Gemini Intelligence Works, but Its Audience Is Still Unclear

Google has launched Gemini Intelligence on Samsung’s newest foldables, but one early test exposed a conflict between technical capability and practical value.

The 9to5Google Google test successfully sent Gemini through Expedia to find a flight from London Heathrow to Dublin. Yet the system never asked about preferred times or airlines before starting its work.

That omission matters because Google presents Gemini Intelligence as a way to delegate everyday digital chores. The early experience suggests users must still supervise decisions that make those chores personal, costly, or difficult to reverse.

Samsung’s Galaxy Z Fold8 Ultra, Fold8, and Flip8 are the first phones carrying the new platform. They give Google a high-profile launchpad before Gemini Intelligence expands to Pixel phones and other Android devices.

The technology can navigate supported apps, interpret screen content, and complete several steps after a spoken request. That makes it more capable than a conventional assistant that merely opens an app or displays search results.

However, successful navigation does not establish that people want software choosing among flights, restaurants, products, or rides. The harder challenge is knowing which preferences matter before taking action.

The launch therefore tests more than Google’s ability to automate Android. It tests whether mobile users will delegate ordinary decisions when checking the agent’s work can consume much of the time it promised to save.

The First Gemini Intelligence Test Exposed the Real Gap

Gemini Intelligence completed the assigned workflow, but completion was not the same as understanding the user’s decision.

Samsung unveiled its new foldables at Galaxy Unpacked in London on July 22, 2026. Google used the event to release the first public capabilities associated with Gemini Intelligence.

The initial release expands task automation from a small collection of services to more than 40 popular apps. Google lists shopping, restaurant reservations, travel experiences, and event tickets among the supported activities.

Gemini can work across multiple screens while showing its progress. Users can leave it running in the background, interrupt a step, make an adjustment, and then return control to the agent.

That design represents a meaningful change from earlier smartphone assistants. Instead of translating a request into a shortcut, Gemini attempts to operate the interface and advance the task itself.

9to5Google’s Damien Wilde briefly accessed that automation on a pre-release Samsung device. Some required services were already installed, including Expedia, which allowed him to recreate a travel demonstration.

Wilde asked Gemini to find a flight from Heathrow to Dublin for a date one week later. According to his Gemini Intelligence test, the system opened Expedia and performed the search successfully.

The result confirmed that Gemini could translate a natural-language request into a multistep interaction. It also revealed how much relevant information the original instruction left unstated.

Gemini did not ask Wilde which airline he preferred. It also did not request a departure time, even though either choice could determine whether a proposed flight was useful.

The user could intervene after noticing the omission. Google designed the workflow to support that correction, and Gemini could resume once the user supplied more direction.

Still, intervention changes the value calculation. The person must watch the process, identify a missing constraint, stop the agent, enter the detail, and review the resulting choice.

Sensitive information created another boundary. Wilde had to take control when the booking process reached the passenger-data screen, although behavior might differ when an app already stores account details.

That safeguard limits unwanted disclosure and unintended transactions. It also means the agent does not simply receive a destination and return with a completed booking.

The experience took place on pre-release hardware with incomplete integrations. It cannot establish how every workflow will behave after the full rollout.

However, it identifies the central issue early. A task can be technically successful while leaving the user uncertain whether the result reflects their actual preferences.

That distinction creates the article’s main tension. Google has shown that an Android agent can operate apps, but it has not shown that routine delegation feels easier than direct control.

Google Built an Agent That Still Needs a Manager

The launch pressures Google to prove that supervision costs less attention than completing the task manually.

Google describes Gemini Intelligence as proactive technology that handles tedious work while keeping the user in control. Its examples focus on chores that require several apps, screens, or pieces of context.

One example begins with a grocery list displayed in a notes app. A user can invoke Gemini over that list and ask it to build a delivery cart.

Another starts with a travel brochure. Google says a user can photograph it and ask Gemini to find a similar tour for a group of six.

These examples are stronger than the flight test because the screen or image supplies useful context. The agent receives a concrete list, picture, or source instead of reconstructing preferences from a short command.

Google’s Android launch details also describe background progress notifications and a final confirmation step. The user remains responsible for reviewing the proposed action.

That structure is sensible for purchases and reservations. An unchecked agent could select the wrong date, duplicate an order, expose personal information, or commit money under unsuitable terms.

Yet every protective checkpoint adds friction. If users must provide extensive instructions and inspect each important field, direct app use can remain faster and more predictable.

