Before ChatGPT, 5 Countries Already Ran on AI. What CX Leaders Can Learn

Customers ask questions to get problems solved. That is the obvious part.

But we all know that some problems with technology cannot always be solved on the spot — not even when the case moves from an AI to a human.

So, what customers often judge, more than the immediate fix, is how you work the problem: whether the right questions were asked, whether the issue was captured properly, whether someone owned the path forward. The process is sometimes more important than a solution that isn’t always there.

Hold that thought, because it explains why the current wave of AI in customer experience is being misread.

Here is the misreading. For most people, “AI” arrived in late 2022, when a chat box could suddenly hold a conversation. UBS analysts estimated that ChatGPT reached 100 million monthly users within roughly two months of launch — the fastest ramp for a consumer app they had seen in 20 years of following the internet. Claude and others followed, and the perception took hold: AI is new, and we are all using it for the first time.

That perception is wrong. What changed in 2022 was not the arrival of AI but its visibility. As even Meta’s chief AI scientist noted at the time, the underlying technology was not a fundamental breakthrough. AI had already been running quietly inside the products, factories, hospitals and cars that businesses and consumers relied on for years — mostly in B2B, mostly invisible. The consumer chat moment simply pulled back the curtain.

FAQ: AI’s History in Customer Experience Before Generative Chat

Editor’s note: These questions address how enterprise AI operated in customer experience, healthcare and manufacturing years before ChatGPT and Claude made generative AI a mainstream topic.

Where AI Was Already Running Before ChatGPT Existed

Long before anyone typed a prompt, artificial intelligence was doing unglamorous, high-stakes work behind the scenes. It sat inside driver-assistance cameras, factory robots, enterprise software, medical scanners and the phone in your pocket. Customers experienced the results — a safer lane change, a faster diagnosis, a resolved support ticket — without ever thinking of it as “using AI.” That gap between experience and perception is the whole story, and it played out across the world’s major technology economies.

South Korea: The Assistant Already in Your Pocket

In March 2017, Samsung introduced Bixby alongside the Galaxy S8 — a voice assistant using natural-language understanding to control apps, translate, and recognize objects through the camera. Millions of people carried an AI assistant for years and simply called it “my phone.” The intelligence was real; the label was absent.

Japan: Intelligence on the Factory Floor and in the Lab

In November 2015, Toyota announced a one-billion-dollar investment to establish the Toyota Research Institute, an AI and robotics research company that began operating in January 2016. Japan’s industrial base — robotics, automotive, manufacturing — had been treating machine learning as an engineering tool for improving safety and productivity long before it became a consumer talking point. The customer felt it as a better, safer product, not as “AI.”

The United States: AI Trusted With Life-and-Death Decisions

Nowhere is the pre-chat maturity of AI clearer than in medical devices, where algorithms were trusted with life-critical dosing years before anyone worried about a chatbot’s tone. The turning point came in September 2016, when the FDA approved Medtronic’s MiniMed 670G — the world’s first hybrid closed-loop insulin system. Its SmartGuard algorithm read glucose every five minutes and automatically adjusted basal insulin, designed to learn an individual’s needs and act to minimize highs and lows.

And the technology never stood still: the current-generation MiniMed 780G (FDA-approved April 2023) adds meal-detection technology and delivers automatic insulin corrections every five minutes — even compensating when a user forgets to bolus or underestimates their carbs. This is adaptive AI making autonomous physiological decisions, on the body, around the clock — and getting materially smarter with each generation.

The approach also diversified. In December 2019, Tandem Diabetes Care’s Control-IQ became the first interoperable automated insulin-dosing algorithm — and it works differently, using continuous-glucose data to predict where blood sugar will be 30 minutes ahead and delivering correction doses before a high or low arrives. One family of systems learns and reacts; the other forecasts and pre-empts — distinct flavors of AI, both entrusted with a patient’s life, both years before the consumer AI moment.

The same trust extended to diagnosis. In April 2018, the FDA authorized IDx-DR, the first autonomous AI system cleared to make a diagnostic decision in any field of medicine — detecting diabetic retinopathy from retinal images without a clinician interpreting the result, at roughly 87% sensitivity and 90% specificity in its pivotal trial. Society trusted AI with a diagnosis, and with insulin dosing, long before it trusted it to draft an email.

