Minimum Viable Product) –  your version 1.0   and check to see if its working

e.g. “Tom and Paul agree that their MVP V1 must have following features & functionalities before deployment:

  1. Include all washing machine types of their company with pictures and serial numbers so customers can easily identify their product
  2. Know the top 10 problems customers can fix themselves and so that the chatbot can provide solutions via images, links or text based on customer input
  3. Must have the function to immediately create a service ticket for the helpdesk to directly call the customer via phone
  4. Must be able to converse in English and German for the aforementioned 3 features since UK and Germany are they key markets and pilot regions

 

Stay Calm. Chatbots are complex. Stay on a non-technical level and focus on your business and marketing objectives.
  • Facebook
  • Twitter
  • LinkedIn

Stay calm. Chatbots are complex.

In the end, it’s like driving a car – you are probably not an engineer – you simply use the technology. If things need to be fixed or improved – you go to the people whose job it is to do just that.

Supervising Your Chatbot’s Learning

An AI powered chatbot needs to be supervised to ensure that it LEARNS properly. In other words – you must TEACH it, give it supporting guidance and directions. That is achieved with defining “confidence scores” for each intent and related answers or information sources. Example:

  • Customer asks: “How is the weather outside?”
    • The bot’s NLU and NLP identify “weather” and “outside” – giving a 99% confidence score that you are asking about the weather.
  • Customer asks: “Is it sunny or rainy?”
    • The bot, in the beginning, will not be able to associate “sunny” and “rainy” with a question about the weather. So, it will give a very low confidence score. It might even respond with a wrong answer.

 

Managing Confidence Scores

Here is where managing confidence scores allow you to manage the responses. You screen the questions asked – filter, based on the automated confidence score the bot gave and begin to fine-tune and manually train your bot. That will take more time in the beginning – but with increased usage – it will take less time and provide better results.

Based on NLU (Natural Language Understanding) and NLP (Processing), the user’s intent is determined and, based on a confidence score – the answer selected is the one that is most feasible. Defining confidence levels is a balancing act   between say “I don’t know” vs giving the wrong answer.

This depends on how important a topic is from the customer’s perspective. And that goes already quite far into AI, machine learning and a bit into deep learning.  You ask:

  • “What is the temperature tomorrow in Dublin” an 80% confidence score for the AI thinking you mean “what is the weather tomorrow in Dublin” is OK.
  • “How long are the shops open today” a 99% confidence score for the AI would be needed to know that you mean “How long is the mall around the corner open today”?

Here are some examples for confidence scores and features you can apply:

If a question is asked and the bot does not fully understand (e.g. confidence score between 60% and 80%) – the bot could clarify it by suggesting topics e.g. “Did you ask about <topic x>”).

If the initial question has a confidence score between 75-90%, but the question has a typo – the chatbot will specifically reply: “You typed Dutsche Bnudselagi – did you mean Deutsche Bundesliga or Deutsche Bundesbank or Deutsche Bundespost?”.  Each option then could be directly selectable.

If the confidence score is below 50% (or any threshold you define), you can have the bot offering to connect directly to a person for a live-chat or simply respond “I am sorry, I do not understand what you ask. I know about <topic 1>, <topic 2>,…<topic x>. A properly phrased question could be this: “What is the cost for a top loader washing machine with energy level B?>”

Confidence scores should be reviewed and evaluated topic by topic and adjusted as needed to avoid giving out false information.

The Route to Success
is to define and create proper topics and help the AI to identify the intent based on keywords. If a question causes a confidence score lower than 50% the bot will basically say “I don’t understand” and send a message to the human team to check (and categorise) the question. Based on our review we then can create a completely new topic, adjust the confidence score to relate to an applicable answer, or connect with our developers in case there was actually a technical issue preventing the bot from answering.

Feed the AI with Questions and Answers
If we had 10 basic questions. And say, possibly 5 variations of each question. This is a very crude example of data set. This could be presented as a 3 column table or an excel sheet (see below).

Creat Variations (or 'iterations') of a basic question
  • Facebook
  • Twitter
  • LinkedIn

Creat Variations (or ‘iterations’) of a basic question

Doing this manually is a citizen development approach but other approaches will often have some manual approaches too.

We attach an answer for each FAQ.   Each question and its variety of similar questions (‘iterations’ which basically ask the same question) will have an answer Linkedin to it. You can write all imaginable iterations of a question and then link it to the same answer.

