Processing of human language via machine programs has been coming to corporate computing for a while now. Some believe it can be not only reliable but also very broadly applicable.
Software startup Moveworks of Mountain View, California, which has been using natural language understanding systems such as Google’s BERT to process IT help desk tickets, Wednesday announced that its automated agent now serves other employee needs throughout a corporation, such as finance, human resources and compliance.
“If there’s a question we’ve been asked since the founding of the company, it is, will you do more than IT,” said Bhavin Shah, Moveworks co-founder and CEO, in a meeting with ZDNet via Zoom.
“The level of service that we’ve been able to provide in IT is starting to make more sense in many areas, like when people have HR questions about a policy update, or a legal affairs question, like, where is the latest corporate NDA,” he explained.
“It shouldn’t take long, these are simple questions with ideally simple answers, but it takes the effort that we are taking to do it right.”
ZDNet has written previously about how Moveworks used “sentence embeddings” to ingest examples of things people ask their help desk systems. Moveworks’s software is then able to automate responding to those sayings in natural language, and then automate the process of getting the user to a resolution.
The company has operated on the principle that “language is the culprit” in multi-stage functions in companies such as the IT help desk ticketing system.
The use of BERT and other language models allows Moveworks’s agent to pop up answers to questions, and propose next steps, inside of workgroup applications such as Slack and Microsoft Teams.
The new offerings consist of modules for HR, finance, facilities and employee communications, which get added to the underlying core program that was developed for IT ticketing, which runs what the company calls its intelligence engine for processing natural language. Early customers for the new modules include Nutanix, Palo Alto Networks, and AppDynamics.
As Shah explains, the idea is to focus on areas of the corporation that have some support aspect built into them.
“The departments we deal with now, HR, finance, legal affairs, facilities, those types of groups, tend to be the ones that have support in their nature,” said Shah. Marketing departments also have conversations, but they have less of a volume of conversation than those other areas, he explained.
Technically, “the shape of the problem is similar to IT,” said Vaibhav Nivargi, Moveworks’s chief technical officer and a co-founder. “Even before we rolled out these capabilities, we would see people say, Let me see if I ask it an HR question, what does the bot do?”
What has taken some effort is to connect the language understanding “core” to the particular domain knowledge of other departments.
“Some of our language understanding capabilities would interpret these issues correctly, given that we have a multi-dimensional intent system that understands this,” said Nivargi.
“But then on the downstream processing side is where things become very specific, with these departments having their own ticketing systems, or queues, and that’s where the whole architecture has evolved, layering on an understanding of language and then expanding to support different domains with a very high degree of accuracy.”
The company is using some pre-built frameworks, such as the Hugging Face Transformers library for NLP models. However, it has developed an infrastructure consisting of a bidding system to find the answers to questions.
That has to be tied into an optimization procedure that takes into account many measurements, not just the language understanding component.
“We know at the machine learning level what is the precision, recall, accuracy, the right metrics for that, and then the conversation metrics, and then the resolution metrics,” explained Nivargi. “We want to make sure the conversation metrics all end in resolution.”
Refining the function of the application, then becomes a kind of economic model of optimization. “Did the model fall short, did the bidder fall short, did we not have a supply,” said Nivargi, describing what he called a “joint alignment.”
“People think it’s just a machine learning problem,” observed Shah, when in fact, he said, what is involved are numerous tasks pertaining to data integrity and many other elements, “the stuff around this to actually make it run,” as he describes it.
Given the investment in developing all that complexity, “IT has been a proving ground for this, but it extends very naturally” to the new tasks, Nivargi said of the technology.
The COVID-19-induced pandemic accelerated the business, said CEO Shah. “The CIO had to make sure that everyone was productive” in lockdown, he explained. “There is a level of focus on employee experience that we’ve not seen before.” That is reflected in the spread of IT roles with titles such “employee experience officer,” etc.
That means, said Shah, “having the walk-up genius bar is no longer sufficient, it’s, how do we get to everyone in this hybrid world, where you are still going to have everyone all over the place.”
And that complexity of reaching and enabling employees has translated into employee frustration and impatience with traditional processes, said Shah.
“I was late for a meeting today by a few minutes, and I was apologizing, because we are so hyper-focused now on our timing,” reflected Shah. “When someone can’t get help for six hours or are waiting for an answer, people get up in arms about it, why these things are so slow.”
The software has become an even broader tool with the so-called communications module, which uses the chat bot function more broadly for sending proactive messages to employees.
“What happened that we weren’t expecting in COVID was that people started to use us as a means of communication,” said Shah. “You had people trying to send updates to employees, such as, here is an update on whether we will reimburse you for your COVID rapid test.”
“That’s where things are starting to evolve where instead of sending an email, and getting no replies, people are using our tool to essentially send an expert into your messaging tool,” he explained. That bot is able to handle questions that arise by the employee, rather than leave that individual hanging after they’ve received the notice.
“Oftentimes, you get this no-reply email, and you have a question, and so before you even do it, get stalled and you don’t take action.”
Shah said the company is seeing “about 70% engagement” on such messages, versus 10% to 15% engagement on email communications.
That suggests Moveworks’s automations could be a form of system of action, as it were, a tool to propel corporate employees to the next action, rather than merely providing information. Given the increasing complexity of multiple systems that have to handshake, including ID systems such as Okta and the likes, handling more and more actions may present a fruitful opportunity for the company.
The next step, Shah says, is to open up the software as a platform that is programmable beyond Moveworks’s own efforts, similar to how the iPhone was initially locked down and then sprouted the App Store.
“A lot of it right now is us doing all the innovation ourselves,” said Shah. “I think there is a mindset that how do we bring these smart folks into the fold a bit, and give them some flexibility to help push this platform forward.”
Moveworks, founded in 2016, has received $105 million in financing in two rounds of venture capital. Shah declined to say what the company’s post-money valuation is now. The company is not seeking any further financing for the moment.
News Source: ZD Net