When we work on data-related products, reviewing policies for storing and managing personal data is just part of the process. It is almost instinctive for GDPR to come up in the first meeting. Data residency gets flagged before a single line of code is written.

But for AI? We don’t ask the question once.

Why have we not thought about checking the policies that govern AI? I wondered.

That realisation sent me down a path to gather knowledge on the AI policies currently defined around the world, not just to enlighten myself, but to bring my team and you along with me.

Most businesses are building AI products in one conversation, and regulators are having a completely different one.

This series is my attempt to put you in the room for both.

Before you read on, here is what to expect from this series:

AI regulation is a global story, and there is no way to do it justice in a single article without either skimming the surface or losing you entirely. So I am not going to try. This is a 3-part series: Part 1 covers the European Union (this one), Part 2 the United States, and Part 3 covers China, the UK, and the rest of the world.

🌍 Why This Matters More Than You Think

AI regulation is not just a compliance exercise; it is a product and business decision. It answers one question that not enough teams are asking: what are you allowed to build with AI? By early this year, over 70 countries have launched or signalled AI policies. Some carry heavy financial penalties. Others read more like voluntary frameworks today, until a regulator decides to make an example, and suddenly they are not.

If your product touches customers in multiple countries, this concerns you too; you are already operating across multiple regulatory environments. Understanding this could save you from fines you never saw coming.

That said, let’s dive in...

The best place to start is also the strictest room in the building, the European Union (EU). The EU AI Act, which is the world's first comprehensive legal framework for AI, entered into force in August 2024. It is rolling out in phases through 2026 and beyond with a series of deadlines arriving one after another until the whole framework is live.

Here is what has already taken effect:

In February 2025, the first phase of the policy kicked in. AI systems classified as “unacceptable risk” became illegal across all EU member states. For businesses, this means several things you might already be doing are now prohibited.

The list includes AI systems that:

1. Use manipulative or deceptive techniques to change user behaviour

2. Exploit vulnerabilities based on age, disability, or financial situation

3. Offer real-time biometric surveillance in public spaces (with very narrow law enforcement exceptions).

4. Predict or detect emotions in the workplace or educational settings.

That last one deserves a pause, because this means if your HR platform, productivity tool, or workforce management software uses any form of emotion-sensing or behavioural inference, even something as seemingly harmless as “engagement scoring”, you are in legally complex territory in the EU.

In August 2025, General-Purpose AI models came under scope. This is the part that matters for any team building with large language models. Providers of foundational models such as GPTs, Claudes and Geminis that power most AI products today must now comply with specific transparency and documentation obligations:

1. Technical documentation (Model architecture, input/output capabilities, training process, its intended and acceptable use cases, etc).

2. Summary of their training data

3. Demonstrate copyright compliance for that data

4. Share information with both regulators and any business building on top of their model

If your product is built on top of one of these models and you operate in the EU, that compliance chain flows through to you.

Now, for what is heading your way before the end of this year.

By August 2026, the comprehensive requirements for high-risk AI systems will become enforceable. High-risk covers a significant amount of what businesses are actively building and shipping today, including AI used in hiring and employment decisions, credit scoring, access to essential services, education, healthcare, and critical infrastructure.

If your product falls into any of those categories, the important question is “what does compliance actually require of you?” By August 2026, you will need to have the following in place before your AI system can legally operate in the EU:

1. Organisations must have a documented, maintained, and up-to-date risk management system throughout the product's lifecycle.

2. There must be evidence that the data used to train or run your AI system is relevant, representative, and free from errors that could lead to a discriminatory outcome

3. There must be a paper trail of how your system was built, what it was designed to do, how it was tested and what limitations it has.

4. AI systems must be designed in a way that a human can intervene, override or shut them down. Any system with no human in the loop is not compliant.

5. Every action your AI system takes needs to be traceable with a log kept for a minimum of six months

6. AI systems must be formally registered in the EU database before it is deployed

7. Specific to some cases, an independent third-party verification of your system might be needed before the AI system can go to the market.

Something to note, this does not just apply to companies based in the EU. If your AI system produces output that affects people in the EU, even if your servers, your team and business are outside Europe, you are in this scope, and the penalties are not theoretical. The penalty framework has been live since February 2025; companies using prohibited AI practices can be fined up to €35 million or 7% of their global annual turnover, whichever is higher.

What this means for your business

If you have users or customers in the EU, you need to classify every AI system you use or build against the EU’s risk framework. Start with a simple audit. Ask questions like: Does this system make decisions that affect people’s lives, livelihoods, or access to services? If the answer is yes, you are likely in high-risk territory. If the answer is maybe, you need a lawyer.

