When selecting a system for your CAI (Conversational AI) system, there are features that will be “must-haves.”  If a solution can’t offer these then they should be dropped from your list.  There will also be features that should strongly influence the decision.   This article started as the 4 key differentiators that matter when making a selection.  It didn’t take long to grow to more than a dozen features to keep at the forefront of your search.

Features You Must Have

Some of these should also be taken into consideration if they might be later requirements.  Be prepared if a company says they are part of the roadmap, but has not yet released them.  After more than a decade involved with tech sales across 3 industries, I can tell you first hand that feature releases almost always slip.  And at times then don’t happen at all.  A good indicator is if the company will let you work it into the contract details related to the feature.  Either an early release clause or a price reduction until the feature is released once the schedule has slipped.

Specific Integrations

One of the biggest ways to keep customers happy and reduce support tickets is when the Virtual Agent can provide information or handle tasks to the human user.  In order to do those tasks, the system has to be able to integrate with the system(s) where the information resides or the system(s) where the tasks need to happen.

These integrations then fall into two buckets:

1. Integrate with the system(s) to get data from.  This might be a need, for example, to integrate with GuideWire or EPIC depending on what industry they are in.
2. Integrate with the system(s) with which to perform some sort of action.  Note that I’ve separated this out since often you can leverage RPA (Robotic Process Automation) to do these tasks.  Many of the RPA platforms have their own integrations.

Some platforms don’t have the integrations built, but have a framework you can use.  This is better than nothing, but keep in mind you’ll be having to build and maintain something that could be quite complex.  Take a look at the API calls for GuideWire as an example.  There is a lot there to build and test if the platform hasn’t yet done so!

Live Agent Handoff

There will be times the conversation needs to be escalated to a human.  If this is part of your strategy you’ll need to confirm it is possible.  There are some other considerations:

• Will the agent get the details of the conversation?  There are different levels of what the system can do here.  Some only give a transcript.  Others can extract key points from the conversation, but if possible this takes additional development effort.
• Can the system differentiate between levels of customers do decide what level of handoff to offer?  For example, for premium or trial customers off a live person.  For ones using the free service offer to create a ticket.
• Can the system provide those waiting for escalation details on how many people are ahead of them and the expected wait times?  For example “You are 5th in line.  Estimated wait time 7 minutes.”

Correct Channels

You need to be able to meet your customers where they are.  Some systems are webchat (text) only, others can do a multitude of text forms of chat – Slack, Facebook Messenger, etc..  Be cautious also of firms that just recently added voice to their offering.

An important item here also is to think about channels that are likely to be desired later.

Visibility and Analytics

There’s an old saying in the Application Performance Management space – “If you can’t measure it, you can’t manage it.”  Your CAI system will always be learning and growing.

Support

Different levels – remember, it isn’t how often you use it.  It is more important how important it will be when you do need it.

Multilingual Support

This could be either a “Must Have” or “Nice To Have” – good segue into . . .

Nice To Have Features

Component Agnostic

Having a modular system where you can leverage other third party offerings for key components gives you a lot of power.  You can pick the price/value level you want.  For example, can you use different TTS/SST (Text To Speech/Speech To Text) providers?  Other NLU solutions?

Exporting/Importing Data and Flows

There are times it is easier to modify something externally. If the platform allows you to export data – what does it look like? Are you able to actually look at the data? Can you modify it using your own tool and import it to another Agent or to update the agent it came from?

Revision Control

There are different levels. One of the items I’ve always liked about Fastly (a CDN/edge compute platform) is that they keep track of changes and you can quickly see the difference between updates. With most CAI platforms, there is next to no versioning and no way to display differences over time or between “releases” – be prepared for this!

Important But Hard To Quantify

Some items are very important, but also very had to measure.

NLU Accuracy

This impacts the user experience. If your Virtual Agent doesn’t understand the user’s Intent from their Utterance, you are creating a bad experience. Less likely the user will return, and more likely they’ll escalate to wanting a human.

Some thoughts on NLG

More to come, along with my thoughts on:

Actual ML and Feedback Loops

Does the system really get smarter?

Ease of use to build/design

Industry Experience