AgaravsGoogle Dialogflow, Amazon Lex, Rasa & IBM Watson

DIY platforms such as Dialogflow from Google, Lex from Amazon, Watson from IBM and Rasa have made machine learning accessible and allow even mid-sized IT teams to quickly build conversational AI. But they have limitations that become apparent pretty quickly. Crossing those limitations is not only difficult, it is expensive and requires specialized talent. Agara helps IT teams - large and small - overcome these limitations and build truly intelligent and effective conversational AI that deliver strong ROI and great consumer experience.

Agara

Purpose-built for Phone Support Automation

Agara has been built with a singular purpose - high-quality, natural autonomous conversations over voice. We have spent the last 2 years identifying the right ingredients, developing them (often from scratch), and fine-tuning them to make them highly effective. Instead of using off-the-shelf solutions from others, we have created custom solutions for the problems we encountered. - Our speech recognition engine is a custom engine trained specifically on phone call data from customer support teams from around the world- Our dialog builder has been built afresh for voice conversations which tend to be long and multi-directional- We have created special modules to correctly identify elements from the customer's voice to eliminate errors The result is a system that outperforms multi-purpose platforms on every metric. It is faster to set up, more effective in its delivery, more useful to the customer and has a lower total cost of ownership.

Dialogflow, Lex, Rasa, Watson

Generic Machine Learning for all purposes

These platforms are like a box of Legos. There are a lot of individual items to play with and interesting projects to build with relative ease. To build a voice conversation engine, you will bring together speech recognition, a dialog manager, and text-to-speech. Some of these components you can train further, for most, you cannot. These components, however, are general-purpose by design. They are created so they can be integrated into any kind of project. The speech recognition is supposed to work across all audio inputs but is not optimized for phone audio (which is rather different). The NLP is suited to short statements rather than long, multiple sentences (which is how we talk). The dialog manager puts all the onus on you instead of bringing the intelligence of actual conversations (which is necessary). The text-to-speech is generic and not tuned to how support agents speak (which can determine the quality and warmth of conversations). The result - it is easy to build simple projects but has a lot of trade-offs when building a production-quality implementation.

Correctly understanding the facts provided by the customer (entities) and the specific assistance they require (intents) is critical. Better recognition can mean the difference between a loyal customer or social media ridicule.

Agara has developed an advanced machine learning-powered Spoken Language Understanding (SLU) module. SLUs are deeply trained units with a single focus - identify entities and intents from the customer's speech. They operate directly on the voice so do not suffer due to poor quality of speech recognition. They perform significantly better when accents are involved or when noise is present. In addition, they can be trained with your data to make it even more powerful when deployed in your setup.

Further reading:
From the blog: How Agara achieves high accuracy in speech recognition
All platforms convert customer speech into text and try to understand the text. However, a lot of detail is lost during the conversion - real-world transcripts are only about 70% accurate. Remember those sitcom scenes where someone is shouting 'Re-pre-sent-ative' into the phone multiple times in the hopes that it catches their point. The result - these platforms often misidentify things that you would expect them to get right. For instance, your name, your order number, or what you want in the conversation. These platforms are built for creating chatbots and, hence, do not solve for the unique problems that are faced in voice conversations.

Further reading:
From the blog: Chatbots + Speech ≠ Voice Bots
Customers DO NOT like fixed scripts. To build engaging conversations, someone must predict all the different possibilities that happen during the call and account for them. This is difficult even for the best of us.

To craft conversations, Agara uses Conversation Blocks and Conversation Manager. Blocks are independent pieces of conversation like understanding the reason you are calling, getting your ticket details, executing a change to your order, or initiating a transaction. We have built scores of these. Each block already includes possible questions that can be asked. For instance, when canceling a ticket, questions about the refund policy, refund timelines, re-issuance, dispute handling, and more are already included in the relevant block. Your IT team no longer needs to spend days identifying every possible question. Further, we train these blocks on real-world data. The result - instead of us hang-programming the conversations, we 'learn' from what customers have asked in the past. Even the smallest edge cases are easily identified and handled. Finally, the Conversation Manager knows how to use the blocks in the best way possible. So your IT team only ever needs to identify what is relevant but not worry about putting them in a hard-coded sequence. They specify what is the best-case scenario and Agara takes care of it when a customer does not follow the path.
IT teams building conversations using these platforms work with the customer support teams to mimic what 'typically' happens in when a customer calls in with a specific requirement. Every customer support person knows that these conversations can go in a hundred different ways. However, it is very hard to build for a hundred different possibilities. The final result is often built for the most common scenarios with the hope that these would account for 80% of calls. These conversations are hard-programmed and force the user to take a very specific path. They also do not allow for any improvisation whatsoever for the customer in case they have a doubt. These conversations, while definitely useful in several scenarios, are not good for handling anything other than the simplest of queries.
Newly designed conversations are not perfect. Based on feedback from customers, they improve and get more relevant and accurate. There must be a mechanism to collect and channelize this feedback in a timely manner.