This information is not to scare you, but to empower product teams to build actively. The businesses that will navigate this well are not the ones spending the most time asking, "What is the minimum we need to do to comply?" They are the ones asking: "What does responsible use of AI actually look like for us, and how do we build that into the product from the start?"

Next week, we are heading to the United States, where there is no single federal AI law; 50 states are each writing their own rules, and the federal government is actively trying to stop them. It is messy, it is fast-moving, and it affects any business with US customers.

See you next week.

Disclaimer: This article is for informational purposes only and does not constitute legal advice. If you are building AI products across multiple jurisdictions, please consult qualified legal counsel.

👋 Just joining us? Last week, we mapped out the differences between Chatbots, AI Copilots, and AI Agents and why deploying the wrong one has real consequences. This week, we put chatbots under the microscope.

Here’s something nobody in the room ever wants to admit when an AI chatbot project goes sideways.

Come closer. I’ll whisper.

🤫 ...the AI wasn’t the problem.

The technology works, and the models are capable. What actually breaks is everything around the AI, the design decisions, system integrations, knowledge management, guardrails or, more accurately, the absence of all of the above.

We have spent years blaming the bot when the real culprits were sitting in the planning meeting the whole time.

This week, let’s walk through the four failure patterns I see most often in CX chatbot deployments, with real examples, honest explanations of why they happen, and what good actually looks like.

The Hard Truth First

Before we get into the failure modes, let’s establish one thing clearly: a chatbot is only as good as the decisions made before it goes live.

A poorly designed chatbot doesn’t just fail to help, it actively drives customers away. It destroys the trust that your customers have in your business or brand. And because it operates at scale, it can damage thousands of customer relationships simultaneously before anyone notices the pattern.

A well-optimised chatbot, on the other hand, is one of the most powerful tools in a CX team’s arsenal. The difference between the two is rarely the AI model. It’s almost always the execution.

“The chatbot isn’t broken. The thinking behind it is.”

With that said, let’s get into it. Four failure modes, four honest dissections.

Failure Mode 01 🔁 The Endless Loop

⚠ High impact

Picture this. A customer contacts your chatbot with a billing dispute. The bot doesn’t recognise the specific issue, so it serves a generic response. The customer rephrases. Same response. They try again with different words, same response. By the fourth attempt, they’re furious, and they still haven’t spoken to a human.

This is the escalation problem. Most chatbots are built with enormous focus on what happens when things go right, and almost no focus on what happens when things go wrong. The result is a bot that keeps looping canned responses instead of recognising that the conversation has hit a wall.

The fix isn’t complicated, it’s just rarely prioritised. Escalation should be treated as a first-class feature, not an afterthought. Customers should never have to fight for access to a human. The bot should be doing that work for them, automatically, when the signals are clear.

✦ Here's what to do instead

• Implement sentiment analysis to detect rising frustration and trigger escalation automatically

• Escalate after two or three failed resolution attempts, don’t wait for the customer to ask

• When a customer explicitly requests a human, that request should be honoured immediately, every time

• Warm handoffs pass the full conversation context to the agent so the customer never has to repeat themselves

Failure Mode 02 🏝️ The Island Bot

⚠ Medium — High impact

I’ll sound like a broken record here, and honestly, I’m fine with that. This point plays on a loop in my head because I keep watching teams make the same mistake and then act surprised by the results.

A chatbot operating in isolation is a chatbot that can only ever be generic. You have customer data sitting in your CRM. Order history in your management system. Account details in your billing platform. And your chatbot has no idea any of it exists.

So when a customer asks, “Where is my order?” the bot serves a generic tracking page link. When they ask about their account status, it draws a blank. It knows nothing about the person it’s talking to, not their history, not their tier, not their last interaction. Every conversation starts from zero.

Generic in, generic out. On repeat. Every single time.

Teams launch chatbots as standalone tools, disconnected from the systems that hold all the context, and then wonder why the responses feel hollow.

The gap isn’t the AI. It’s the architecture around it.

✦ Here's what to do instead

• The fix is straightforward even if the implementation isn't, connect your chatbot to your data systems. Your CRM, order management, billing platform, and knowledge base should all be feeding context into every conversation. A bot with access to the full picture will always outperform one working in the dark.

Failure Mode 03 📚 The Stale Knowledge Problem

⚠ Medium impact

A chatbot is only as accurate as what it’s been taught. Feed it an incomplete knowledge base, and it fills the gaps with guesswork. Feed it an outdated one, and it confidently serves customers information that is no longer true. Both are damaging. The second one is often more so, because the customer has no reason to doubt a confident, well-worded wrong answer.

