Executive summary
Customers do not need AI experiences that simply feel faster or smoother; they need experiences that help them make better decisions with confidence. When customers are evaluating an AI recommendation, the product should slow them down just enough to support reflection through clear explanations, confidence signals, counterarguments, or prompts that prevent blind acceptance. Once they have made a decision, the experience should become effortless, making it easy to act, correct, override, or provide feedback. Customer-centred AI design means putting friction where it protects customer judgment and removing it where it blocks customer control, while measuring success not only by adoption or acceptance rates, but by whether customers genuinely engage with, challenge, and improve the AI’s output.
When Seamless AI Stops Serving The Customer
For two decades, one philosophy has dominated digital product design: seamlessness. The goal of every interface was to remove every obstacle, eliminate every moment of friction, and make the experience so smooth that the user never had to stop and think. And if you could make it a little addictive along the way, even better.
The results are all around you. Swipe to unlock. Swipe right to like. Infinite scroll that never requires a decision to continue. One-click purchase. Face ID. Auto-play. Autocomplete. Each is a small piece of design craft, a trick that removes a moment of friction and makes the desired behavior easier, faster, and more automatic.
Most of the time, this is good design. For AI, it is exactly the wrong default.
The Two Brains You Are Designing For
To see why, it helps to remember how the brain actually works. The intuition is older than software, and Daniel Kahneman gave it its modern shape in Thinking, Fast and Slow: you have two cognitive systems running at all times, and they are not equal.
The lower system, sometimes called the reptilian brain or System 1, is the older and faster one. It pattern-matches, it associates, it predicts, it acts. It runs almost entirely below conscious awareness. It is what you use when you recognize a friend's face in a crowd, pull your hand back from a hot stove before pain even registers, or sense the mood in a room the moment you walk into it. It’s fast, effortless, and often right.
In an AI interface, the reptilian brain is what scrolls past the confidence score, clicks through the warning, accepts the first suggestion that appears, and approves the recommendation because it looks roughly correct.
The higher system, the mammalian brain or System 2, is newer and slower. It reasons. It weighs evidence. It catches the lower brain when its shortcuts go wrong. It is what you use when you work through a logic problem, read a contract carefully before signing, compare two job offers side by side, or change your mind because new evidence has contradicted your first impression.
In an AI interface, the mammalian brain is what reads the explanation behind a recommendation, compares the AI's reasoning with its own, notices that the output does not quite match the situation, and modifies the draft rather than accepting it as is.
Both brains are always available. They are not, however, equally easy to engage. The reptilian brain is faster, cheaper, and almost always wins by default. The mammalian brain takes energy to activate, tires quickly, and gives up easily. Most human cognition, in the wild and at the keyboard, runs on autopilot. Deliberation steps in only when something forces it to.
That asymmetry is the central design question for AI. Where do you want the reptilian brain to lead, and where do you want to wake the mammalian one?
The frictionless paradigm was Built for One Brain
Frictionless design is, at its core, a conspiracy with the reptilian brain. That is the whole craft of it. It is the craft of making actions feel so effortless that deliberation never gets a chance to intervene. For most of the products that defined the past two decades, this was exactly right. You do not want anyone deliberating every time they unlock their phone.
But somewhere along the way, the philosophy escaped its appropriate context and became a universal principle. Designers stopped asking what kind of cognition the moment actually called for. Seamlessness became the goal in itself.
And then came AI.
An AI interface is not a phone lock screen. It is a system that produces outputs, often confident-looking ones, that the user is meant to evaluate, accept, modify, or reject. Apply the frictionless philosophy uniformly to that, and you get something that looks like adoption and feels like trust, but is actually neither. You get users on autopilot, accepting whatever the model produced because the path of least resistance is to accept it and move on.
This is automation bias, and human factors researchers have been documenting it for at least four decades. The pattern was studied in aviation autopilots and clinical decision support long before it had a foothold in generative AI. What is new is the scale: a frictionless AI interface puts automation bias in the hands of every knowledge worker on the planet, every day.
The dark pattern industry has known this for years. The unsubscribe button is buried five screens deep. The pre-ticked opt-in. The cancellation flow is so annoying that most people give up. These work because they exploit the same mechanism great design uses: removing friction from the path you want the user to take.
Friction is not bad. Friction is a tool. The question is never whether to use it. The question is where, how much, and in whose interest.
When Should Frictionless Design Add Friction, And When Should It Remove It?
The human-AI interaction is not a uniform experience. At times, the user is asked to do mammalian work: to understand what the AI has produced and to apply independent judgment to it. At other moments, the user is being asked to do reptilian work: to act on a judgment already made and to receive the feedback that signals their action mattered.
These are two very different cognitive tasks, and they call for two very different design philosophies.
Friction For The Thinking Moment
When the user has to think. This is the territory of meaning and decision. The user needs to understand what drove the AI's output and how much confidence to place in it. The user needs to evaluate the recommendation critically rather than accept it automatically.
