HR Chatbots - It's all about language

In his excellent blog post “Chat bots in Human Resource”, Adit Jain lists three persuasive reasons why building an AI chatbot makes sense:

Employees can connect with HR via mobile devices
HR can respond in real time, resulting in faster employee decisions
Chat bots can collect data to create more efficient processes

Building an HR Chatbot to respond to HR queries – and as a front end that can initiate automated standard HR processes – is a compelling idea. CHROs dream of increased employee engagement. Shared Services Leads (and Finance behind them) rush to model a new world of efficiency improvements and cost reductions. And they’re none of them wrong to have that vision.

In this article we outline the key things you need to consider before building an HR conversational chatbot

You need different competencies

But there’s no silver bullet. Moving from an experience which includes a set of FAQs or a Service Agent at its centre, to one where a chatbot interacts with employees, requires a major shift in mindset as well as investment in a completely new set of competencies.

In just the same way that transitioning from Waterfall to Agile meant that the people who wrote Requirements Documents had to learn to write User Stories, the first step on the journey to a successful Chatbot is to understand that using the language of, and the techniques used to create FAQs and Service Agent scripts aren’t going to work: you are going to have to learn (or buy in) the skill of crafting natural language conversations where intent is recognised. Think of the difference between a technical writer and a playwright.

You need to train a Bot to understand Intent

It may sound obvious, but the key to successfully replacing FAQs and Agents with a chatbot is to create an experience where the intent of the user’s plain language query is understood and they get the answer they want or can initiate the process/carry out the action they want.

When there’s a human involved (either an employee or service agent) there’s no problem because humans understand language. The employee knows the intent of their question and searches through a set of FAQs until they find a match. The language used within the FAQ will be quite different from what they would use if they were talking to an agent, but it doesn’t matter because they have the natural ability to match the question in their head (x) with the written question (y): x=y, therefore this is the question I’m looking for. In a call to a Contact Centre it’s the agent who interprets the intent of the caller and matches it to an issue within their script.

Chat bots work in the same way but, just like humans, they can’t (yet) magically develop comprehensive linguistic capabilities overnight. They need to be trained, and the trainer needs to be a human. Fortunately, unlike humans, they learn incredibly quickly.

The people you Need

You’re going to have to set up a team, and not just for the initial project – for ever!

Apart from your developers, that team must include a Content Owner (HR) and a Creative Copywriter whose job it is to convert FAQs/Scripts into natural language conversations.

The Content Owner must be an HR Expert who understands the subject and can confirm that the Copywriter is crafting a correct answer.
The Creative Copywriter should be someone who understands the linguistic nuances of your users. If you have an English copywriter creating conversations for a Philippine audience, it’s not going to work. Get someone who understand Philippine English. It absolutely shouldn’t be someone who normally writes FAQs – nobody speaks FAQ-ese!

HOW YOU DO IT

Initial Training

Don’t try to convert every single FAQ into a conversation (you only produced that many to avoid a Change Request with your outsourced partner!). Rather, analyse the calls to your Contact Centre and choose the top 100 use cases at first.

For each of your 100 use cases the Copywriter needs to craft the Question and Answer. Ideally, they need to craft 20-30 different natural language versions of the same question (utterances) so that the language AI has a good place to start in determining the user’s intent.

If you’re using Microsoft’s QnAMaker.ai you only need one question, but you then need to test the FAQ with lots of different versions of the same question until you’ve reached 20-30. This is done by a human sitting at a computer (preferably the copywriter).

Refining and more training

You’re now ready to release to a group of testers, but before that your developers need to ensure that two things are in place so that you can quickly improve the effectiveness of the BOT:

A feedback request after each question is answered along the lines of “Was that the right answer, (Y/N)?”
A daily report which lists all the user questions that a) the user response to the feedback question was “No”, and, b) the AI didn’t feel confident enough to provide an answer for.

Each day during the testing period (or in real time), the Content Owner needs to view the report and either retrain the BOT where the intent wasn’t recognised; or update the Knowledge Base with a new answer (written by the copywriter) where appropriate.

Ideally, you should get to the point where you have 200-300 versions of the same question to feel confident that the Bot can answer anything; but remember, this is an intelligent system so often fixing one problem fixes ten!

Deployment and BAU

You’re finally ready to deploy and move to Business as Usual which means running the daily report and checking where you need to make tweaks because someone will always come up with a new way of asking a question! That’s why you need your team for ever.