Anyone who has shopped online in the last few years knows the drill. A chat window opens before you’ve even started browsing. A scarcity banner flashes: “only two left.” A pop-up asks for your email in exchange for a discount you didn’t request.
Real people shop in fragments, preferences, half-formed ideas, and even a blend of languages.
For years, the industry has been geared toward optimizing clicks and basket size, often at the expense of understanding what customers actually want. It works, sometimes. But it’s also exhausting. Now, as privacy expectations rise and shoppers grow weary of pushy AI bots, nearly 45 percent abandon their purchases before making it to the checkout.
What if there was a better way to do business? One that prioritizes intent modelling and ethical “pull-based” design over those aggressive (and annoying) “push” triggers. Toronto-based cloud, AI, and DevOps company Dedicatted is betting on that approach. Its bilingual assistant for Kuwaiti e-commerce platform Taw9eel lifted average order value by six percent, sped up product discovery, and improved conversion rates—proof, the company states, that sales and customer trust go hand in hand.
From keywords to intent
Real people shop in fragments, preferences, half-formed ideas, and even a blend of languages.
That messy mismatch between how shoppers actually behave and how most bots process language—is what Dedicatted set out to design around in partnership with Taw9eel, said Serhii Semenchenko, the company’s Chief Technology Officer.
Traditional chatbots run on fixed scripts, with little room to adapt. “You have specific steps, and based on that, it provides some answers,” he explained. The Taw9eel bot was built to do the opposite. It reads intent, weighs context, and pulls shoppers to their destination, rather than pushing.
“It acts more like a proxy that helps you find what’s right for you,” he explained.
“It acts more like a proxy that helps you find what’s right for you.”
Consider someone buying a TV. A shopper might know the room is bright in the early evening, that they sit about three metres from the screen, and that they like a certain brand or style. They might switch languages, misspell words, or change their mind halfway through a sentence.
Traditional systems struggled with this kind of complexity, said Semenchenko. Dedicatted’s next-gen assistant, meanwhile, is designed to figure out what matters—whether, in the case of the television purchase, that’s brightness level for glare or screen size for viewing distance—and narrow from there. If the request is too vague, it follows up with questions instead of dumping a long list of items and hoping one sticks.
Under the hood, that experience is the result of semantic intent mapping, said Semenchenko. Instead of matching keywords, the system converts language into tokens and vectors, a numerical representation of meaning, to interpret the request. For the user, it feels seamless: you explain what you need, and the system works out the details.
The same architecture enables it to move smoothly across languages. Taw9eel shoppers routinely switch between Arabic and English in a single sentence. That kind of code-switching can trip up most systems. An intent-driven system reads between the lines and works out what the shopper means.
“The idea is that the specific word in Arabic and in English converts to pretty similar tokens,” Semenchenko said. “The LLM works with tokens, not with real text.”
Building trust over pressure
Every design choice on Taw9eel comes back to the same variable: trust. Without it, no amount of intent modeling or clever UX will bring a shopper back, according to the company.
Image courtesy Dedicatted.
The team tested that theory on the platform, measuring the familiar pressure tactics against their absence. Removing things like urgency prompts, countdown timers, and scarcity banners barely made an impact on sales. Ratings and review counts, on the other hand, dropped site performance quickly. Shoppers rely on those verification signals to make a decision, regardless of whether they arrive at a product through the chatbot or the classic way. When those signals are missing, they’ll go looking for verification and reassurance elsewhere, “and you lose momentum,” said Semenchenko.
The same trust logic shapes how Dedicatted manages data privacy. Most companies assume they need deep personal data to deliver a personalized experience. Taw9eel’s assistant draws on anonymized patterns rather than personal data, like the broad signal that people who buy bread also tend to buy butter. That means customers wary about giving away personal information can use it without wondering what the site is storing about them, said Semenchenko.
Trust also shapes how the assistant manages its own limits. Rather than returning the wrong answer, the assistant is designed to measure its own response confidence. When that score drops below a set threshold, the conversation hands off to a human with the full context intact. Critically, the shopper doesn’t have to start over or rephrase the question, which, over time, builds further trust in the process, said Semenchenko.
Dedicatted uses similar systems in financial workflows where documents require human review before a decision is finalized, and in internal tools where employees are navigating hundreds of pages of policy. The goal is always the same, said Semenchenko. “Not just to have AI because of AI, but to provide some real ROI and really streamline some manual processes.”
PRESENTED BY
Curious what pull-based AI looks like in practice? Read Dedicatted’s full Taw9eel case study
Feature image courtesy Dedicatted