-------
In a digital and AI-driven universe, one sector still surprises by being extremely non-digital, which is industrial distribution. Even with all the development around chatbots and customer portals, a large part of the B2B commerce occurs in the phone. That idea is what prompted Smit Dagli, co-founder of Kanava AI, to start a company aimed at automating voice interaction in this frequently-neglected industry. In an interview with Smit, we got a detailed insight of the same as he shared the story of Kanava, the hurdles of applying AI to old-world industries, the technical and cultural difficulties of applying AI in an old-world industry, and just what customer trust looks like in a market, which is accustomed to doing things the old-fashioned way.
We (my co-founder and I) spotted this huge opportunity in industrial distribution, this massive gap where so much business is being done over the phone. Placing part orders, inventory calls and inquiry into the status of an order all contribute to calling customer service lines and are difficult to scale. We understood that although consumers had access and were experiencing chat and automation in their industries, B2B distribution was still trapped in industry-unfriendly systems and human processes. Kanava was created to plug that hole- with actual automation of phone transaction, real automation, not just point-question-answer.
It is all about speed and habit. Customers in a high proportion of these businesses are on site at a job or a warehouse, so they do not feel like sending an email or searching a web site. It happens that they just make a phone call, order a part, and want it to be handled. There is something in that immediacy which is difficult to substitute. Voice is better, too, at managing ambiguity; people describe components in shorthand, use legacy names, or refer to historical orders informally. A human (or a clever AI) is needed to comprehend that.
It is all a mess. You have thousands of custom SKUs, regional nomenclature, antique ERP systems, some of which are operating over AS/400s dating to the 1970s. General purpose LLMs are not enough. We created retrieval-augmented generation pipelines so that our responses are anchored in real-time customer data. This implied integrations into ERP and CRM, both deep and fast. Whenever it comes to voice, there is no room to lose time, once the AI gets more than a second to react, people hang up.
At first, we imagined we would create an intelligent FAQ agent. However, when the early pilots tested it, they discovered that, unless the AI had the capability of carrying out real transactions, aka placing orders, checking stock, then it would be no good. That caused us to be directly integrated into backend systems. We also constructed a hybrid escalation model. The AI manages 60-80 percent of calls through the whole process after which it can seamlessly transfer the call to the human. That has played a key role in creating confidence.
We have piloted with a number of middle sized distributors in North America and India. There have been encouraging signs even at an early stage whereby, on the average, our AI agent is able to resolve or cause triaging of most incoming calls without any human involvement. We are also looking at good inbound interest of manufacturers that need to augment their sales and support with AI. YC has also accelerated our feedback loop a lot, and we now have finer-grained integrations and improved performance.
Latency and integration are tied. Latency as the voice is ruthless, you have to respond in sub-seconds to have natural conversations. Integration since all customers will have another stack. We created a modular integration layer capable of communicating with both new API-driven applications and legacy on-premises systems. That has been a game changer.
Get a painful, narrow use case that you can solve end-to-end. In our situation it was through automating the placing of orders through the telephone. Second, be close to the actual customer processes, laboratory testing will only take you so far. And last but not least, measure everything. In voice AI, your north stars are latency, intent accuracy, and task completion rates, particularly.
We are optimising the product in the sense that we are increasing and decreasing automation and control of the human being. In B2B there is trust and transparency and so much of it, and to that extent, we are trying to make the escalation to be more cooperative and more natural. Others include nearby verticals such as equipment rental and parts distribution which we are also experimenting with. Whenever there is complexity and phones ringing, that is where voice AI is needed.
The following wave is not chatbots that represent FAQs. It is concerned with agents dealing with workflow and are able to make actual transactions. B2B-trade is long overdue for such an upgrade. When used correctly, AI will not be an assistant tool, but an ultimate component of business operations.
As Dagli and his colleagues move ahead in Kanava AI, they are showing us that even the most traditional industries can use the latest AI, provided you are willing to meet them at their levels.