🧩 The Messy Middle: Why AI in Retail Isn’t a Straight Line 

If you’ve ever tried something new, you’ll know it rarely — actually never — goes exactly according to plan. 

That’s as true for learning to ride a bike (yes, here we go again with the bike analogy) as it is for transforming a business. 

A typical “AI transformation journey” — including several I’ve personally been part of — usually looks something like this: 

👉 In PowerPoint: 

Data → AI → Automation → Value. 

👉 In reality: 

Data → Confusion → Alignment → Learning → Rework → Partial Value → Culture Shift → Real Value (maybe). 

The reason? Change is hard. 

Executives often underestimate the gravitational pull of “how things have always been done.” 

They overestimate how ready their data, processes, and people are. And they almost always hit a wall where the technology is no longer the problem — the organization is. 

💾 It starts with data… and humility 

Most “AI journeys” begin with a painful realization: your data is worse than you thought. 

Not because people have done a bad job — but because retail data was built for reporting, not for reasoning. 

You find product life cycles that don’t make sense, duplicate SKUs that no one noticed, inconsistent hierarchies, multiple “truths” for the same product, and systems that speak entirely different languages. 

And then there’s the data that simply isn’t used — and therefore never checked. 

But here’s the first important point: that’s not failure — that’s the starting point. 

The smart move isn’t to wait for perfect data. It’s to design systems that improve data as they learn. 

👉 Progress beats purity, every time. 

👥 Then come the people 

The second wave of resistance isn’t technical — it’s emotional. 

When AI enters the room, roles shift (or at least, they should): 

• Planners become validators. 

• Analysts become interpreters. 

• Managers become facilitators. 

And here’s where many companies miss the target entirely. 

If you blindly chase AI for its own sake, you’ll miss the mark. 

But if you forget about the people, you’ll miss it by even more. 

Just to illustrate: how many of you, in your last technology discussion, said something like “It needs to have AI”? 

I’ve personally seen plenty of RFX documents that simply ask, “Does your solution have AI?” — as if that single checkbox were the measure of intelligence. 

But what does that even mean? 

And when you finally get this “magical AI,” what happens? 

Your users hesitate to use it — because it feels like a black box, spitting out answers they don’t fully understand or control. 

That’s not innovation; that’s alienation. 

It’s not about building mysterious systems. 

It’s about using AI to enhance human reasoning — helping teams draw conclusions faster and connect dots they couldn’t see before. 

Just as you trust your Product Manager or Logistics Manager to make a call you don’t fully grasp, you need to extend that same trust to AI — within clear guardrails and a shared understanding of the goal. 

This is where it often gets uncomfortable. 

But it’s also where real transformation happens — when organizations stop asking, “Can we trust the machine?” and start asking, “Can the machine trust us?” 

• Because the quality of AI outcomes depends entirely on the quality of human input — the signals, context, and guardrails we provide. 

And here’s a practical step forward: 

👉 Start asking your technology providers how they leverage AI to support users, instead of simply, “Does it have AI?”. 

That one shift in thinking will put you on the right path. 

⚙️ Process before prediction 

Here’s another truth: most AI failures are process failures in disguise. 

If pricing, promotions, and supply decisions are still made in silos, it doesn’t matter how smart your models are. 

Even though AI can overcome certain gaps – AI isn’t a patch for bad process — it’s an amplifier.It scales whatever you already have, good or bad. 

Example: 

If replenishment runs on daily human overrides because pricing updates come too late, your “AI optimization” will simply automate inefficiency. 

Fix the timing, connect the data, and suddenly the same model delivers value. 

That’s why the “AI revolution” in retail won’t be won by data scientists alone. 

It’ll be won by companies that rethink their operating model around decisions, not departments. The next blog in this series will deep dive into this topic. 

🚧 Embrace the ugly phase 

The messy middle is where the real magic happens: 

• When the pilot that looked perfect in one category falls apart in another — and you realize the issue isn’t the model, it’s the assortment logic. 

→ Solution: Build adaptive models per category instead of a one-size-fits-all rollout. 

• When a planner overrules an AI recommendation — and the team takes the time to understand why. 

→ Solution: Treat overrides as insights, not errors. Feed that reasoning back into the system. 

• When IT, operations, and merchandising finally realize they’re solving the same problem from different angles. 

→ Solution: Establish shared KPIs around outcomes (availability, margin, waste) instead of function-specific metrics. 

That’s the phase most organizations try to skip — the one you have to survive before you scale. 

The difference between those who win with AI and those who don’t isn’t technology maturity. 

It’s organizational stamina. 

💡 The real question 

Everyone wants to know, “How fast can we get value?” 

The smarter follow-up question is: 

“How fast can we learn?” 

Because every failed pilot, every integration headache, and every uncomfortable meeting is part of the learning curve that separates hype from impact. 

The messy middle isn’t a detour. 

It is the journey. 

🏁 Final thought 

AI in retail isn’t a straight line from data to ROI. 

It’s a loop — a constant evolution of data, trust, and process. 

 The winners will be those who stop trying to skip the mess — and start mastering it.