AI Has Changed the Definition of a Connected Building
I spent hours in my own basement, on the phone with my internet provider, because the basement is where the modem is.
There is no landline in the house. We cut the cord years ago, the way most people have, which means the house runs on the connection I was calling about, and the phone I was calling on runs on cellular. The basement is fully below grade. The modem sits in the center of it, which is the right place for the equipment and the wrong place for a person holding a phone.
The call dropped. I went up the stairs to find a bar, redialed, and came back down to read the lights on the box. Then it dropped again. By the clock it took hours. By my estimate at the time, considerably longer.
None of that was a technology failure. It was concrete, and grade, and the fact that the one room I had to be standing in was the one room the signal could not reach.
That is a house, and the stakes were an evening. The same scene plays out in apartment buildings, office towers, hospitals, and every other kind of property, every time somebody walks into a room that nobody thought about for connectivity when it was built. Those rooms are where the equipment lives, which makes them where people go when something has gone wrong.
Every conversation about AI runs on the same short list. Model capability. Compute. Data quality. Algorithms. All of it real, all of it worth arguing about.
None of it is what stops AI at the front door of most buildings.
That is connectivity. Secure, reliable, always-on cellular connectivity, inside the buildings where work and care and commerce actually happen. AI started as a tool for people sitting at desks. It is becoming an operational layer in nearly every job and nearly every room. The assumptions we made about in-building network access were written for the first version and never updated for the second.
Wi-Fi was not built for the world AI is building. More to the point, it was never built to serve everyone in the building. In-building cellular was.
That is not an argument for ripping out Wi-Fi, and it belongs at the top rather than buried at the end. Wi-Fi is the organizational network. It is good at that job and most buildings will keep running it. What buildings do not have is a second layer that works without organizational membership, and cellular is the one already sitting in most people's pockets.
The gap is not new. It is newly expensive. For more than a decade the industry's working estimate has held that 70 to 80 percent of wireless data traffic originates or terminates indoors. Read that as a long-standing estimate rather than a current measurement. The precise figure matters less than the direction, which nobody disputes: most wireless activity already happens inside. Research CommScope commissioned from Coleman Parkes in 2015 found that roughly two percent of commercial buildings had a dedicated in-building cellular system. Almost all of the demand is inside. Almost none of the infrastructure is.
That gap sat there for a decade and cost almost nothing, which is the honest reason nobody fixed it. A dead zone in a plant room was an annoyance. AI did not create the gap and it does not change the physics. It changes what falls into the hole.
Every Building Has Three Populations: In, On, and For
Why staff, contractors, and the people a building serves have different connectivity needs
Here is the part most technology plans miss. This is a change in how every person in a building interacts with that building, no matter who signs their paycheck or why they walked in, rather than an upgrade for a defined group of workers.
Walk any complex building on a Tuesday and count who is actually there. Three populations show up, every time, and the only thing separating them is one small word.
In. The people who work in the building. Permanent staff. Clinical teams. Knowledge workers, managers, operators. These are the people enterprise technology budgets were written for, and they are already getting AI: ambient documentation, augmented workflows, decision support. Their need for connectivity is understood and funded.
On. The people who work on the building. Contractors, trades, facilities technicians, security, service vendors, maintenance crews. They keep the place running. They also work in the worst rooms in the building: plant rooms, service corridors, roofs, sub-basements, loading docks, active construction zones. They are not on the corporate network. They are not on managed devices. And they are showing up with AI of their own: assisted inspection and diagnostics, field service management, augmented reality guidance during a repair, automated compliance capture. They need connectivity that follows them, not connectivity they have to ask permission for.
For. The people the building works for. Patients. Customers. Guests. Fans. Students. Residents. They arrive with their own phones, their own apps, and now their own AI. A patient running a health app during a visit. A traveler with an AI concierge. A student with an AI tutor between classes. What they can do inside your walls is decided by what your building lets through.
In, on, and for. Three populations, one building, at the same time. Almost every network in almost every building is designed for the first one.
Wi-Fi covers that first group well, because that group is what it was designed around. Serving the second and third means provisioning people the building does not employ, forever. Cellular reaches all three for one reason that outweighs the rest: it does not ask anyone for credentials. It follows the device instead of the location. It is what everyone's phone already speaks.
