Earlier in June I had the opportunity to talk to Barry Coleman, CTO of Agent.ai, an about 2-year-old company at the time of writing this. The company spun off of manage.com, a very different business that enable the delivery of in-app advertisements. In order to support this mission more and more, first internal, then external support capabilities were needed.

At first they built chat functionality for internal and for support purposes. Then there was the question of how to efficiently provide 24/7 support. This resulted in giving birth to a bot structure that can help customer service agents in an assisting mode, called co-pilot mode, and an autonomous mode, called autopilot. And it gave birth to Agent.ai.

Agent.ai’s mission is to enable “exceptional customer service for all”.

While this mission is not particularly unique, their approach is. First, Agent.ai has built its customer service software around a machine-learning platform. Second, the company provides their solution without asking their clients for a huge upfront investment or the need to have of AI-proficient developers in house. Third, they wanted to avoid the pitfall of inflated expectations. With AI and machine learning being very hyped topics at the moment, this is a very valid concern.

Going backwards through the objectives, Agent.ai opted for offering very specialized bots first. As there is no general AI yet, this is pretty straightforward. Specific, tightly framed topics are far easier to support with AI and exposed by bots than broader bodies of knowledge. For example, specializations include the handling of order inquiries or of support call closure surveys.

The second objective was achieved by doing all the heavy lifting, including the customer specific training of the AI in their own system, by providing specialized bots, and by offering APIs for their customers to implement own specialized bots.

One interesting aspect is that Agent.ai’s software fabric allows the individual bots to collaborate with each other and communicate internally with agents and externally with customers. This collaboration is necessary due to the strong specialization of the bots and is mainly controlled by a ‘central’ AI-based bot that resides in the Agent.ai infrastructure, called ‘AVA’, which is an abbreviation for Automated Virtual Agent. AVA is the brains of the system.

The job of the AI bot is to understand speech and to identify a user’s intent using NLP, neural networks, and deep learning. This intent could be a request for information or a call to support an incident.

With this done the AI bot dispatches the incoming request to the corresponding specialized ‘intent’ bot that can take up the transaction and hand it over to another bot, or escalate to a human agent in case they get stuck.

The system is trained from a variety of sources, such as FAQ, existing documentation, and e-mail trails.

Chat transcripts prove to be especially valuable as they allow for identification of both, problem and a solution. These transcripts also offer an excellent means for continuously training the bots while being in co-pilot mode, the mode in which they suggest answers, along with a confidence level in the answer, to human service agents. The usage of chat protocols along with the service agents choosing to use bot recommendations or not, allows for constant recalibration of suggestions’ confidence levels.

Which leads to the topic of trust; user trust as well as agent trust – and to the question when a specific bot can be put into the wild and work autonomously. The answer to this is surprisingly simple although there is no explicit measurement: If suggestions consistently exceed a defined high confidence level then the bot is good to go unsupervised and escalates issues it cannot answer itself to a human service agent. Another possibility of identifying trust levels is the change of customer sentiment in the course of a transaction.

Working in co-pilot mode, with the ability to have bots work unsupervised, human agents free up the time to work on novel problems. Typically, these can be the issues that bots haven’t been trained for, and maybe cannot be trained for. Barry emphasizes that “human-machine cooperation is really important”.

My Take

Agent.ai has an interesting story to tell. The idea of offering an affordable infrastructure to provide 24/7 mobile in-app customer service using bots that are driven by machine learning and AI is probably not new but consequently implemented. Bots can considerably speed up the support transaction by continuously listening to specific queues. With well-trained bots this can lead to positive support experiences by showing that a customer’s time is valuable. This also applies to the co-pilot mode, when the bots can already prepare suggestions along with confidence levels that help the service agent prepare herself for an issue.

In addition to providing a toolkit for mobile in-app support, Agent.ai is supporting nearly all major messaging platforms, which allows for richer customer profiles as well as for a wider reach for both, Agant.ai’s customers, and Agent.ai itself. Agent.ai’s customers can offer their customers availability on the channels they prefer without being in the need to look for additional vendors to cover different messaging channels.

Agent.ai’s bot-driven mobile first approach puts the company into an interesting position. Mobile in-app specialists normally do not support messaging services with the correct argument that the service engagement can be made far more personalized. This is due to more information being available to the service agent via the SDK. It simply can provide more information than the messaging service will ever do. On the other hand there will be many users who simply do not want to install vendor apps.

Integrations with Zendesk and Salesforce give exposure to the world of the ‘big guys’. Zendesk does believe that bots are not yet far enough to be really useful in customer facing service interactions. Meanwhile Salesforce does not have any bot capabilities either, as far as I know. Both companies offer integration into major messaging apps, with Zendesk also offering an app SDK, though. Still, this leaves an opportunity for innovative vendors.

I believe that the strategy of covering the breadth of mobile along with the ability to cover small to big customers is pretty strong. It puts Agent.ai dead into a spot that no major vendor covers, while at the moment having a technological advantage.

However, there are also some concerns. I suspect that the approach of taking away the ‘heavy lifting’ from customers may lead to consulting services, which do not scale well. In addition Agent.ai is a young company, which always raises the fear of viability. Agent.ai says it has more than 1,000 customers from small to large in different industries. These customers are using SDK and web client most, followed by the Facebook Messenger. While this sounds like a big number there is no information on actual users.

Still, this is a company to watch.