A flight search makes this problem unusually visible. Departure time, airport, baggage allowance, connections, refund rules, airline, and loyalty membership can all affect the right choice.

The original request contained only two airports and a date. Gemini therefore faced an under-specified decision, not merely an interface-navigation problem.

A human travel agent would normally clarify important preferences before presenting options. A digital agent needs a comparable method for deciding when to ask and when to proceed.

Asking about everything would make the interaction exhausting. Asking about nothing transfers too much judgment to a model that does not necessarily know the user’s priorities.

Personalization could narrow that gap. Google’s broader direction connects Gemini with information from apps and services, potentially giving it access to established routines and preferences.

However, inferred preferences carry their own risk. A past purchase does not always represent a lasting rule, and a familiar airline might not be suitable for a particular trip.

The agent therefore needs more than memory. It needs calibrated uncertainty, which means recognizing when existing context is insufficient for a consequential choice.

This is why the “agent that needs a manager” criticism matters. The problem is not that Gemini failed to press the correct buttons during the demonstration.

The problem is that the user still supplied judgment, monitored execution, handled sensitive data, and checked the result. Those activities contain much of the cognitive work behind the original task.

Google’s immediate pressure is to make that management burden visibly smaller. A polished animation or successful sequence cannot substitute for measurable time and attention savings.

Samsung also has something at stake. Its foldables are the first devices associated with the Gemini Intelligence badge, so buyers will encounter Google’s early product decisions through Samsung hardware.

Samsung says its new Galaxy lineup combines partner AI services with transparency, privacy, and user control. Those principles are necessary, but users will judge the phones through actual reliability.

An automation error can reflect poorly on both companies, even when a third-party app or incomplete user prompt causes it. The partnership distributes the opportunity and the reputational risk.

9to5Google Google Findings Challenge the Convenience Pitch

The central reversal is simple: the agent works, yet its successful performance can still feel unnecessary.

Google is not presenting Gemini Intelligence as a laboratory project. It is positioning the platform as a practical answer to routine “life admin” across Android.

The 9to5Google Google findings challenge that pitch because the demonstrated task already has a familiar, reasonably efficient interface. Users know how to enter airports, dates, and filters in a travel app.

Automation becomes valuable when it removes coordination that people dislike. Copying a long shopping list across apps is a clearer example because the agent can eliminate repetitive entry.

A complex research task offers another promising case. The user might need to extract items from an email, locate them in another service, compare availability, and prepare a reviewable result.

Booking one straightforward flight is different. The user often wants to compare options directly, and the comparison can be more important than filling the search form.

In that situation, the interface is not merely friction. It is where the person discovers tradeoffs, changes priorities, and develops confidence in the final choice.

The same issue appears in food ordering. Repeating a known order from a trusted restaurant is suitable for delegation because the desired outcome is already defined.

Choosing dinner for several people involves dietary restrictions, delivery timing, substitutions, and changing preferences. Automation can move through screens without resolving those human constraints.

Google’s agentic AI model, meaning software that plans and performs several actions toward a goal, therefore has an uneven field of useful tasks.

High repetition and low ambiguity favor automation. High consequence and subjective choice favor direct user involvement, even when the technical workflow is easy to automate.

This division explains why Wilde described the technology as closer to a compelling demonstration than an essential service. The agent’s visible activity can look impressive without changing the user’s outcome.

The launch also highlights a difficult product-design question. Google must decide whether Gemini Intelligence should replace app interactions or prepare them for faster human review.

The second model appears more credible today. Gemini can gather information, populate a cart, or narrow options while leaving the final judgment with the user.

That approach provides assistance without pretending the agent possesses every preference. It also aligns with Google’s final-confirmation requirement.

However, preparation must still save enough effort. A result that demands line-by-line verification can become another inbox that the user has to process.

This is where reliable personal context becomes important. People already struggle to retrieve project details, travel requirements, and decisions scattered across different services.

A searchable personal knowledge base can reduce that retrieval problem. It does not eliminate the need to verify an external agent’s proposed action.

Google’s challenge is broader because Gemini must act across systems it does not fully control. Every supported app has distinct layouts, account states, confirmation screens, and error conditions.

More than 40 integrations make the launch sound substantial. The useful measure will be how many workflows finish correctly under ordinary, messy conditions.

A restaurant might have no availability. A product might be out of stock. A ticketing service could introduce a queue, authentication request, or changing seat map.

These exceptions separate dependable agents from scripted demonstrations. They also determine whether users remain comfortable letting the software continue in the background.

