April 2026
Designing AI products that earn customer trust
In my previous article I explored how trust, perceived value and effort are the core behavioral drivers behind whether people actually adopt AI products. Not whether people poke around — but whether they truly adopt AI tooling into their workflow (and pay for it).
A lot of products assume the path to AI adoption is obvious: slap on a copilot, give users a glowing purple prompt box and let the magic happen. ✨
But research — and likely your own experience — suggests that’s a risky strategy. Not least because traditional copilot experiences shift who does the thinking: the human or the machine. And that shift has a knock-on effect on how expensive these experiences feel when they fail.
“When AI absorbs cognition, it also inherits responsibility. And when it inherits responsibility, the cost of its mistakes rises.”
That’s the uncomfortable reality behind many copilot-style experiences. These systems promise extraordinary value — but in return they ask users to invest time, effort and trust. The more investment a system asks from users, the more value they expect in return.
That might be acceptable for free tools — nobody is surprised when a free website generator busts out a generic landing page that doesn’t make much sense. But in paid products, where the cost of failure is higher, the moment a copilot produces something confidently wrong, trust evaporates. In some cases this can even trigger a kind of Broken Window Effect — where one poor outcome leads users to question the reliability of the entire system.
The good news is that copilots are only one extreme end of a much larger design spectrum. If you step back and look at AI features through the lens of visibility — which is how obviously they present themselves within a product — things become much more interesting to design around.
Spectrum of visibilityDirect link
Some AI experiences present themselves directly to the user as collaborators — copilots, assistants or agents. These are highly visible systems — what we might consider foreground visibility. This is the class of AI products that do things on behalf of people, and in some cases they are the product — think something like ChatGPT, or Figma Make.
In the middle of the spectrum, AI is visible but contained. Users know they’re interacting with an AI-powered feature — a summary, a suggestion, an insight — but cognition still largely sits with the human. Success does not depend on the user donating autonomy — in fact if you removed the AI entirely, the product would still function perfectly well.
At the other end, AI operates invisibly — usually at the system or infrastructure layer. These types of features enrich, classify, predict and score without announcing themselves. Users benefit from them without consciously interacting with AI at all — they ask very little. This might look like auto-suggested tags, the right reply surfacing at the right moment, automatic prioritization of a conversation or something as mundane as intelligently removing a long email signature in an incoming message.
The visibility of your AI — whether it feels like a collaborator, a background assistant or simply a more intelligent system — determines how much 💪 effort users invest, how much 💰 value they expect in return and how quickly 😡 trust collapses when things go wrong.
Visibility changes the cost of failureDirect link
The more investment you ask from a user, the less tolerant they become of mistakes. You already know this — it’s well established UX. Research from Nielsen shows that as user commitment increases — through time, effort or trust — tolerance for failure drops dramatically.
Historically we’ve measured this commitment through time or friction. But in an AI era this increasingly shows up as attention, cognitive effort and trust. These are the main human levers we’re pulling.
Consider invisible AI in the form of a background spam filter — it can make the occasional mistake and most people won’t notice. Even if they do, they won’t care too much. The investment is low, so the consequence of failure is also low.
But if a copilot (foreground visibility) sends an email reply to a customer containing incorrect or misleading information, the consequence is very different. Now the user’s investment was high (this is a real customer!), so the cost of failure is high too. Users start checking facts, they challenge every output, they begin to question whether the system can be trusted at all. A broken window.
Same technology… very different tolerance for error! This is where the visibility spectrum starts to matter.
UX complexity correlates to visibility
Trust isn’t the only reason foreground visibility is risky. They’re also incredibly complex to design — and this complexity seems to correlate neatly with where AI sits on the spectrum.
When you start digging into foreground AI experiences you realize they are extremely UX-heavy systems disguised by a deceptively simple glowing purple text box. As usual, the most simple design tends to be the most complex.
Speaking from experience, while a copilot AI looks simple on the surface, behind the scenes they require product teams to solve entirely new interaction problems with no playbook or tried-and-tested patterns. For example: How do users manage context? How does the system explain uncertainty or explain reliability? How do users review and edit AI-generated output? How do we move new users past the empty page problem?
These aren’t easy UX/UI challenges.
And all of this sits on top of constantly evolving user expectations about what these tools should be capable of and, how they work. In fact I’ve designed AI experiences that have aged dramatically within just a few years — that’s entirely new territory!
So why do most teams start here?
The economics of positioningDirect link
Perception, of course. On the other side of this complexity trap is enormous potential upside.
Like most big projects, foreground AI features offer high-risk, high-reward. When they work well, they can define an entire product and outperform competitors. And when they appear so deceptively simple, it’s hard for any leadership team, client or investor to pass them by.
Plus there’s no denying that foreground AI features are hot right now. They’re easy to market, easy to demo and they match what consumers expect AI to look like. A shiny visible copilot promises power and productivity. That’s compelling!
Build it and they will come… right?
Of course, the cost of that promise is significant. High perceived value requires high cognitive effort of users. It requires high trust, high UX investment and a huge cost of failure when things go wrong.
Invisible intelligence works differently. Expectations are lower, failures less dramatic, and the behavioral cost is smaller because users still feel in control. But invisible systems come with their own challenges — they’re harder to sell, harder to screenshot and much harder to package up as impressive “innovation” in a feature release.
Which means positioning on this spectrum is not just a design decision. It’s a strategic one. The question isn’t just how ambitious your AI feature should be… it’s how much visible failure cost your product — and your customers — are willing to tolerate.
The ‘Trust-Acquisition Loop’Direct link
But there might actually be a way to have it all.
A product can support powerful foreground AI that asks a high investment from users — without carrying such high risk. Just not immediately.
Instead of jumping straight to highly visible (and costly) AI experiences, teams can earn their way there by accumulating user trust through increasingly visible surfaces. These are slightly more passive, but might look like an insightful summary, flawless categorization or surprisingly accurate insight that spots something a human missed.
Individually they feel small. But together they prove the system keeps making good decisions. Users may not consciously attribute these improvements to AI — but they begin to develop confidence in the system.
And over time, that confidence turns into trust.
Trust is the north star.
Through the consistent delivery of these small but meaningful “aha!” moments, the relationship between the user and the system begins to evolve. Eventually the product earns the right to ask for more — without such a high consequence for failure. This creates what I’ve been calling a Trust-Acquisition Loop:
👋 Introduce intelligence invisibly — where it’s cheap and easy to implement.
✅ Deliver reliable outcomes that help customers succeed on their own terms.
👍 Build trust through repeated success and repeated “aha!” moments.
♻️ Gradually surface more visible capabilities.
Each cycle allows the system to move slightly further along the visibility spectrum. Adoption may grow slower at first, but it becomes far more stable and tolerant to risk. Instead of asking users to trust the AI immediately, the system proves its intelligence through small, consistent wins. Those wins compound.
Trust can be accelerated in other ways too. Observability and transparency features help humans build confidence in machine decisions — tools for testing, rating and improving outputs all reinforce the sense that the system is behaving predictably.
Equally, establishing clear AI Design Principles can help teams build consistent systems that reflect their values. When products behave in ways that feel coherent, explainable and intentional, users are far more willing to trust them.
Ultimately the goal isn’t just to ship impressive AI features. It’s to build systems that people believe in. Tools that are useful, and convert.
Over time, the most successful AI products won’t necessarily be the ones that launch the flashiest copilots. They’ll be the ones that built real trust first — and only surfaced their intelligence once users were ready for it.
Because great AI products don’t begin with autonomy.
They begin with trust.