Digital Marketing & Sales Suck (And Also What to Do About It)

This post was originally published in a free Chat Secrets training course ‘Marketing & Selling Using Chat’ . It is republished with permission.

It sucks because most marketing and sales solutions today were designed to automate business processes, and not to offer the intelligence needed to make informed decisions. They were built for the first wave of digital disruption (process improvements) and not what is coming (intelligence improvements)

As a result current on-line web marketing is not very granular. For example

Show everyone who lands on some page a free-plus-shipping offer

is often used as a way to qualify both subscribers and buyers. Web pop ups pester all users to sign up to an offer and are so annoying that millions run blockers

What if we could introduce such offers based on what a user says rather than what she happens to be doing or looking at? 
So can we use an AI to introduce our offers for us? Can we train our AI to introduce our funnels for us? Can our AI pitch the right product for the right customer at the right time?

The Scenario

Imagine we are a fitness business with a chat bot which helps improve our users health & fitness (think MyFitnessPal or the FitBit app). Our chat bot allows users to record and add notes to activities like walking or cycling, get the amount of time remaining in those activities as well as help with weight loss and healthy eating. Also imagine that we have sourced some health & fitness related products from Alibaba which have good profit margins so we are drop-shipping them from Amazon.

We want to offer our users the opportunity to try them out and we want the AI to introduce the offer for us when it thinks that is what the user wants

The Challenge

The thing is that users ask for things in different ways. For example, one user may say ‘I need some new running shoes’ and another ‘My running shoes need replacing’ and our AI is supposed to resolve these two to the same ‘send running shoes’ offer.

So the AI needs to be trained in order to handle these variants. But then the problem becomes — where does that training come from?


It is sufficient at the start when creating a chat bot to do the training manually. However once it is deployed into the wild it needs continuous training in order to be able to understand a wider range of phrasing and vocabulary. Doing all of the model’s training manually is time-consuming and hard to scale long term so one solution is to create a feedback loop from the chat bot.


The technique we used was suggested in this article and we modified it to work with the Bot Framework

What is not easily available (unless you are Google/Amazon/Facebook) are reliably large-scale high quality data sets on consumer digital purchasing decisions. These are exactly the things we need to train our AI so we need to be creative and find another solution.

We need somehow to create a feedback loop so that our users can correct the AI. The problem is that this will soon get annoying if the user is constantly asked if the AI is correct or not. What we can do is to hack the confidence score to only show the user feedback loop if the score is below some threshold value (like 0.7). This works because the confidence score gets generated each time we send a query to the AI so we can use this to decide if we need any further feedback from the user or not.

The AI is not sure of the users intent so asks for feedback based on the low confidence score

Once the user provides feedback then the AI updates with the new information


Next we need to work out how we can use this hack in the wild for some time before letting our actual target users play with it. The more training we can give the AI the better it will work. The solution to this was to use Amazon Mechanical Turk (MTurk) which is a crowd sourcing internet marketplace where people can make use of human intelligence to perform tasks that computers are currently unable to do.
We set up a job asking workers to help train ourAI.

We chose to spend around $20 at a few cents per job. We soon started to get some really interesting results

Our worker wanted to see an offer but the AI got it wrong. Our worker corrected this automatically for us (in red)

Something else which is interesting is that we have started to build relationships with our MTurk workers who chose our tasks (since they can email us through the system)

We always thank them for working on Chat Secrets and try to respond quickly to them. Even though they receive only a few cents for our tasks it is important to let them know their work matters to us.

The actual offers themselves are pulled from a bot called which helps connect users to products they will love. It also allows you to monetize conversational experiences and is well worth checking out. Here is a video of our chat bot presenting an offer

Our Chat bot has worked out our users intent and also shown them an appropriate offer.

If you have a small business and are interested in working with us to implement such an AI in your particular domain please let me know —

Particular areas of interest for us are Real Estate, Dietary Supplements and Drop Shippers

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.