This failure is especially common in organisations that treat the knowledge base as a launch asset rather than a living document. Products change. Policies evolve. Pricing updates. The chatbot doesn’t know this unless someone tells it.

The result? Fragmented, unreliable responses. And customers who stop trusting the bot entirely after one bad experience.

✦ Here’s what to do instead

• Treat the knowledge base as a live product, assign ownership and a regular review cadence

• Run the bot against real customer scenarios regularly to catch gaps before customers do

• Mine failed or low-confidence interactions to identify what the knowledge base is missing

Failure Mode 04 — The Big One

🎭 Hallucination & The Governance Gap

🔴 Critical

This is the one that keeps CX leaders up at night, and rightly so. AI hallucination is when a model generates a response that sounds completely confident and coherent, but is factually wrong. The AI doesn’t know it’s wrong. It just sounds right.

Research puts hallucination rates somewhere between under 5% for straightforward questions and over 25% in complex, multi-step scenarios. Even retrieval-augmented systems (RAG), the ones designed to ground the AI in real documents, can fail if they pull stale as we discussed earlier or irrelevant content.

In customer service context, this isn’t just embarrassing. It’s legally and financially dangerous. Regulators in multiple markets are now treating AI-generated misinformation as a potential deceptive business practice. If your chatbot misleads a customer, even unintentionally, your company can be held responsible.

The uncomfortable truth about hallucination is that there is no perfect solution. But there is a meaningful difference between organisations that have thought seriously about this risk and those that haven't.

✦ Here’s what to do instead

• Deploy an assurance layer i.e a system that provides visibility into who the bot is talking to, why, and when

• Build detection for when answers deviate from approved, verified sources

• Establish clear compliance boundaries i.e what the bot is and is not permitted to discuss

• Log and review low-confidence responses regularly, not just customer complaints

• Treat prompt injection as a real attack vector and test for it before launch

Before You Launch — A Quick Honest Check

If you are about to deploy a chatbot, or if you have one live and are wondering why it’s underperforming, run it against this list. Not as a formality but as a genuine diagnostic.

The deployment health check

🔁Does your bot have a clear, automatic escalation path, or does it loop?

🔗Is it connected to your data systems?

📅When was the knowledge base last reviewed and updated?

👤Does the bot personalise responses using customer history and context?

🛡️Do you have a governance layer monitoring for hallucinations and prompt injection?

📊Are you reviewing failed interactions regularly, not just overall satisfaction scores?

If any of those feel shaky, that’s where to focus. Not on the AI model. Not on the interface design. The fundamentals above will move the needle far more than switching to a newer model ever will.

Chatbots are not failing because AI is overhyped. They’re failing because we keep treating the fundamental problem as a technology problem when it’s actually a service design problem.

The best chatbot aren’t the most technically sophisticated. But the most thoughtfully built, with real escalation paths, real integrations, a live knowledge base, and a team that actually reviewed what went wrong regularly.

That’s it. That’s the secret.

That’s my view for the week. See you next week 😉

I had to pause and say: look, I know they all run on AI, I know they all feel like “talking to a machine”, but they are doing fundamentally different jobs.

And once I walked through the distinction, something clicked. The confusion wasn’t a knowledge gap. It was a framing gap. Nobody had ever told them the jobs were different.

That conversation stuck with me because if someone close to the technology can mix these up, then the confusion is everywhere. And it matters. Choosing the wrong tool doesn’t just waste money. It sets the wrong expectations, frustrates teams, and leaves real value on the table.

So here’s the clearest breakdown I can give you: three tools, three very different jobs, and why understanding the difference will change how you think about AI at work.

The Chatbot: The One That Talks

Chatbots have been around longer than most people realise. At their core, they are programs designed to simulate conversation, written or spoken. The humble chatbot on a retail website asking, “Can I help you find something?” is one. So is the voice that picks up when you call your bank.

But not all chatbots are created equal. There’s a whole spectrum:

• Menu-based chatbots are the simplest kind; they present buttons and follow a decision tree. Click “Billing,” then “Make a payment,” and off you go. Great for transactional tasks, not great for anything else.

• Rule-based chatbots work with “if this, then that” logic. They can handle a wider range of questions, but only if someone has already predicted those questions. Anything outside the script and they freeze up.

• AI-powered chatbots use natural language understanding (NLU) to interpret what you’re actually saying — not just looking for keywords. They can ask clarifying questions. They can handle ambiguity. They remember context within the conversation.

