When Ada launched its service; Integrated Reasoning Engine (RE) in Februarywhose goal was to help businesses build, manage, and optimize customer-facing AI agents using a single set of instructions. Businesses can apply these instructions across channels, from email and chat to voice.
“Thanks to the integrated inference engine, [businesses don’t] Manage a voice agent in one location, an email agent in another location, and a third AI agent in another location. ” mike murchisonthe CEO and co-founder of an agent customer experience platform vendor told No Jitter. “If I tell my agents to be more empathetic in an email, Ada can tell them to be more empathetic on the phone, and now I have one task that I had to repeat in five different places.”
RE allows Ada AI agents to work in parallel with each other, Murchison explained. For example, one Ada AI agent is having a voice conversation with a customer while other agents are working in the background. This is similar to how human contact center agents multitask. For example, if a customer’s flight is canceled and they call to rebook, a human agent will maintain the conversation, verify information, find a new flight, rebook and send a confirmation email.
“The tasks that our customers automate with us are increasingly long in duration and complexity, and it is critical that we can continue to engage with them, stay informed, and resolve other issues while we work in the background,” said Murchison. “I think this year we’ll probably be automating the first task, which is about 14 hours of work for a human team.”
No Jitter spoke to Murchison about how the Ada platform works and coordinates activities between AI agents. The following conversation has been edited for length.
NJ: Can you tell us a little more about how the Ada platform works?
Mike Murchison (Murchison): At the highest level, the architecture includes two different language models. One is the “talker” model, which allows for very fast, conversational interactions. The other is the “thinker” model. This is a deep reasoning model that is involved in performing complex multi-step tasks.
These two systems interact with each other in a way that is not exposed to the end user. Enterprises don’t have to think about all the different layers of agent interaction that exist within Ada. They only manage one agent, which is like one member of a team. [while] End users only think about a single experience, a single agent, that they are interfacing with.
We believe in simplicity. That’s because for many companies, it’s an overwhelming world of orchestrating and coordinating to find coordination between dozens, and now hundreds, of different types of agents. We simplified this into a system where only one customer-facing AI agent is deployed and managed.
Yes, technically there is a swarm of agents under the organization. [Ada] Foods work together to perform a variety of tasks, and we want to take that complexity away from our customers.
NJ: Where are you going? ada’s playbook Does it fit?
murchison: One of the most difficult problems in customer-facing AI is how to leverage the creativity of language models while meeting the strict compliance and determinism expected of businesses, especially those in regulated environments.
Ada is an autonomous agent that can creatively take actions based on all the tools it integrates, allowing it to power highly dynamic and rich experiences while adhering to high-level policies. Use case-based also allows you to operate in a very deterministic way, in a way that traditional rigid, incremental systems do not.
This is done using playbooks, a source of natural language instructions that can be created by literally dragging and dropping flow diagrams from traditional standard operating procedures or traditional IVR systems. Simply drag and drop your PDF or screenshot into your playbook environment.
Next, it automatically generates a playbook that is preconfigured with the actions you configured in Ada. ADA follows its instruction set very reliably.
NJ: How will companies know that the handbook is being followed? Also, if an exception occurs, how do you track the error?
murchison: Use language models to annotate all conversations, including inference logs. We give our customers complete observability on things like: Where is this handbook compliant? And if, for example, an API goes down, at what stage did it break?
We also identify and implement improvement opportunities to our customers through our compliance oversight agents.
This is more of an internal story because the customer doesn’t really think that way, but we have another agent who oversees all conversations related to the playbook, checks whether explicit steps are followed, and presents it to the customer in the form of a dashboard so they can take action.
There are also reviewer models that are great for understanding resolution. We annotate every conversation to help you understand it, but was this a relevant conversation? [Was it] Exact conversation? Was it safe? Did the customer do that? [issue] actually Do you want to solve it?
