“There’s nobody home. There is nobody on the other side. This is just a prediction engine that predicts what would make sense in this context, right? And then expectations are too high and the solution falls flat.” — Julian Krenge, Co-Founder & CPO, parcelLab
By the end of 2025, most retailers have deployed some form of AI chatbot. The technology has become table stakes: delivering decent answers for generic questions is no longer a fancy tech show-off but the baseline. Yet genuine advanced implementations remain vanishingly rare.
As our founder & CPO Julian Krenge puts it: “One in a thousand” retailers actually create opportunities to go beyond basic deployment. The rest are stuck in a loop of underwhelming results and growing frustration.
The expectation–reality gap is nothing short of Grand Canyon scale. Retailers deploy AI expecting their customer service ticket load to drop by a third. But no matter how hard they manifest, it doesn’t happen. Why? Because AI-generated answers leave so much room for confusion that tickets still get created. They’re just more narrowly scoped.
“My favorite example for this is you can get ChatGPT to a state where it tells you, ‘I cannot find out right now. Let me research a bit, and I’ll get back to you tomorrow.’ It will say this, but it will not do any research in the background. This does not exist. But people think this is real because it feels like it’s a human. What everybody still has to realize, ultimately with AI: there’s nobody home.”
A skilled customer service agent knows that when Cathryn asks, “Where is my order?”, she actually wants to know, “Will these clothes arrive before Christmas?” AI answers literally, not intelligently. And the tickets? They keep coming in, burying hopes and aspirations under tons of workload.
This creates a cascade of consequences. Management keeps knocking at all windows: The “do something with AI” choir intensifies, and often without any clear strategy attached. Many retailers have already cycled through numerous tools, each time hoping the next vendor will finally deliver. Investment piles up. Returns stay flat. Customer frustration persists despite the technology sitting right there on the website.
The root of all this evil? A false assumption. “If we just find the right tool, AI will work.” Retailers believe that switching vendors will eventually solve their performance issues. This belief is the fundamental mistake that keeps companies stuck — and it’s exactly where we need to start.
The expert: Julian Krenge on what actually works
If the problem isn’t the tool, what is it? Julian Krenge has spent years watching this pattern unfold. As Co-Founder and CPO at parcelLab, he works with hundreds of retailers across different sizes and verticals. He sees their AI implementations firsthand: the ones that fail, the ones that stall, and the rare ones that deliver (oh, how we wish to be them, huh?).
What sets his view apart is pattern recognition. Retailers like True Classic, YETI, and Pandora have all tackled AI implementation in different ways. Julian can point to specific decisions that made the difference: namely, concrete choices about data, scope, and evaluation. His conclusion? The problem is never the tool. It’s always the context and data infrastructure surrounding it.
The tool-switching trap
“What really keeps them stuck at stage zero is switching tools rather than enabling the tools. Fundamentally, even if you look at all of the benchmarks and the underlying AI models, ChatGPT and Gemini and Claude are identical. They are as good as the other. The limiting factor is how do you put in more knowledge, more context, so it can give the right answer.”
Retailers cycle through two, three, sometimes four AI providers, each migration accompanied by promises of better performance. And each time, the results plateau at the same mediocre level. We’re gonna hold your hand when we say this: The tool wasn’t the bottleneck. It never was.
The real issue: context and data.
“I think the biggest gap is how much time and effort still has to be spent on giving the AI the proper context, the proper information, the proper data, the proper access to really work on these.”
AI can only work with what it’s given. Without real-time shipping data, carrier information, and order history, even the most sophisticated language model is reduced to guessing. And here’s where it gets messy.
“The typical story for every retailer is that, yes, we have all this data here. And then you talk to another person at the same company and they tell you, yes, we have all the data. It’s always here — and it’s a different system.”
Everyone believes the data exists. It does. But, like the perfect turkey stuffing, everyone has their own special taste of handling format, size, and flavor. Email addresses formatted differently, duplicate customer profiles, legacy databases nobody wants to touch. This data chaos isn’t a minor inconvenience you serve to your grandparents on Thanksgiving. It’s the core obstacle to a functioning database. At some point, you must settle on a common recipe.
The 3 pillars of successful AI implementation
Success doesn’t require the newest model or the most expensive vendor. It requires a shift from tool-first thinking to context-first thinking.
Pillar 1: Context & data first
Before selecting any tool, start with a data audit. Identify your top five data sources needed for AI to perform.