Israel: The AI That Watches the Road

Founded in Jerusalem in 1999, Mobileye built computer-vision systems that let a single camera detect vehicles, lanes, and pedestrians. By 2016 its EyeQ chips were installed in roughly 16 million vehicles, and in 2017 Intel acquired the company for about 15 billion dollars. Every driver who received a collision warning was relying on decades-old machine-vision AI — and almost none of them called it that.

Germany: AI Inside the Enterprise and the Plant

In January 2017, SAP launched Leonardo, a machine-learning foundation embedded across enterprise applications. Tellingly for CX, one of its early services — Service Ticket Intelligence — already categorized customer tickets automatically and proposed solutions. Alongside it, Siemens built MindSphere to apply AI and machine learning to industrial equipment. Enterprise AI was resolving service cases in 2017 — the same task everyone thinks generative AI invented.

How Accurate Was the FDA-Cleared IDx-DR Diagnostic System?

IDx-DR, authorized by the FDA in April 2018 as the first autonomous AI diagnostic system in any medical field, detected diabetic retinopathy from retinal images with roughly 87% sensitivity and 90% specificity in its pivotal trial.

Related Article: Human First, AI Smart: The Customer Experience Balance for 2026

How AI in Customer Experience Evolved From Embedded to Agentic

Seen across time, the pattern is clear: AI moved from invisible infrastructure toward visible, conversational presence — and is now moving again, toward autonomous action inside the enterprise. Each phase was real; only the last one felt like “AI” to the average customer.

Era	Phase	What the customer actually experienced
Pre-2010	Embedded & statistical AI	Recommendations, fraud checks, speech recognition, call routing — unseen.
2010–2016	Industrial, perceptual & clinical AI	Driver-assistance cameras, factory robotics, and the first autonomous insulin dosing (Medtronic 670G, 2016).
2017–2019	AI goes ambient & clinical	Voice assistants (Bixby), enterprise ML (Leonardo), autonomous diagnosis (IDx-DR) and predictive dosing (Control-IQ).
2022–2023	The perception shift	Generative chat (ChatGPT, Claude) makes “using AI” a conscious, everyday act.
2024→	Agentic & orchestrated AI	AI that acts, not just answers — inside B2B workflows and customer journeys.

What Year Did AI Move From Ambient to Agentic, According to This Timeline?

The timeline marks 2024 onward as the shift to agentic and orchestrated AI, when systems begin acting inside B2B workflows and customer journeys rather than only answering questions.

Why the Chatbot Is Actually the Simplest AI Application

Here is what makes the perception gap almost paradoxical. The thing that finally convinced the public that “AI is here” — a company chatbot answering questions — is, in engineering terms, one of the most basic applications of these models. It is pattern-matching over text: retrieve, phrase, respond. Compare that to an algorithm that decides how much insulin to release into a living body every five minutes, a vision system that distinguishes a pedestrian from a shadow at highway speed, or a model that reads a retina and rules on disease. Those carry real-world, irreversible consequences and demand far more of the technology.

A customer service bot is the visible tip of a very deep iceberg. The same underlying advances in machine learning power fraud detection, demand forecasting, predictive maintenance, drug discovery, logistics optimization and the autonomous decisions inside the devices above. Mistaking the chatbot for the whole of AI is like mistaking a light switch for the power grid.

The interface is simple; what stands behind it — and what came before it — is not.

View All

Why Is a Customer-Service Chatbot Considered ‘Simple’ AI?

A chatbot performs basic pattern-matching over text — retrieve, phrase, respond — while the same underlying machine learning also powers far higher-stakes systems like autonomous insulin dosing and retinal disease diagnosis.

Why B2B Adopted AI Years Before Consumers Noticed

B2B lived with AI first because the stakes and the data were there: a factory, a hospital, a fleet of cars, an enterprise support queue; the discipline of shaping how a system talks to a person grew up in this era, too, through interactive voice response, early chatbots, and voice assistants, well before generative models made dialogue fluent. The craft is not new; only its raw material improved.

