Travel has always been a story of how much friction we’re willing to tolerate to get somewhere.
A century ago, a long-distance journey began at a train station thick with smoke and shouting — porters calling out platforms, conductors checking tickets by hand, departure boards updated by men on ladders sliding metal placards into wooden frames. To find your train, you asked. To know if it was on time, you asked. To navigate the station itself, you read paper signs, listened for announcements, and trusted strangers in uniforms. Information lived in people, on walls, and on paper. Getting it requires your active attention at every step.
Flash-forward to today. Most of us have been there: sprinting through an unfamiliar airport on a tight connection — gate changed, wrong terminal, boarding in nine minutes. Scanning digital signage for your flight details, checking your airline app on your phone, asking a stranger for directions you’re not sure you can trust. The infrastructure has become extraordinary in the century since the great train stations. Vast real-time departure boards. Multilingual overhead announcements. Phone apps pushing gate changes before you’ve thought to check. Ride-shares that know you’ve landed. Using face recognition to pass you through security. None of this felt like a revolution as it arrived, but each one is the first glimpse of something bigger: intelligence that reaches you before you reach for it.
Now imagine where this is heading in the future. You walk into a terminal that no longer looks like an airport. The departure boards are gone. So are most of the large signs, the overhead voice paging passengers, the long banks of monitors that have defined air travel for decades. The space feels more like a hotel lobby: more human, more open, more conversational. As you cross the threshold, a heads-up-display in your contact lenses gives you a subtle indication for the direction you need to go, and a voice in your ear says, “Gate B47, ten-minute walk. You have time to take your next call before boarding. Here’s a lounge to your right where you can take it.” Every traveler in the room is moving through their own personalized layer of information, invisible to everyone else, delivered exactly when and where it matters.
Across a century of travel, infrastructure has gotten smarter by getting quieter. Train stations became airports. Departure boards became phone apps. And now, the apps themselves are about to dissolve into the environment.
This isn’t a thought experiment. It’s the direction enterprise AI is already moving. And it raises a question that my team at Salesforce AI Research has been working to answer: what happens when AI stops being a tool you go to, and becomes a presence that’s simply already there?
The Four A’s of Ambient Intelligence
Ambient intelligence is not a single product or a single technology. It’s a design principle—a set of properties that distinguish it from the AI interactions most of us use today. Four key attributes define it, and for illustrative purposes, we’ll use B2B scenarios we come across daily at Salesforce.
1. Always-On.
Traditional AI operates in discrete exchanges: a prompt goes in, a response comes out, the interaction ends. Ambient intelligence operates on a continuous stream — audio from a sales call, visual input from a field technician’s camera, telemetry from thousands of enterprise systems, signal from any source that generates data over time rather than at a single moment. The shift sounds technical, but the experiential difference is significant.
This distinction matters because the most valuable moments in a workflow rarely announce themselves. The sales rep doesn’t stop mid-conversation to file a help ticket. The operations manager doesn’t always know which of ten dashboards to check first. Ambient intelligence has to be watched before it can be useful, which means the underlying system must be able to process streams natively, not just snapshots.
2. Aware
A signal on its own is rarely enough to produce a good decision. What the signal means depends on what else is happening around it: history, prior conversations, external events, background conditions. A spike in server demand means one thing on an ordinary Tuesday and something entirely different during a long holiday weekend or on the day a major news event changes customer behavior. A context-aware system widens the aperture beyond the data point itself to incorporate everything that might explain it.
This is where ambient intelligence separates from traditional analytics. A dashboard shows you the signal. An ambient system is aware of the signal in context — and can therefore tell you not just what happened but why, with enough confidence to know whether it matters.
3. Adaptive
Because ambient systems remember, they can tailor and adapt, creating a highly personalized experience. The alert one employee receives is not the same alert another employee receives, even from the same underlying data, because the system has learned what matters to each of them — which regions they cover, which customers they’re focused on, which patterns they’ve already reviewed. Broadcast notifications send the same message to everyone; personalized ambient intelligence sends the right message to the right person at the right moment.
This is a property of memory vs. interface. A system that cannot remember cannot personalize. Persistence is what turns raw signal into relevant insight.