• Hybrid chatbots sit between the rule-based and AI tiers, combining structured menu logic with machine learning capabilities. The menu handles predictable, high-volume interactions cleanly; the AI layer picks up where the script runs out. The result is a bot that is reliable enough for routine tasks and flexible enough for messier ones. In practice, this is what many production chatbots look like today.

• Generative AI chatbots are the newest tier. They don’t just retrieve pre-written answers — they generate new responses, adapt to your tone, and can create content on the fly.

It’s worth pausing on the hybrid model for a moment, because it’s easy to skip over. There’s a temptation to treat this spectrum as a straight upgrade path — as if everyone should race straight to generative AI. But structure has real value. A well-designed hybrid gives businesses the speed and predictability of rules where they need it, and the intelligence of AI where they don’t. That balance often matters more than chasing the most advanced option available.

The chatbot’s job is to steer the conversation. It answers, guides, and hands off. It is reactive. It waits for you to speak first and responds from there.

Think of a chatbot as a highly knowledgeable receptionist. Brilliant at answering questions, terrible at doing your tax returns.

The AI Copilot: The One That Helps You Work

Now we move up a level. The AI Copilot isn’t just having a conversation with you, it’s sitting inside the tools you already use, watching what you’re doing, and helping you do it better.

The name is deliberate. Like a copilot in an aeroplane, it doesn’t take the controls, you do. But it’s watching the instruments, flagging anomalies, and handing you the right information at the right moment so you can make better decisions faster.

A copilot is a conversational interface built on large language models (LLMs) that pulls context from across your software ecosystem, your emails, your CRM, your documents, your data and uses that to guide and assist you in real time.

In practice, copilots are the things that:

• Draft that follow-up email based on your meeting notes

• Summarise a 40-page report into three bullet points before your 9 am

• Spot that a customer account is showing churn signals before your sales rep notices

• Suggest the next field to fill in a CRM form based on what you’ve already typed

Crucially, the copilot doesn’t do things for you without asking. It suggests. It drafts. It highlights. You stay in the driver’s seat; it just makes the drive smoother.

“An AI copilot isn’t meant to replace people, it’s there to help your team find answers quickly, avoid redundant tasks, and make smarter decisions.”

For customer experience teams, this is transformative. A support agent with a copilot can triage complex tickets faster, access a customer’s full history in seconds, and get real-time suggestions on how to resolve an issue, without ever leaving their inbox.

The AI Agent: The One That Gets Things Done

And now we get to the one that’s genuinely changing the game.

An AI agent is not a conversational tool. It’s an operational one. It’s a system designed to complete goals autonomously, across multiple steps, using whatever tools and data it needs to get there.

Where a chatbot waits to be asked, and a copilot helps you act, an AI agent acts. It can break a complex task into subtasks, execute each one, track what worked and what didn’t, and adjust along the way, all without a human holding its hand through every step.

Here’s what that looks like in practice. You tell an AI agent: “Process all new customer onboarding requests that came in this week.” The agent:

• Reads the intake forms

• Cross-references your CRM for existing customer data

• Creates accounts, assigns reps, and sends welcome emails

• Flags any edge cases that need human review

• Updates your dashboard when done

You didn’t click a single button. The agent planned, executed, and reported back.

This is possible because agents combine several capabilities at once: they use LLMs to reason and plan; they call external tools and APIs to take real actions; they retain memory across steps to maintain context; and they apply guardrails, permissions, approval flows, and validation checks to stay within safe boundaries.

If a chatbot is a receptionist and a copilot is a highly capable colleague, an AI agent is the department that runs itself checking in with you only when it genuinely needs a judgment call.

Which One Do You Actually Need?

If you’re figuring out where to start, here’s a rough guide based on your team and use case:

• Support, IT helpdesk, customer experience at scale: Start with a well-designed AI chatbot. Handle volume first.

• Sales, analysts, operations, engineering: A copilot will dramatically speed up your team’s existing workflows without changing how they work.

• Finance, compliance, back-office automation, complex multi-step ops: This is agent territory. The ROI is real, but so is the implementation work.

Chatbots talk. Copilots assist. Agents execute.

The future of customer experience will be built by the teams who understand that difference, and deploy accordingly.

That’s my view for the week. See you next week 😉

I’m not proud of that. But I suspect I’m not alone.

Over the years, I’ve come to realise that Customer Experience is so much bigger than support tickets, success metrics, or NPS scores. It’s the entire feeling a customer has every time they interact with your brand — from the first ad they scroll past to the third time they email your support team about the same problem (we’ve all been there).

