How To Avoid

Pyramid Of Doom

When Designing

Bot Workflows

Alex Bunardzic

October 14, 2016

Vienna, Austria

me

• 26 years of software development
• 50% engineering, 50% process
• 50% business analysis
• fan of purity in programming (no side effects)
• intensely dislike self-serve (will be explained later)

Two Families of Bots

1. Concierge

2. Butler

Concierge

Loyal to the Corporation

Butler

Loyal to the Individual

Butler

• Focused on learning about its owner's habits
• Focused on detecting recurring behavioural patterns
• Trained to respect the owner's preferences

Butler Bot

• Stars in many Hollywood movies
• Needs heavy duty AI

Concierge

• Focused on learning about the corporate goals
• Focused on consistent customer service
• Trained to detect business opportunities (i.e. up-sell/cross-sell)
• Sets the tone of the conversation

Concierge Bot

Establishes the conversational protocol

Concierge Bot

Doesn't need beefy AI

Conversation with a narrow focus

Example: Medical Receptionist

Typical experience with a Medical Receptionist (Concierge)

• Receptionist sets the tone of the conversation
• Establishes the conversational protocol
• Asks for pertinent information (name, insurance, phone, date of birth)
• Impervious to any attempts to spark up irrelevant conversation
• If a visitor refuses to comply, receptionist may ask them to leave

Concierge Bot model requires rudimentary NLP

Self-serve

Technological advancements have ushered the era of

self-serve

High tech self-servitude => instead of human cashiers, self-serve checkouts

Consumers are duped into doing all the legwork while still paying the full price

Only businesses benefit from the self-serve model

Concierge bots are the next step in the evolution of high tech self-serve paradigm

What's the difference between self-serve and full-serve bots?

Using a self-serve bot is similar to explaining your problem to a toddler

Using a full-serve bot is similar to explaining your problem to a mathematician

Self-serve/full-serve dichotomy is similar to the imperative/functional programming dichotomy

Challenges

Challenges

• Even rudimentary NLP is a huge challenge!

Google Assistant

Challenges

• Even rudimentary NLP is a huge challenge!
• Luckily, there are commodities and neural networks to help

Challenges

• Even rudimentary NLP is a huge challenge!
• Luckily, there are commodities and neural networks to help
• One challenge remains largely unaddressed

Workflow

Workflow

• Bot workflow is an uncharted territory

Why?

Bots are Stateless

Stateless Bots are Useless Bots

How to Endow Bots with State?

Implement Workflow!

Bot Workflow persists the state of conversation

Workflow can get very complex!

Is there a prescribed solution?

Yes, there is. It's called REST

How does REST implement workflow?

HATEOAS: Hypermedia As The Engine Of Application State

Work flows through piecemeal discovery

All the information necessary to decide what's the next step is

in band

There is no need to establish the purpose of the conversation beforehand

There is no need to reach out for any additional information that's

out of band

Nice. Why don't we use REST for bot workflow then?

Unlike apps, bots have no routes

In REST, business logic is compartmentalized in routes

In bots, business logic is piled up in a big

bowl of spaghetti

Also known as the Pyramid of Doom!

So what's wrong with the Pyramid of Doom?

Difficult to read

Difficult to reason about

Very hard to modify

Extremely hard to troubleshoot

Extremely hard to debug

Almost impossible to document

Is there a solution to this?

I wasn't able to find ready made solution for bot workflows, so I rolled up my sleeves

Fight the Pyramid of Doom!

Divide and Conquer!

Bot Architecture

Two Layers

Commodity layer

Core competence layer

Commodity Layer

• Messaging channels (SMS, Messenger, WhatsApp, Slack, etc.)
• Natural Language Processing/Understanding (NLP/NLU)
• Computing infrastructure (servers, clients, message queues...)

Core Competence Layer

• Business application domain-specific code
• Business entities
• Business operation

Business Operation

1. Dispatching
2. Authentication/authorization
3. Validation
4. Business rules/processing logic
5. Persistence
6. Response

Business Entities

1. User
2. Conversation
3. Context

Conversation

1. Belongs to User
2. Remembers messages received from the User
3. Remembers replies to the User
4. Declares the next step
5. Drives the workflow

Next  step?

Determined by the business logic

How is next  step implemented?

If using imperative approach, results in Pyramid of Doom!

More desirable to implement next step using declarative approach

How to implement

declarative approach?

Compartmentalize processing logic into isolated components

Similar to REST and HATEOAS

In REST isolated components are implemented as resources

HATEOAS provides the next step

Example (simplified)

• User asks the bot to see the menu
• The bot messages back with the menu
• User is supposed to pick an item and send it to the bot as a message
• Bot checks if the user chose an existing menu item, or if the user said something unrelated 

Implementation

(gory details)

Create a component ProcessMenuSelection

Declare next step to be "TakeItemSelection"

Create a component TakeItemSelection

Declare next step to be "TakeItemQuantity"

And so on...

How to execute this Workflow?

Create a main diver component:

Entry with a single method: dispatch

The main engine runs on two lines of code:

1. clas = next_step.constantize

2. clas.new(params)

OK, but where is the next step coming from?

next_step = user.conversations.last.next_step

Wrap up

Bots for Business https://medium.com/bots-for-business

Welcome to Web 3.0!

Questions?

Parting is a sweet sorrow

• https://twitter.com/alexbunardzic

• https://medium.com/@alexbunardzic