4. Anticipatory
The fourth attribute is the one that most clearly distinguishes ambient intelligence from the AI most people use today. Ambient systems anticipate rather than react. They surface the insight before the human realizes they need it. They take action within defined parameters without waiting for instruction. The shift is from “ask and receive” to “anticipate and deliver, and that inversion is what makes the paradigm genuinely new.Proactivity is also where the design challenge gets hardest. Acting without being asked is powerful when it’s calibrated correctly. It’s disruptive when it isn’t. Which is why the attributes that enable ambient intelligence must be matched by a discipline that governs when and how it actually speaks.
A system that’s always on but poorly calibrated fails in two directions. In the worst case, it interrupts at exactly the wrong moment, which disrupts focus, breaks a customer interaction, or pulls a decision-maker’s attention away from something consequential. In the everyday case, it simply pesters: a steady drip of low-value nudges that trains the human to ignore the system entirely. Anyone who lives in the modern work environment knows: after enough red badges and pop-ups, you file them under “later” — which really means “never.” Ambient intelligence has to earn every interruption it makes.
We changed the engine, not the factory floor
To understand why this shift to Ambient Intelligence matters (and why it’s happening now) it helps to look at a moment from industrial history that rhymes uncomfortably with where we are today.
When electricity replaced steam power, factory owners installed the new motor where the old engine sat and changed nothing else. For decades, productivity barely moved. Only when engineers asked “how would we design this factory from scratch with electricity?” did everything shift—new layouts, new workflows, new relationships between worker and machine. Productivity skyrocketed. Economists Erik Brynjolfsson and Andrew McAfee documented this pattern in The Second Machine Age, and it maps with uncomfortable precision onto the current state of enterprise AI, well over three years since the release of ChatGPT.
Of course, the models have gotten dramatically better since then. Through research and advancement, intelligence is genuinely more powerful. But the way humans interact with that intelligence has barely changed. You still open a tool. You still type something. You still wait. The leap from one model generation to the next, once felt as a revelation, has become increasingly imperceptible to everyday users — not because progress has stopped, but because we keep pouring better capability through the same old interface. You don’t notice the 10x better model when you’re still typing prompts into the same chat window. We changed the engine, but we didn’t change the factory floor.
Ambient intelligence is the new factory layout. It asks: if we were designing the relationship between human and AI intelligence from scratch, knowing what we now know, would we really build a system that requires humans to stop what they’re doing, open a separate tool, and articulate a precise request? Or would we design intelligence that’s already present in the workflow, already aware of context, already ready to act?
Three places it’s already happening
Ambient intelligence isn’t a single product or a single use case. It’s a design principle that expresses itself differently depending on the environment. Here are three places where we’re actively seeing it take shape.
In sales and service conversations.
Most business happens at the human level. It’s really about conversation between two people. So how can AI interject without disrupting that dynamic? In a traditional sales interaction, there are often too many details for a sales rep to remember, or have immediate access to. This can often lead to delays in the deal cycle, or worse, losing the deal. And in service calls, there is a similar challenge, where the information needed to serve the customer needs to be manually accessed, leading to long wait times, transfer to different departments and low customer satisfaction.
What if ambient intelligence could change that? Our team has been experimenting with an approach we call PISA – Proactive In-Meeting Support Agent. PISA listens to the conversation and scans the screen to understand the context in real time. And because it has access to customer data, meeting notes and other contextual information, it surfaces information just at the right time, and in a format that is quick and easy for the human sales or service agent to digest.
In operational environments.
Enterprise systems generate millions of signals every hour— latency metrics, error rates, log lines, pipeline movement. Most is routine; a fraction matters: service degradation, customer churn, security anomalies. The traditional dashboard requires a human to decide which one to check first. Ambient intelligence watches continuously at machine scale.
Ambient intelligence in operational environments changes the equation. Instead of a human deciding which dashboard to open, the system itself watches — continuously, at machine scale — and surfaces only what genuinely requires a human decision. Our team’s work on what we call Deep Insights, which effectively generates a data analyst agent in the background, that investigates the data and brings up the right insight, to the right leader, at the moment it matters.
In the field.
Most enterprise AI lives behind a screen. But a significant share of the work that keeps modern businesses running happens in physical environments where workers can’t pause to type a question — a technician on top of a wind turbine, a service engineer inside a customer’s data center, a field worker diagnosing a malfunction in a hospital basement. These are the moments when expertise matters most and where access to information is hardest to get.
