Drayton Wade, the head of operations at Kognitos, talks AI in retail and where the vendor sees its tech providing the most value.

The opportunity window for retailers between Black Friday and holiday shipping deadlines can seem narrow. The high-stakes weeks put pressure on operations and customer service to get the most out of every resource. And those resources can include new investments in AI retail strategy. Operations and customer service are both areas of focus for the business automation platform Kognitos. The startup closed a $20 million Series A round of funding in November and is looking to grow.

Drayton Wade, the head of operations at Kognitos, shared the San Jose, California-based startup’s perspective on AI strategy in retail. Wade is based out of Charleston, South Carolina, where the company is currently expanding its presence. He worked previously for the automation platform UiPath, eventually joining Kognitos in 2022.

Kognitos has already worked on use cases in retail and consumer packaged goods, as well as manufacturing and logistics. Companies using its tech include PepsiCo, Wipro, Century Supply Chain Solutions and Norco Industries. Wade explained to Digital Commerce 360 how he thinks Kognitos’ existing capabilities fit into AI strategy in retail. He also talked through the problems that the technology in the startup’s platform is designed to solve.

Current demand for AI solutions

“One of the main things we hear from executives, CIOs, CFOs, CEOs,” said Wade, is “‘Hey, I know I need to have an AI strategy. I know that I want to implement some form of generative AI. But I have various concerns about safety, or I don’t know where to begin.’”

That perspective at the executive level reflects known sentiments in the business world. A 2023 study from IT solutions integrator Insight and research firm The Harris Poll found that 73% of leaders at Fortune 500 companies expect to incorporate generative AI within the next three years to improve employee productivity. Also on the list of respondent priorities were customer engagement (66%), along with supply chain (41%) and inventory management (40%).


In the same survey, however, respondents acknowledged concerns about AI, with worries about quality and control (51%) and safety and security risks (49%) showing up most often.

Visibility and auditability

For Kognitos, Wade sees the way forward being to increase visibility where a high degree of documentation exists, focusing on accessibility through a natural language interface. The company touts its AI strengths with logic and pattern-recognition capabilities, but transparency and auditability are areas where Wade sees a chance to respond to common worries.

“When you type something into ChatGPT and it gives you an answer, you don’t know why it came up with that answer,” Wade explained. “Even if you’re fine-tuning a model, you don’t really know why it did what it did. And neither does OpenAI or anyone else, even though they have amazing technology.”

Protective vs. opportunistic needs

He believes that can be a way to differentiate their approach while anticipating organizational needs to govern and explain what Kognitos’ platform is doing at the client level. Visibility, meanwhile, is a demand he sees coming from both protective and opportunistic needs.


“No bank’s gonna do mortgage applications if they can’t see why it’s doing what it’s doing,” he said, citing an example of protective motivations. “It’s too much of a liability, because if [the AI] hallucinates once, you’re going to have serious issues.” Hallucinations are instances where an AI model sees patterns or conditions that don’t reflect what a human user would see, resulting in factually incorrect output.

From an opportunistic perspective, Wade sees shortening paths to internal data and information as a challenge that AI can answer, both in terms of time and reduced numbers of steps. Those needs for speed and efficiency may extend beyond AI in retail, but it’s a sector where Kognitos is already in use.

“We’re already working with some customers on that where they can go back through all the runs and query, ‘What was my average margin on all orders in November?’” he said. The next step on the platform would be to “ask the simple question and actually query on that data because now it’s all logged in English in a way that’s very easy to query,” he explained.

Reducing technical work to plain-language interactions

“We have built an interpreter for English, which effectively makes English code,” Wade said. “Then we use large language models in different ways to enhance the user experience.”


LLMs are one of the most common types of deep learning models for AI in use now from Amazon, Google, Meta, OpenAI, and other tech companies. Wade stressed that the interpreter is the “core” of what Kognitos is offering. On top of that, they use multiple LLMs to generate output.

“It’s a combination of fine-tuning LLMs like [Meta’s] Llama,” he stated. “And then we also use things like [OpenAI’s] ChatGPT at times.”

According to Wade, this approach allows Kognitos to avoid becoming too dependent on one single LLM. Use cases for leveraging them, though, can include asking questions about fields in non-standard forms. Those could be purchase orders coming from different vendors. In some cases, another search system might encounter an error because it is not aware of one vendor’s format. Kognitos hopes to do better by helping non-technical users troubleshoot through conversation.

Examples of output

“[The Kognitos platform] would actually create an English question of ‘Hey, I can’t find this purchase order or this purchase order number. Can you help me?’” he said. “And then the business user can respond in free-flowing English saying, ‘For Amazon, the purchase order is always directly below the date.’”


The system could also create an output for a response to the customer, based on internally available data.

“The other way large models come in is there’s times when we want to generate content or we want to use large language models to do things like translation or context,” Wade said, talking through the process flow. “We call it Koncierge, but we’ll say, ‘Ask Koncierge to create an email based on a summary of the customer support ticket. And we’ll say ‘Use the model GPT4.’” Alternative LLMs such as Llama or Falcon would also be options.

Customer service

Customer-facing impact could also potentially come from more efficient call-center experiences. This is an area where Wade expects to see an intersection with AI in retail.

“I think you see a lot of this in retail and ecommerce as well — both have large contact centers,” he explained. “We’re actually doing a use case with a contact center where we are live-transcribing, or we’re connected with the transcription tool that is live-transcribing the call. It then goes in and writes a note and uses a large language model to create a summary of that conversation into Snowflake.”


This is where documentation comes into play, because the platform can refer to known standard operating procedures.

“That information and then taken and based on that summary, we look up a standard operating procedure of what the agent should do in that particular scenario,” he said. “Based on that standard operating procedure, it will then take action and create an email to send out to the customer with what actions are being taken as a follow-up.”

Fraud prevention

One other area of focus for Kognitos is fraud prevention. Wade discussed that use case in the context of a loyalty points project with a Fortune 500 consumer brand.

“They have a program where you take a picture of your receipt, and it shows their products on that receipt, and you get points,” he said. “The issue is receipts from every single gas station and every single grocery store all are different formats. It’s largely unstructured.”


Again, unstructured data is where the platform is intended to provide value.

“With Kognitos, though, we’re able to actually take that information, extract it from all the different receipts, classify it into an Excel format that then gets uploaded into their marketing software for market research, look at the things that they’re buying with those particular products, and from there even run fraud detection.”

The intended outcome of that process would be the elimination of duplicate receipts being submitted. That outcome clears up multiple problems on the client’s end.

“We actually use large language models and some custom models we build, and we can run it within the interpreter to put those fraud signals in and make sure they’re not paying out extra points in those scenarios or corrupting their market research data.”


Ultimately, that will be part of the interpreter’s true test. Its mission will be to turn those signals into actionable insights that are easy to understand for general business users. That will involve not just speaking to people, but understanding them as well.

“We’ve always been teaching people how to communicate with machines, which is really backwards,” Wade stated. “Instead, we needed to build it to where machines can understand people at the end of the day.”

Do you rank in our database?

Submit your data with this quick survey and we’ll see where you fit in our next ranking update.

Sign up

Stay on top of the latest developments in the ecommerce industry. Sign up for a complimentary subscription to Digital Commerce 360 Retail NewsFollow us on LinkedInTwitter and Facebook. Be the first to know when Digital Commerce 360 publishes news content.