And right now? We’re in the middle of the biggest transformation this industry has ever seen.

So let’s talk about it.

🕰️ A Brief History of CX (Bear With Me. This Part Actually Matters).

To understand where we're going, we need to understand how we got here. Because AI doesn't create a new foundation — it amplifies the one that already exists.

📠 The 1980s–90s: Efficiency First. Early CX was basically a stopwatch exercise. How fast can we answer the phone? How quickly can we close a ticket? Metrics like wait times, handle times, and first-contact resolution (FCR) ruled the day. The goal was managing operational service efficiently, not crafting an experience.

🌐 The 2000s: Digital Arrives (And Things Get Interesting). The internet changed everything. Suddenly, customers had options to contact with businesses via email, websites, Social media, Live chat, e-commerce and businesses had to follow them. This era gave us customer journey mapping and the Net Promoter Score (NPS). For the first time, companies started asking: What does our customer actually go through? A shift from service management to experience management.

📱 The 2010s–2020s: Welcome to the Experience Economy. This is where CX really grew up. Amazon taught us that one-click ordering could become a standard expectation. Netflix taught us that a personalised homepage is table stakes. Spotify taught us that a platform that knows your mood feels like magic. Data-driven, personalised experiences weren’t a competitive advantage anymore; they were the minimum bar.

🤖 2020s–Present: The AI Era CX 2.0. And here we are. Real-time insights. Emotional analysis. Predictive, proactive service. The platforms that once helped businesses respond to customers are now helping them anticipate customers. The shift from reactive to predictive is genuinely massive.

🚀 So What’s Actually Changed?

Here’s the thing about AI in CX: it’s not just making old things faster. It’s rewriting what’s possible.

• We went from basic FAQ bots (remember those?) to AI copilots that help your human agents respond smarter and faster, to fully autonomous AI agents that can resolve complex queries end-to-end. That’s not a small step. That’s a leap.

• Customers today interact with brands on their phones, on mobiles, online, via social media, live chat, in store, sometimes all in the same week. Excellent customer experience now means a consistent, intelligent persona across every single one of those touchpoints. AI is what makes that coherence scalable.

• Before AI, businesses collected data. Now, AI and machine learning evaluate real-time customer data, predict behaviour, and enable truly personalised approaches at a scale no human team could manage. The recommendation engine that knows you want oat milk before you’ve said a word? That’s CX intelligence at work.

• The emotional connection layer of CX is something that doesn’t get talked about enough: the best CX isn’t just efficient, it’s empathetic. Businesses that create emotionally connected experiences see customers who feel genuinely understood and valued, and that drives loyalty in ways that no loyalty points programme ever could.

AI has moved fast. Faster than most CX strategies have kept up with. And that gap? That’s where things get messy.

Now the Fun Part: 🔮 What Does the Future Actually Look Like?

A few things on the horizon that genuinely excite me:

Journey Mapping 2.0: Gone are the static journey maps we used to build in workshops. AI now enables dynamic, personalised, predictive journey maps that update in real time based on how customers actually behave. That’s a completely different tool, one that requires a completely new way of tracking and evaluating customer behaviour, and honestly, a whole article on its own. Coming soon 😉.

Hyper-personalisation at scale: Not just “Hi [First Name]” in an email. We’re talking about omnichannel experiences built from unified customer profiles, where every interaction adapts to who you are and what you need right now. CDPs and modern data architecture platforms have never been more critical; the infrastructure finally matches the ambition.

The Metaverse layer (yes, really): Before you roll your eyes, think about the show Upload (premiered May 2020), where CX is used as a bridge between dimensions and each client has an “angel”, essentially a dedicated agent assigned to meet their every need. Sound familiar? This shows what agent-customer interaction could look like in a three-dimensional, immersive environment, which is genuinely fascinating. Personalised virtual spaces. Gamification. It’s not mainstream yet, but it’s coming, and CX will be there when it does.

💭 My Take

We are living in the most exciting era of customer experience has ever seen. The platforms that were once simple communication tools are becoming intelligent decision systems capable of anticipating needs, personalising at scale, and creating emotional connections that actually last.

But the organisations that will truly win aren’t the ones who bolt AI onto a broken strategy. They’re the ones who build the right foundation, embed customer-centricity into how they’re governed and how they invest, and then let AI do what it does best.

The question isn’t whether to use AI in your CX stack. That ship has sailed.

The question is: what are you building it on top of?

See you next time 😉.