The reptilian brain will fail at both. It will pattern-match the output to what feels familiar, find no immediate reason to disagree with the AI's recommendation, and adopt that recommendation as if it were the user's own conclusion. Genuine understanding and genuine deliberation require waking the mammalian brain.
The standard answer to this problem is explainable AI (XAI): surface the model's reasoning, show the features that drove the prediction, render the confidence interval, and trust the user to engage with what you have shown them. It helps. It does not go far enough. An explanation panel nobody opens is still a frictionless interface. A confidence score the user can scroll past is still a frictionless interface. Making information available is not the same as making the user engage with it.
This is where strategic friction earns its place. Researchers call these patterns cognitive forcing functions. A forced pause before a high-stakes confirmation. A prompt that asks the user to articulate their reasoning before approving. A design that surfaces the counter-argument before the conclusion. A confidence score that the user has to actually read before they can proceed. These are not obstacles. They are cognitive forcing functions. The speed bump outside a school is not a failure of road design.
This is also where trust calibration lives. A confidence score is only useful if the user is required to engage with it. A counterfactual ("the score would have been 78% if +1 year experience") is only useful if it lands before the conclusion does, not after.
Seamlessness For The Acting Moment
When the user has to act. This is the territory of control and feedback. The user has already deliberated. They have perhaps decided that the AI is wrong and want to override it. What they need now is not more deliberation. They need effortless execution. An override button that requires three confirmation screens and a written justification is not a trustworthy design. It is friction in the wrong place, at the wrong moment, serving no one.
The same logic applies to the feedback the system gives back. If a user provides input, flags an error, overrides a recommendation, and the interface gives no visible signal that anything changed, that user will quietly disengage. The reptilian brain is wired to repeat behaviors that produce visible reward. Honor that wiring and engagement compounds. Frustrate it, and the user goes silent.
Two cognitive zones. Two opposite design philosophies. One discipline.
How Do You Tell Adoption From Overreliance?
Whether you design these systems, deploy them across an organization, evaluate their outputs day-to-day, or simply live with the decisions they shape, the question to carry away is the same. It is not whether the interface in front of you is friction-free. It is whether the friction it does or does not contain has been matched to the cognitive work the moment requires.
The standard adoption metric, active users, sessions, and completed workflows, tells you how often a system is being used. It does not tell you whether the user is engaging or rubber-stamping. A user who readily accepts every recommendation looks, in the dashboard, identical to a user who is critically evaluating each one. They are not the same user. The first is on autopilot. The second is doing the work that the interface was supposed to support. When something goes wrong, only the second one will catch it.
If your adoption is dominated by speed and acceptance at every step, you do not have engagement. You have overreliance. The interface has done what frictionless design always does. It has put the reptilian brain in charge of work that requires the mammalian one. That is not adoption. That is autopilot wearing the costume of adoption.
The fix is not to make AI harder to use. The fix is to design each moment for the brain it actually needs. Frictionless design where it honors execution, deliberate friction where it wakes judgment.
Friction where it wakes deliberation. Seamlessness, where it honors execution.
That is the discipline trustworthy AI demands. Every interface, every dashboard, every workflow, and every metric should be judged against it. The systems that get it right will produce people who genuinely understand what the AI is telling them, genuinely engage their own judgment in evaluating it, and genuinely feel in control of what happens next. The systems that get it wrong will produce the appearance of all three, and the substance of none.
What To Do On Monday?
At ML6, we build agentic AI systems for clients who care about whether their users are actually thinking. The discipline below is what we use as a first pass when we audit an AI product for the frictionless-by-default failure mode. None of it takes longer than a week to start.
- Inventory the cognitive moments. Walk through every user flow in your AI product and label each step as a thinking moment (the user is evaluating an AI output) or an acting moment (the user has already decided and is executing). Most teams discover that the split is roughly 30/70 and that every single thinking moment has been designed as if it were an acting moment.
- Add one cognitive forcing function per thinking moment. A confidence score the user must actually open. A counterfactual surfaced before the recommendation. A prompt that asks the user to record their own assessment before the AI's score is revealed. Pick one. Ship it. Measure.
- Remove all friction from the acting moments. The override is one click. The correction is one field. The feedback is visible within the same screen. If your users have to leave the flow to disagree with the AI, your design is telling them not to bother.
- Replace the adoption metric. Stop treating sessions and acceptance rates as signs of trust. Start measuring edit rate on AI drafts, override frequency, and time spent on confidence indicators. Those are the metrics that distinguish engagement from rubber-stamping.
- Audit against the EU AI Act. For high-risk AI systems, "effective human oversight" is not a design preference; it is a compliance requirement. Most of the patterns above are also the patterns regulators are about to ask you to demonstrate.
This is the operating manual we apply across our Responsible AI and Agentic AI engagements. If you want to talk through what it looks like in your product, that is where to start.