Our team has been working on what ambient intelligence looks like in this environment: lightweight smart glasses paired with an always-on visual reasoning system. Imagine a technician who has spent two hours on a job and needs to ask about something that happened in the first twenty minutes. The system has to hold that entire session in context, not just the last exchange. Or consider the moment when a technician asks, “Show me how to fix this” — and the system generates guidance using the exact equipment visible in front of them, not a generic diagram from the manual. That level of personalized, contextual, long-horizon memory is what separates ambient intelligence from a search engine with a camera.
The technical challenge is real and quite different from the sales context: visual context is messier than audio, equipment varies, lighting changes, a given component might be partially occluded or installed in a configuration that doesn’t match the manual. Solving it means combining real-time visual understanding with deep knowledge of the equipment, the customer’s service history, and the technician’s own training level, all fast enough that the guidance arrives while the technician’s hand is still on the part.
The result is the same pattern we see in PISA and in operational environments, expressed in a different physical form. The AI is always paying attention. It understands context. It tailors its response to the person and the moment. And it acts before being asked. The technician is the one making the decisions; the system is the one making sure they have what they need to make those decisions well.
Earning the “right to interrupt”
Describing systems like these, there is a temptation to make them sound frictionless. They are not. The design challenge at the center of ambient intelligence is genuinely difficult, and getting it wrong produces something worse than the problem it was meant to solve. A system that surfaces too much becomes noise; the human it was meant to assist learns to ignore it. A system that surfaces too little becomes irrelevant; the moments that mattered slip past. Between those two failure modes is a narrow, context-dependent threshold — the moment at which intelligence becomes genuinely useful rather than intrusive.
Hitting it requires the system to understand not just what is happening, but what kind of moment it is. Context comes in two forms. There’s spatial context: the events happening around a specific signal that explain it. A bandwidth spike on a Wednesday afternoon means one thing; the same spike on the Friday before a long weekend, when traffic has its own seasonal logic, means something else entirely.
There’s also temporal context: the history that gives the present moment its weight. A latency anomaly that follows last week’s failed deployment is a different problem than the same anomaly arriving out of the blue. Or, in a sales context: a customer’s question about pricing means one thing in the first conversation and something quite different in the seventh, when the deal has stalled twice already. Without both kinds of context, even the best ambient system raises false alarms or misses real ones — what researchers call “precision and recall,” and what engineers more bluntly call “crying wolf.”
This is a research problem as much as an engineering one. And it connects directly to the principle that guides all of our work at Salesforce AI Research: humans at the helm. Ambient intelligence is built to amplify human judgment — to give a sales rep the preparation advantage of the most experienced colleague, to give an operations team awareness at a scale no human team could sustain alone, to give a field technician the answer that would have taken a phone call to headquarters.
It cannot be overstated that trust is the prerequisite for all of it. Ambient systems are always on, always listening, always observing. All of that requires the highest standard of data security, privacy governance, and enterprise-grade compliance. The deeper challenge, though, sits one layer below the technical safeguards. When AI is genuinely ambient, it inevitably encounters information that should not travel: a private comment in a sales call, a sensitive detail in a service interaction, a moment of frustration that wasn’t meant for the record. Teaching a system to recognize what’s shareable and what isn’t, and to act accordingly, is one of the open research questions our team is actively working on. The answer matters because no enterprise customer hands over the context that makes ambient intelligence possible without it.
The airport without signs
The future airport scenario in my introduction is not a science fiction take on travel. It is a model for what every enterprise environment is becoming. The most powerful version of ambient intelligence is the one that makes the signs unnecessary, because the system has already found you, already understood what you need, already delivered it before you thought to ask. The environment doesn’t announce itself. It simply works.
The arc of travel has been a steady march toward quieter infrastructure. The crowded train station became the modern airport. The departure board became the phone notification. And the phone notification, in turn, will dissolve into something more ambient still — present in the moment, absent the rest of the time, calibrated to the person and the work in a way that no broadcast system ever could be.
That’s the future our team is building toward. Always-on, aware, adaptive, and anticipatory enough to be useful; quiet enough to be welcome. Enterprise AI that recedes into the fabric of the work, giving people back something genuinely scarce: the space to focus on what only humans can do.
When AI becomes invisible, the technology has not failed. It has finally arrived.
I would like to thank Doyen Sahoo, Junnan Li, Daniel Lee, Zeyuan Chen, Juan-Carlos Niebles, Itai Asseo, and Karen Semone for their insights and contributions to this article.
