Why Good AI Sometimes Needs to Get in the Way

For years, technology has been designed to remove friction. Fewer steps, faster answers and less human intervention have generally been seen as signs of progress. 

AI has accelerated that thinking. But as businesses race to make AI faster and more autonomous, there’s a danger that we start treating every pause, check or human intervention as a failure. 

Sometimes the opposite is true. Good AI isn’t designed to do everything. It’s designed to understand the boundaries of what it should do. 

The Best AI Knows When to Stop 

Imagine someone calls a hotel and asks for a table for two on Friday at 7:30pm. 

That’s exactly the kind of structured, transactional conversation AI can handle well. It can check availability, make the reservation, confirm the details and let the customer get on with their day. 

Now imagine the next caller says they’re planning their wedding reception at the same hotel. They want to talk about possible dates, guest numbers, menus and what the venue could offer. 

Technically, an AI could continue the conversation. It could keep gathering information and asking questions. 

But should it? 

At some point, that conversation stops being a straightforward transaction and becomes something that requires judgement, expertise and relationship-building. The most useful thing the AI can do at that moment is recognise the change and bring in a person. 

That’s not AI failing. It’s AI understanding its role. 

Some Friction Is There for a Reason 

We tend to think about friction as something customers dislike, and much of it is. 

Nobody wants to repeat information they’ve already provided, navigate unnecessary steps or wait for approval when an AI is perfectly capable of completing a routine task. 

But there is another kind of friction that serves a purpose. 

It’s asking someone to confirm before an irreversible action is taken. It’s verifying information when there is uncertainty. And it’s pausing an automated journey when the consequences of getting something wrong become more significant. 

The key question shouldn’t be, “Can we remove this step?” It should be, “What is this step protecting?” 

If the answer is nothing, remove it. If it’s protecting accuracy, security, trust or the customer’s intended outcome, that friction may be valuable. 

Knowing When to Quit Is An Engineering Problem 

Getting AI to stop at the right moment sounds simple. In practice, it’s one of the harder parts of designing a good AI experience. 

First, you need to define the boundaries. 

What can the AI handle confidently? What should trigger an additional check? What types of conversation, decision or request should result in escalation? 

Those boundaries shouldn’t be left entirely to the model to improvise. The important business rules around permissions, routing and escalation need to be deliberately designed and governed. 

Then there is confidence. 

An AI system needs to recognise when it has enough information to continue and when uncertainty has become too high. Sometimes the right response isn’t another generated answer; it’s acknowledging the limitation and changing course. 

And finally, there is the handover itself. 

A Customer Shouldn’t Have to Start Again 

This is where businesses can inadvertently replace one form of friction with another. 

If the AI spends several minutes gathering someone’s name, requirements, dates and preferences and then transfers the conversation to a person who knows none of it, that’s not a seamless escalation. 

It’s asking the customer to do the work twice. 

A good handover should carry the context with it. The person taking over should understand who the customer is, what’s already been discussed and what they’re trying to achieve. 

That principle applies beyond customer conversations too. If an AI system escalates an internal decision for human review, the person needs enough context to understand why it was flagged and what they need to decide. 

Human intervention shouldn’t mean going back to the beginning. 

The best experiences are the ones where AI and people each do what they’re best at and the transition between them is almost invisible. 

Stop Measuring Success by How Rarely Humans Get Involved 

There’s a wider measurement problem here too. 

One common way of assessing conversational AI is containment: how many interactions can the AI complete without involving a person? 

It’s an understandable metric. But taken too far, it can encourage exactly the wrong behaviour. 

If the objective is simply to maximise containment, every human handover starts to look like failure. That creates an incentive for the AI to keep going when the better customer outcome might be to stop. 

Instead, we should be asking whether the customer achieved what they came to do. 

Was the issue resolved? Was the booking made? Did the customer reach the right person? Did the technology make the journey easier? 

An AI that hands over 20% of conversations at exactly the right moment may create far better outcomes than one that contains 95% but leaves customers frustrated. 

The goal shouldn’t be to keep humans out of the process. It should be to involve them where they add value. 

Design The Boundary, Not Just the Automation 

As AI becomes more capable, the question of what it can do will become less useful. 

Increasingly, it will be able to do a great deal. 

The more important questions will be what we want it to do autonomously, where we want it to verify before acting and where we want a person to take over. 

Those decisions need to be made deliberately, rather than discovered when something goes wrong. 

So before removing another step from an AI journey, ask why that step exists. 

Some friction is pointless and should disappear. Some is there because the technology or process hasn’t been designed well enough. But some friction protects the customer, the business or the outcome. And good AI needs to understand the difference. 

Sometimes it should act. Sometimes it should check. And sometimes the most intelligent thing it can do is recognise that it has reached its limit, stop and hand over to a person. 

That’s not a failure of automation. It’s what good automation looks like.