That history carries a practical lesson for anyone building AI into customer experience today. Customers were never really reacting to the technology. They were reacting to whether their problem got understood and owned. The generative wave makes AI feel new and conversational, but it does not change the underlying test — and it raises the stakes, because a fluent answer can hide a hollow one.

How Did Conversation Design Predate Generative AI in B2B?

Conversation design — the discipline of shaping how a system talks to a person — developed through interactive voice response, early chatbots and voice assistants well before generative models made dialogue fluent.

What’s Next for Agentic AI in B2B Customer Experience?

The next phase is agentic: AI that does not just answer but takes action inside workflows — opening cases, processing claims, updating records. In customer experience, that shifts the human’s role toward the ambiguous, high-stakes, judgment-heavy moments the machine should not close alone. The organizations that win will not be the ones that deploy the most AI; they will be the ones that decide, in advance, what the AI may resolve and where a human must own the outcome.

What Should Organizations Decide Before Deploying Agentic AI?

The article argues that winning organizations decide in advance which workflow actions AI may resolve and where a human must own high-stakes, ambiguous or irreversible outcomes.

How Enterprise CX Leaders Should Build AI Into Operations

The history carries a hard lesson for anyone running a company today. If AI has quietly created value for decades, why do so many current projects fail to? The evidence is blunt: MIT’s State of AI in Business 2025 found that roughly 95% of enterprise generative-AI pilots deliver no measurable impact on profit and loss, and McKinsey reports that only about a fifth of organizations have redesigned even a single workflow around AI — most simply layer it on top of processes built for a pre-AI world. The winners are not the ones with the best model; they are the ones who rewired how work flows.

The lessons below are drawn from that divide, and they apply well beyond customer experience and service — to operations, finance, supply chain, design, processes, engineering and R&D alike.

• Fix the process before you automate it. Automating a broken process only lets you do the wrong thing faster. Map how work actually flows today, remove the redundant steps, then apply AI to what remains. The value comes from redesigning the workflow end-to-end, not from bolting intelligence onto a legacy path.
• Aim AI at the back office, not just the shop window. MIT found the biggest returns in unglamorous operations — reconciling invoices, forecasting demand, screening documents, scheduling maintenance — while the money poured into flashy sales and marketing bots delivered the least. Follow the friction and the cost, not the visibility.
• Set an outcome target, not an AI target. “Deploy AI” is not a goal; “cut invoice-processing time by 40%” or “reduce unplanned downtime by 20%” is. Tie every initiative to a specific operational or financial number, and kill the ones that cannot name theirs.
• Treat data as the asset that compounds. Models are becoming commodities; your proprietary, well-governed data is not. Most failed pilots fail on dirty, disconnected, or unusable data long before the model is at fault. Invest in clean, connected, permissioned data first — it is the foundation everything else stands on.
• Buy and partner more than you build. Externally sourced systems reached production roughly twice as often as internal builds, because specialist partners bring the scars of dozens of prior implementations. Reserve scarce in-house engineering for what is genuinely proprietary; buy the rest and integrate it well.
• Design the human-in-the-loop by decision, not by default. High performers are far more likely to have defined, in advance, exactly which outputs a machine may finalize and which require human validation. Decide where a person must own the call — high-stakes, ambiguous, irreversible — and build that checkpoint into the system rather than discovering it after a costly error.
• Remember the rewiring is 80% organizational. The technology is the smaller part; governance, incentives, training, and change management are the larger one. Put a senior owner on AI governance, retrain people for the tasks that remain, and communicate wins internally. A pilot that ignores the operating model will stall no matter how good the model is.

None of this is unique to one industry. Whether the output is a diagnosis, a delivery route, a financial close, or a resolved customer cases, the pattern is identical: the organizations that benefit are the ones that redesign the work around the intelligence, own their data and decide deliberately where humans stay in charge.

Which returns us to where we started. When a problem cannot be solved immediately — and some cannot — what the customer remembers is the quality of the process: the right questions, the issue captured accurately, a human who owned it rather than passing it on. AI has been quietly improving that process for over two decades. The task now is not to be dazzled by how new it suddenly feels, but to design it so the moments that require ownership still find a human ready to give it.

Learn how you can join our contributor community.