Then ask:
Where does this data live?
What’s the quality like?
Can we access it in real-time?
Take basic WISMO queries, for example. The AI needs order status, package count, carrier information, delivery dates, dispatch status, and signature requirements. Six data sources for one use case. Add sizing help – e.g., product catalog and sizing charts – and complexity multiplies fast. Most retailers underestimate how much groundwork is needed before AI can deliver meaningful answers.
Pillar 2: Use case clarity
“AI in post-purchase works when you give it the right context and information, and have a clear understanding of the use cases that you want the AI to help with. What will not work is throwing AI in general at customer service.”
Most retailers think “AI in CX” equals “AI in Customer Support.” That’s narrow. They miss pre-purchase support (sizing, product guidance), automated communications (confirmations, dispatch notifications), and legal obligations (returns processing, recalls).
For each use case, ask: What’s the best case, worst case, and expected impact? How certain am I that we can actually pull this off?
Use cases with high potential but low confidence need to be discounted. Start with the highest-scoring use case with high confidence. Get it working. Then expand.
Pillar 3: Continuous evaluation
AI will never identify that it’s doing the wrong thing. If you want to know whether it’s performing, you have to check in — constantly.
Monitor thumbs-down feedback systematically. When customers signal dissatisfaction, trace failures back to their source. Was it missing data? Poor quality? A knowledge gap? Each analysis reveals what needs fixing.
One critical question any vendor should ask themselves in front of the mirror tonight: “How do I evaluate if the AI is performing?” If you can’t answer clearly, that should be a wake-up call.
Strategic principles for eCommerce and CX leaders
How do smaller retailers put this into practice without enterprise budgets? Quick answers beat slow answers, even if they’re less right.
A basic chatbot linking to your FAQ still reduces support volume. Deploy imperfect solutions. Don’t wait for the perfect day. It will never come. Fast iteration beats waiting for perfect implementations that never ship. Complexity determines feasibility.
A brand selling phone cases via Amazon? Basic FAQ automation might resolve 50% of WISMO queries. IKEA with its own delivery fleet? That same setup solves nothing. Know where your business sits on the complexity spectrum before deciding how deep to go. Don’t be dazzled by purple-pink gradient marketing.
“The industry, while there is a massive opportunity, is also a hype to a certain degree. Don’t be dazzled too much by shiny slides and nice demos. Those demos are always the best case scenario.”
Ask vendors for specific success metrics. If they can’t give concrete numbers, the slides are doing more work than the product. Reassess every six months.
“Whatever you thought six months ago, this doesn’t work? We tried something six months ago, put it into an AI, gave it a bunch of context. It didn’t work. You wait six months, take a newer model, it works.”
Build reassessment into your roadmap. The solution you dismissed last quarter might be viable today.
The bottom line
Retailers who keep failing share a pattern: they switch tools hoping the next vendor will solve their problems, deploy AI broadly without a data infrastructure, and evaluate once during implementation before moving on.
Retailers who succeed do the opposite. They invest in data integration before tool selection. They prioritize specific use cases. They treat AI supervision as ongoing work, not a launch-day checkbox.
The technology has matured: ChatGPT, Gemini, and Claude are essentially identical in capability. The competitive advantage lies in how well you provide context and data to whatever model you use. The retailers pulling ahead aren’t winning because they found a secret tool. They’re winning because they did the unglamorous work of connecting systems and cleaning data.
Before your next vendor meeting, skip the feature comparisons. Audit your data instead. Identify your top five data sources. Ask yourself: Can we get this to the AI in real-time? Is it clean enough to be useful?
That question will do more for your AI success than any tool comparison ever could.
Your questions, answered
AI chatbots fail because they lack proper context and data, not because of the tool itself. Most retailers deploy AI without integrating real-time order data, product catalogs, or customer history. The result: generic answers that don’t actually resolve customer issues.
No. Switching tools is the most common mistake retailers make. ChatGPT, Gemini, and Claude perform nearly identically in benchmarks. The difference lies in how well you feed them data and context, not which model you choose.
To resolve “Where is my order?” queries, AI needs access to six data sources: order status, package count, carrier information, estimated delivery dates, dispatch status, and signature requirements. Without real-time access to this data, AI can only link to generic tracking pages.
At Stage 1 maturity, where AI can look up real order data, retailers can expect a 20-30% reduction in WISMO inquiries. These are non-value-adding contacts that get resolved before ever reaching a human agent.