Agara has in-built mechanisms to track the points where the customer encountered a problem, did not get recognized by the system, or expressed frustration. These mechanisms provide real-time information to the team about specific areas that need improvement as well as how those improvements can be brought about. In addition, Agara's team comprises linguists, speech UX specialists, and machine learning scientists. Linguists ensure statements/questions are well-formed and meaningful. Speech UX folks focus on the best, most effective way to handle every part of the conversations. Machine learning scientists ensure that improvements can be driven not just from better-formulated questions but also by creating newer better models.
To ensure that customer feedback is collected, logged, analyzed, acted upon, and implemented, there are several components that will need to be built in addition to the core conversation. It is important to remember that the feedback comes in the form of conversations that did not go as planned - insights must be extracted from this raw data. This requires a dedicated team that is always on the lookout for opportunities to improve. The team must build dedicated data pipelines that can ingest, cluster, and analyze customer calls. They will then need to turn insights into modified conversations, build them, test them, and deploy them. This is a virtuous loop that differentiates great conversational AI from the barely put together ones. Importantly, the team will encounter problems that cannot be fixed by feedback alone. For instance, say your AI has problems correctly identifying the date of birth used for authentication. This requires delving deep into the machine learning models and cannot be handled purely by the platform. Problems with the system correctly understanding the customer is the #1 reason why customers sometimes do not like autonomous systems.
Effective conversations require better machine learning models. While platforms provide the tools to train these models, they also require large amounts of context-sensitive data, extensive expertise and a fair bit of time to get the right results.

Agara uses SLUs to understand more. Agara uses SLUs to understand more. Agara uses SLUs to understand more. Agara uses SLUs to understand more.
Agara uses SLUs to understand more. Agara uses SLUs to understand more. Agara uses SLUs to understand more. Agara uses SLUs to understand more.

Build or Buy?

Buy from Agara
Build in-house

With Agara, your spend is 60%-80% lower than current manual spends, highly predictable and in direct proportion to the volumes processed. There are no fixed annual charges, no hardware/software based escalations and no limit on number of workflows being used. All upgrades to the system and increases in accuracy through machine learning investments come at the same predictable price too!

While initial costs can be low, high quality infrastructure and dedicated staff adds up in cost rapidly. These costs also tend to be fixed in nature irrespective of the output or result. Any upgrades and even routine maintenance require constant investment in machine learning qualified staff, experimental IT infrastructure and long lead times. Overall, the cost of a production-grade build over a 1+ year timeframe can cost 20%-50% more than buying from Agara.

Agara is a machine learning focused company. With a current team of 10 (and constantly expanding), we bring advanced skills to the product on day 1. The combination of speech, NLP, user experience and linguistic experts on the roster means you get the best help everytime. If needed, we even dedicate one of our machine learning experts to ensure the best results are achieved for you.

Simple projects on these platforms are easy. However, for anything more complex, you will require machine learning talent, either in-house or contracted from outside. Quality machine learning talent is hard to find, in high demand and, therefore, very expensive. In addition, everyone has a lead time before they can make real impact for you. Lead times can be significant for anything more than the basic conversations leading to projects going off-track.

Agara has been building and implementing large-scale conversational AI products for global enterprises for the past 3+ years. We have worked with several systems across several countries. As a result, we are able to deliver new implementations at a rapid pace and bypass many potential pitfalls.

Building conversational AI workflows using the platforms is only step 1. To take the project to production, one needs to solve extensive topics like infrastructure, user testing, latency, workflow comprehensiveness, analytics and more. Since this is a new area, each of these topics throws up unexpected results that must be dealt with. Unforeseen challenges in building production-ready solutions can create serious unpredictability in the project and throw timelines off-track.

Agara is designed for fast rollouts leading to faster returns on your IT investments. You would be testing new workflows internally within days. With standardized APIs, integration to your cloud systems is a breeze. Even large enterprises can expect to be live in as little as 4 weeks.

Week 1

- Use case discovery- Workflow development and customizations- Internal testing underway

Week 2-4

- IT integrations- Workflow tweaks and fine-tuning

Week 5

- Phased launch- Further fine-tuning based on feedback


Integration period mentioned above assumes systems where APIs are already available. For instance, Salesforce / Zendesk for CRM and Amazon Connect for telephony. The period may take longer depending on your specific systems and policies.

Start optimizing your voice support today

One of our friendly and knowledgeable experts is always at hand to partner you through your transformation journey.

Schedule a Demo

Learn how Real-time Voice AI can help you

Schedule a Demo

Have an urgent requirement and wish to talk to a sales rep? Call +1-512-333-4634 now