Designing AI systems people can trust

AI in product design is often presented as a feature: a chatbot, a smart search bar, a generated summary, or a recommendation card.

But in complex products, AI cannot be treated as something sprinkled on top of the interface. It has to be designed as a system.

A trustworthy AI system needs to understand context. It needs to know when to speak and when to stay quiet. It needs to explain why something is being shown. And most importantly, it needs to help users move forward without overwhelming them.

Recently, I’ve been thinking a lot about how to design AI that supports recommendations, next steps, and contextual guidance. The challenge is not only making the AI “smart.” The challenge is making the AI feel understandable, predictable, and useful.

To make this more concrete, I’ll use a property app as an example: something where users might search for homes, compare listings, submit rental applications, review documents, schedule viewings, pay fees, or decide what to do next.

It is familiar enough that most people understand it, but complex enough to show why AI design needs structure.

Trust starts with grounding AI in real objects

One of the most important design decisions is making sure AI recommendations are connected to real product objects.

When AI creates a recommendation, a task, an update, or a summary, the user should be able to understand what it is about.

A recommendation should not feel like it came out of nowhere. It should be grounded in something recognizable:

• a property
• a listing
• a viewing
• an application
• a lease
• a document
• a fee
• a neighborhood
• a saved search
• a mortgage estimate
• an agent conversation

The user should be able to answer:

• What is this about?
• Why am I seeing this?
• What can I do next?

Without that grounding, AI feels vague. It may sound helpful, but it does not build trust.

For example, this recommendation is weak:

You may want to review your housing costs.

A stronger version is:

This apartment may need review because the monthly cost is higher than your saved budget once fees are included.

The second version works better because it connects the AI message to real objects: the apartment, the user’s saved budget, and the additional fees.

That object grounding matters. It helps users understand that the AI is not just producing a generic suggestion. It is responding to something specific in their experience.

AI content needs clear jobs

A lot of AI experiences become noisy because every message feels like a recommendation. But not every AI-generated item has the same purpose.

I like to separate AI-supported content into simple jobs:

Task – something the user needs to do.

Status – something that is true right now.

Recommendation – something the AI suggests the user consider.

Update – something that recently changed.

This creates a clear system. For example, imagine a rental application is missing proof of income. That same source object could appear in different places, but each appearance needs a different job.

As a Task: Upload proof of income. Your application cannot be submitted until this document is added.

As a Status: Application incomplete. 1 required document is missing.

As a Recommendation: Prepare your references. This landlord usually asks for references before approval.

As an Update: Application saved. Your draft application was updated yesterday.

The same object can appear in more than one place, but the same message should not repeat. That distinction makes the product feel more intelligent and less noisy.

AI should get more specific as the user moves deeper

Another important trust pattern is progressive specificity.

At a high level, AI should summarize. At a list level, AI should help users choose. At a detail level, AI should explain. At the exact object causing the issue, AI should help resolve.

For example, imagine an issue starts inside a rental application because one required document is missing.

At the home page level: 1 application needs attention.

On the applications list: Action required.

On the application detail page: Proof of income is missing.

Inside the document section: Upload proof of income.

Each level has a different job. The home page creates awareness. The list page helps the user choose the right object. The detail page explains what is happening. The deepest actionable object owns the resolution.

This prevents the same AI prompt from appearing everywhere. It also helps the user build trust because the system gives the right amount of detail for the context.

AI prompts should not repeat everywhere

One pattern I try to avoid is putting the same AI prompt across every level of the product. Imagine the same prompt on every screen:

Ask AI: Why is my application incomplete?

On the home page. On the applications list. On the application detail. In the document section. The exact same prompt, everywhere.

This feels repetitive and lazy. It makes the product feel like AI was pasted onto screens instead of designed into the system.

A better version would be:

Home page – 1 application needs attention.

Applications list – Action required.

Application detail – Proof of income is missing.

Document section – Upload proof of income. Ask AI: What document should I upload?

Same issue. Different level. Different job. The deeper the user goes, the more specific the AI can become.

Standard search and AI help should not do the same job

Search is another place where trust can break quickly.

If a user searches for “apartments,” they may simply want to find listings. That is standard search.

If a user asks, “Can I afford this apartment?”, they are asking for help understanding a decision. That is AI guidance.

I like to separate these modes clearly:

Standard Search – find this.

AI Guidance – help me understand this.

Standard search should retrieve known objects:

• listings
• saved searches
• neighborhoods
• viewings
• applications
• documents
• agents
• leases
• fees

AI guidance should help interpret, summarize, or recommend next steps.

For example, standard search can return:

2-bedroom apartment in Brooklyn – $3,200/month · Available July 1

AI guidance can say:

This apartment is within your monthly rent range, but the upfront cost may be high because it requires first month, security deposit, and a broker fee. You may need around $9,600 before move-in.

Those are different levels of system behavior, and the interface should make that difference visible.

AI recommendations need rules, not just prompts

A trustworthy AI product is not just a good prompt. It needs rules. Some of the rules I define early are:

• When should AI show a recommendation?
• When should something become a task instead?
• When should AI stay quiet?
• When should it show evidence?
• When should it route to a human?
• When should it avoid showing something because confidence is too low?

For example:

• If the user must do something, it should become a task.
• If AI is suggesting something optional, it can become a recommendation.
• If something simply changed, it is an update.
• If something is ongoing, it belongs in status.

These rules help designers, product managers, and engineers make consistent decisions. They also make AI feel less random to users.

Trust also comes from restraint

A common mistake is making AI visible everywhere. Every card gets a prompt. Every page gets a suggestion. Every screen gets a generated summary.

But more AI does not automatically create more trust. Sometimes it creates noise.

AI should appear when it adds value:

• when something is unclear
• when something needs action
• when the user may not know what to do next
• when multiple objects need to be connected
• when a recommendation is genuinely useful
• when human support may be needed

AI should not appear just because the system can generate something. Restraint is part of trust.

Designers need to define the system, not just the screen

Designing AI systems is not only about interface design. It is also about defining the relationships between:

• objects
• user goals
• system signals
• recommendations
• evidence
• prompts
• actions
• handoffs
• trust moments

The interface is only the visible part. Behind it, designers need to define the logic of when AI appears, what it knows, how it explains itself, and how it helps the user move forward.

For me, this is where AI product design becomes most interesting. Not:

Where can we add AI?

But:

What system would make AI useful, understandable, and trustworthy here?

Closing

The future of AI in product design is not just smarter answers. It is better systems.

Systems that are grounded in real objects. Systems that separate tasks from updates and recommendations. Systems that become more specific as users move closer to the source of a problem. Systems that explain themselves. Systems that know when not to speak.

That is how AI starts to earn trust: not by sounding intelligent, but by helping people understand what is happening and what to do next.