How to build a Google Assistant platform, from someone who’s actually done it

Officeworks was the very first Australian retail brand to get access to the Google Assistant platform. Speaking at Mumbrella's Retail Marketing Summit, the team detailed the trials and tribulations of playing guinea pig.

If you’ve been to a marketing conference lately, you will have heard speakers wax lyrical about how great voice is.

But while it’s all exciting in theory, the nuts and bolts of putting something as complex as a voice assistant together can leave marketers wondering if it’s really all worth it.

Matt Ware, head of operations at First worked with Daniel Gladki, head of eCommerce at Officeworks, and believes voice is worth the struggle. They worked together on Officeworks’ Google Assistant project, which officially launched in November.


Ware (left): “We broke this assistant in the test environment, at least 15 times”

“We broke this assistant in the test environment, at least 15 times,” says Ware. “The reason that we did this is we were incorrectly building the way that conversational flows were structured inside the assistant. The problem is that traditionally most people have looked at things like this as almost being like an IVR inside the speaker, or as a traditional AI. ”

Ware explains that in order to understand how voice assistants work, it’s essential to understand the difference between a traditional AI, and what Google has built with their machine learning technology.

The next step is actually understanding how the machine learning ‘thinks’, and adjusting your programming accordingly.

“You need to ignore every declaratory statement. ‘I want’, ‘I need’, ‘tell me about’ is actually superfluous to the outcome that you are trying to achieve for that consumer,” Ware says.

Ware makes it clear quite how strict Google is about ensuring the user’s microphone is switched off as soon as possible, perhaps as a response to those who fear a Black Mirror-style invasion of privacy from owning voice-activated assistants.

“There can be potentially 500 or 600 conversational end points inside of your assistant. Google will follow every single one of these until they hit a conversational end point. You then have to turn the microphone off within a specific set of time if the user does not interact with it.”

Google’s strict guidelines proved a problem for the team, especially when they had so many conversational threads to contend with.

“[Google] found three instances where we did not shut the microphone off within the prescribed couple of seconds of the user not interacting with the assistant, and rejected us. They are very tight on the ability for these assistants to be able to listen in on what other people are doing.”

Ware advises that anyone looking to build their own assistant tool must “follow the guidelines and structure against it. Think about exactly what it is that will impact those people who are talking to the assistant to enable them to still have the trust to be able to have it inside of their home.”

The team was surprised how relatively easy it was to get set up analytics within the system, after an initial period of confusion when partners weren’t sure exactly how much analytical data Google would be able to provide.

“You are able to put your analytics platform behind the assistant – most people didn’t realise this when they first put it together, but we discovered that you can actually set it up with a standard installation,” says Ware. “You can treat users as they come through as new users versus returning users. You can see their session times. You can see what intent was triggered.”

Officeworks also discovering customers were talking to their assistant with unnatural speech patterns – an issue the team hopes naturally rectifies itself over time, as customers become more familiar with how a voice assistant works.

“People were speaking to the assistant, and trying to treat the assistant like a website. Someone would get inside the assistant, and say ‘product search’, expecting that you had say ‘product search’ to trigger the ability to talk about the product, as opposed to understanding that this is a native environment,” Ware says.

“Voice is actually probably one of the first truly omni-channel products that we’ve seen, that enables you whilst you’re mobile, to be able to take it from search, to discovery, to in store in one system.”

Daniel Gladki, head of eCommerce at Officeworks says: “Invest in your data structure, but before that, figure out what problem you’re trying to solve. Start with something that’s easy to build that can be launched quickly, to then help you understand how your customers are actually using your device and how they talk to you and talk to your brand. That then informs where to invest further or where to improve.”

Gladki: ‘Invest in your data structure’

Ware notes how listening to your customers was a theme that ran throughout the Retail Summit, and explains how voice provides “a very simple way to truly be able to listen to your customer”.

“Being able to tap into every single person in store who gets a question asked of them by a customer as they come through… If you’re able to see trends of questions coming through faster on a regular basis, then you can identify sooner the problems that you as an organisation might have, or areas that you might want to expand into,” he says.

Another nugget of advice Ware provides is to “have a long-term plan for the channel”.

“This isn’t something you can build once and leave for the next three years like some apps. You actually are going to have to go in and continue to look at this and iterate upon it.

“If you want to build these to be structured and informative for users, you’re going to have to listen to them and make changes based on the information they’re asking for.”



Sign up to our free daily update to get the latest in media and marketing