The early flight test does not prove Gemini will fail those situations. It shows that successful execution alone cannot answer whether the product deserves routine trust.

More Than 40 Apps Do Not Guarantee Everyday Trust

Coverage measures where Gemini can act, while trust depends on how it behaves when instructions, interfaces, and consequences become uncertain.

Google’s expanded app support is the clearest numerical improvement in the first Gemini Intelligence release. The previous task-automation beta supported only a handful of food and rideshare services.

The new version reaches more than 40 popular apps, according to Google’s Unpacked announcement. It also combines app control with screen and image understanding.

Those capabilities widen the range of possible requests. They do not ensure that every app supports every workflow, or that each flow handles interruptions consistently.

Google’s support documentation says the feature is available through the Gemini mobile app on Android. It can also run when users activate Gemini by touch or voice.

On Samsung’s new Fold8 and Flip8 devices, Gemini can request that the user take over a task. If the user does not respond within ten minutes, the automation stops.

That timeout illustrates the hybrid nature of the product. Gemini acts independently for part of the process, but some states still require immediate human judgment or authentication.

This handoff can protect the user. It can also interrupt the exact moment when background automation was supposed to free their attention.

Trust will depend on whether these interruptions occur at understandable boundaries. Payment confirmation and passenger details are reasonable places to require control.

Unexpected handoffs during simple navigation would feel different. Repeated interruptions could teach users that delegating a task creates more uncertainty than doing it themselves.

Google also needs consistent explanations. Users should know what the agent selected, which information shaped that decision, and which alternatives it discarded.

A final screen is not always enough. If Gemini presents one flight without explaining the search parameters, the user must repeat the comparison to evaluate its recommendation.

There is also a meaningful difference between reversible and irreversible actions. Adding groceries to a cart is easy to undo, while submitting a booking can create fees or limited recovery options.

A mature system should adapt its autonomy to that difference. It can move quickly through low-risk steps and slow down when consequences become harder to reverse.

Samsung emphasizes control and privacy in its Fold8 announcement. Google likewise says Gemini acts after a user command and stops when the task is complete.

Those statements describe intended safeguards. Independent testing must establish how predictably they work across supported services and unusual app states.

The hardware requirements also restrict the initial audience. Google associates Gemini Intelligence with advanced devices that have on-device AI models, recent media capabilities, and at least 12GB of RAM.

Samsung’s new foldables meet those requirements and introduce Gemini Nano 4. This on-device model supports local AI functions that can complement cloud-based Gemini services.

According to device requirement details, Google is also evaluating launch quality and field performance standards.

A limited hardware rollout gives Google a controlled starting point. Users buying new premium foldables, however, are not necessarily the people who need app automation most.

They might tolerate experimentation because they actively seek new technology. That can produce enthusiastic demonstrations without proving durable adoption among mainstream Android users.

The original review acknowledges this limitation. Wilde had incomplete integrations and only brief access on pre-release hardware, so his experience was not the finished product.

That caveat should temper conclusions about reliability. It does not erase the question of audience because the tested workflow was one Google and Samsung chose to highlight.

Marketing examples reveal the job a company believes its product can perform. If the example feels slower or less trustworthy than direct app use, the value proposition needs refinement.

The Best Use Cases Are Narrower Than Google’s Vision

Gemini Intelligence looks strongest when the desired outcome is explicit, repetitive, reversible, and expensive to enter manually.

The grocery-list example fits those conditions better than travel booking. A written list defines the items, while the cart remains reviewable before purchase.

Reordering familiar food is another plausible case. The agent can reuse a known selection and ask only about changes, reducing both ambiguity and manual input.

Ride booking also has potential when the destination is already clear. The user still needs to confirm the pickup point, timing, vehicle type, and estimated arrival.

Accessibility provides a more consequential opportunity. Voice-driven automation can reduce the effort required from people who find complex touch interfaces difficult to navigate.

That benefit should not be treated as a niche afterthought. An agent that operates cluttered app interfaces could remove real barriers, even when it saves only a small amount of time.

The product could also help users manage several sources at once. Google’s example of finding a syllabus in Gmail and adding required books to a cart involves retrieval and transfer.

This workflow contains mechanical labor that users may gladly delegate. The list still needs review, but the agent can reduce searching and repeated data entry.

By contrast, open-ended shopping invites subjective decisions. “Find a good laptop” requires a budget, workload, portability preference, operating system, and tolerance for compromises.