Say you have 5 versions (iterations) of a question the chatbot identifies the INTENT (from the use of the keywords in the question) and then provides the answer.

Answers as well can be in various types and formats e.g. pictures and/or videos or text attached to an answer? E.g. take the live weather feed from youtube.

RESOURCES (the 4Ms)

What Resources Do You Need (The 4 Ms)?

Men and Women (human resource) + money (budgets)  + minutes (timescales) +megadata (data – structured and unstructured). ‘Resources Required’ depend on the company’s maturity and readiness in various areas.

MEN AND WOMEN  A different company that created an internal chatbot had the following resources: a strategic leader, 3 subject matter experts from the team, 4 external developers and 4 Microsoft specialists who’s support was included in the contract with Microsoft. They needed 9 months from version 0.0 to deploy version 1.0.

MONEY  (budget) A project manager can allocate a number of hours each week to the project. This can be fully costed. Then there are also license costs which vary enormously e.g. from $400 pm to $4,000 pm for, say, 10,000 requests p.m. Alternatively, a Flat Fee can be fixed at whether you have 1 enquiry or 1m enquiries.  Ask the question: ‘If I get x000,000 viewers/enquiries how much will it cost?’

Hosting – you need to check if the above costs include the cost to host the application (the bot itself) on (a) CSP (Cloud Service Provider) like Microsoft, Google, IBM, AWS plus give the bot access to your data or (b) on-premises (data centre). You can easily spend hundreds of thousands on a solution that makes AMAZON jealous with dozens of developers and features and functionalities and with connections to data sources. Here is where a Solution Architect, Solution Designer or a representative from a vendor can help to calculate and estimate costs and feasibility.

MEGA-DATA Data includes all data – both structured and unstructured. This effectively includes all data and information that can be used – whether (a) a database of customers (and their preferences plus their previous purchases) for personalising answers,  (b) Q&A lists (for recognising questions and ‘intent’ as well as the sending the right answers (c) data readiness.

Establishing information & data readiness for chatbot is a critical step in any AI (or even CRM project).   Is your data ready to be used (is it clean and consistent in, say, the use of first name, second name with first letter in capitals, plus does every customer list the type of washing machine they bought?) etc. If the data is not OK – you may need dozens of people cleaning it up. This can take weeks or even months. However – if everything is OK, you may need no more than a handful of people in total.

Data Readiness is more complex than capital letters for names. It means that data is ready to be used at different levels and angles:

  1. Accuracy – is any data we provide correct, consistent, cohesive, always up-to-date and owned by us?
  2. Security – can data access be misused to breach our network? Accessibility – chatbot is given access to read, write, modify and create content (access includes diaries)
  3. Compliance – are we only displaying data that is needed?
  4. Technical – is accessing the data actually possible?

MegaData – ensure all the data and information is ready for the chatbot to communicate to users. This includes Conversation Flows/Decision Trees.

Security  –  protect your business from ‘malicious intents’ (a) hackers accessing your data (b) attacks from say an aggressive competitor 10,000 enquiries per second which creates –   DDoS (Distributed Denial of Service) = system overload = systems crash.  There are 7 layers of security including network security, application security and data security.

MINUTES  / TIMESCALES   how long does it take to get a chatbot up and running?  Again depending on the solution you want to see after completion. And how fast you want things done. So using a basic project framework setup triangle might bring everybody in alignment (resource & scope & cost à Quality). In uncertain/ambiguous situations like these – going agile for execution is the best option.  Set an exact number of hours p.w. on this project.  Assign people and say “GO! See what you can do with the money we have in 12 months”.

Incidentally, Speed of Bot-Response should be agreed: 3-5 seconds or instantaneous + volume of enquiries/interactions from visitors (5 an hour or 500,000 a day) affects costs & solution design and architecture.

—end—

Many thanks to Tom Sickert. This is an early draft and so we welcome your comments, queries, challenges or improved examples. Please do post a comment.

If you liked this you might also enjoy:

Chinese GirlBot With 465m Boyfriends

Artificial Influencers – Meet Shudu & Miquela

Artificial Influencers Use My Magic Marketing Formula (IRD)

Here Come The Really Clever Bots – where AI meets customer needs

Join me in Clubhouse in my club called SOSTAC® Plans any Friday 3.30pm – 4.00pm BST for a chat, Q&A, observations about SOSTAC(r) Plans and any other marketing related issues including AI Driven Bots.