The agent can ask those questions, but a long interview reduces convenience. It can infer answers, but silent inference increases the chance of an unsuitable result.

This creates a practical rule for early use. Users should delegate preparation before delegating judgment.

Ask Gemini to assemble options, transfer explicit information, or fill reversible fields. Keep direct control over preferences, sensitive information, and final commitments.

That model is less ambitious than Google’s vision of proactive technology working throughout the day. It may also offer a more credible path toward habitual use.

Trust often grows through small, predictable successes. Users might first accept help with carts and forms before allowing the agent to handle reservations or purchases.

Google’s broader rollout could support that progression. The company says Gemini Intelligence features will arrive in waves on Samsung and Pixel phones before reaching watches, cars, glasses, and laptops.

Cross-device access raises the stakes. A request that begins through smart glasses might be convenient, but reviewing detailed alternatives on that interface could be difficult.

The agent would then need to understand when to act, when to summarize, and when to move the decision to a screen. Interface selection becomes part of intelligent assistance.

That future also increases the importance of personal context. Users will not want to restate the same preferences on a phone, watch, car, and pair of glasses.

Google can reduce repetition by connecting account data and Personal Intelligence features. It must also provide clear controls over which information informs each action.

A useful agent cannot simply know more. It must make the boundary between remembered facts, inferred preferences, and current instructions understandable.

For knowledge workers, this distinction resembles the difference between recall and authorization. Finding an old decision is helpful, but acting on it requires confidence that the decision still applies.

Tools for work memory can help people locate prior context. Gemini Intelligence goes further by trying to convert context into external action.

That additional step creates both its appeal and its risk. The agent saves more effort only when users trust it with more control.

The early evidence supports a narrower conclusion. Gemini Intelligence has crossed the technical threshold for operating several apps, but it has not crossed the social threshold for broad delegation.

Three Signals Will Show Whether Gemini Intelligence Matters

The next test is not another polished demonstration, but whether ordinary users repeatedly delegate tasks without recreating the work through supervision.

The first signal is the quality of the full production rollout on Samsung’s Fold8 family. Reviews should track completion rates, interruptions, corrections, and time saved across multiple supported apps.

One successful flight search proves that the agent can navigate a prepared path. Repeated testing will show whether it handles account prompts, unavailable options, changing layouts, and incomplete instructions.

The most revealing comparisons will measure total user attention. An automation that runs for several minutes can still be useful if it needs only a brief review.

The same workflow loses value when the user watches every screen and repeatedly intervenes. Time alone does not capture that difference, so reviewers should document active supervision.

The second signal is how Google handles preference clarification. Gemini must learn which missing details require a question before it begins an important task.

If it starts asking concise, relevant questions about flights, purchases, and reservations, the early criticism will weaken. That behavior would show improved judgment about uncertainty.

If it continues selecting defaults without explanation, users will keep treating the system as an interface operator. They will hesitate to view it as a dependable personal agent.

Google should also make the final review more informative. A good summary would identify chosen constraints, unresolved assumptions, sensitive steps, and alternatives worth checking.

The third signal is adoption beyond launch-week demonstrations. Google’s Pixel expansion and later cross-device rollout will reveal whether app automation becomes a recurring behavior.

Feature availability is not adoption. Useful evidence would include repeated task delegation, broader third-party support, and workflows that users initiate without promotional prompting.

Pixel users provide an important test because Google controls the hardware, operating system, assistant, and many connected services. That integration should reduce inconsistencies found on third-party devices.

If Gemini Intelligence remains associated mainly with premium foldable demonstrations, its audience will stay narrow. The technology may remain impressive without becoming a standard Android habit.

If users begin delegating repetitive, explicit tasks several times each week, the product will have found its foothold. Google can then expand autonomy from that base.

The 9to5Google Google test does not settle the future of agentic Android. It supplies a useful early warning about confusing capability with value.

Gemini Intelligence can perform a multistep task, display progress, accept corrections, and return control at sensitive stages. Those are meaningful product advances.

Yet people do not delegate merely because software can click through an app. They delegate when the result is predictable, the handoffs are clear, and oversight takes less effort.

For now, cautious users should test the system with reversible tasks whose desired outcomes are already explicit. Lists, familiar orders, and information transfer are better starting points than consequential choices.

Then ask a direct question after each attempt: did Gemini remove work, or did it merely turn that work into instructions and review?

Google’s answer will emerge through product behavior, not launch language. The audience for Gemini Intelligence will become clear only when users stop managing the agent and start trusting the outcome.