Which points at the real argument, and it is worth planting here rather than saving. The problem is not that buildings have bad Wi-Fi.
The problem is that connectivity has been treated as a tenant service while it was quietly becoming a property characteristic.
Everything below is the case for that sentence.
AI Is Not a Desk Job Anymore
Where in-building AI is actually deployed: hospitals, industrial sites, venues, and commercial real estate
The first wave of AI tools landed in the easiest possible place. Copilots, summarizers, code generators, delivered to people sitting still. Access points in the ceiling, fiber to the building, connectivity solved because the work never moved.
The work moves now, and most of it does not look like an office.
Hospitals need coverage for all three populations at once. Diagnostics at the point of care, ambient clinical documentation, AI-guided surgical systems, continuous patient monitoring. Every member of the care team needs signal. So does the vendor servicing the chiller, and the patient who arrived expecting to participate in her own care.
Manufacturing and industrial sites run AI where the coverage is worst. Quality control, predictive maintenance, autonomous mobile robots, machine vision. Picture the vibration analyst who comes in twice a year on the packaging line. Her diagnostic tool now streams sensor data to a model, compares it against a fleet baseline, and returns a bearing recommendation before she leaves the floor. She is standing between a conveyor and a concrete wall, forty feet from the nearest access point, on a network she will never be issued credentials for. The tool is only as good as the bar on her phone. When she loses it, the finding goes in a notebook and gets typed up in the parking lot, and the model gets none of it.
Venues concentrate the entire problem into ninety minutes. Operations staff, event contractors, security, catering, and sixty thousand guests, all on AI-connected devices, all at once, in a structure built out of concrete and steel.
Commercial real estate is being asked to deliver it for everyone else. Office towers, mixed-use, campuses, retail floors, hotels. Owners now field requirements for tenant collaboration tools, tenant operational AI, building automation, and every person who walks through the lobby.
The requirement is not set by the employee roster. It is set by everyone who is in the building.
The Wi-Fi Assumption
Why the Wi-Fi assumption breaks outside the enterprise, for contractors, visitors, and AI workloads
There is a line in almost every enterprise technology plan that nobody writes down, because everybody assumes it: inside the building, Wi-Fi handles it. Commodity layer. Solved problem. Next slide.
That assumption held when the heaviest thing on the network was email and a browser, running on a known set of managed laptops. It does not hold when the network carries real-time AI inference for a crowd of workers, contractors, and visitors who each brought their own device and their own software.
Wi-Fi's limits here are not tuning problems. They are architecture.
Wi-Fi requires membership. Connectivity comes from joining a specific network. You can provision staff. You cannot practically provision contractors, vendors, and guests without guest network management, credential issuance, and segmentation, and then maintain all of it forever. When AI use goes universal, that overhead stops being an annoyance and becomes unworkable.
Coverage breaks in real buildings. Hospitals, plants, warehouses, and mixed-use campuses are not open-plan rectangles. Stairwells, elevator shafts, sub-basements, loading docks, plant rooms, and radio-hostile construction produce dead zones, and they produce them exactly where mobile workers spend their day. The physics are measured, not anecdotal. Report ITU-R P.2346, the international compilation of building entry loss measurements, runs from 10 to 25 dB of attenuation through wood up to 35 to 40 through metal. NIST measured the same effect material by material in Electromagnetic Signal Attenuation in Construction Materials (NISTIR 6055). A building does not have to be trying to block signal to block it.
Handoffs interrupt. Cross a large facility and Wi-Fi passes you from access point to access point. Mishra, Shin and Arbaugh instrumented the process at the University of Maryland for ACM SIGCOMM Computer Communication Review in 2003 and measured a range running from roughly 40 milliseconds to nearly 600, well past the 50 millisecond ceiling voice traffic tolerates. Fast-roaming extensions have narrowed that considerably since. What they have not changed is the architectural distinction: Wi-Fi mobility depends on roaming between local access points, where cellular mobility is fundamental to the network design. For AI that holds state, a clinical note in progress, a monitoring stream, an agent mid-task, an interruption is not an inconvenience. It is a failure.
The security perimeter is hard to hold. Multiple user classes, personal devices, contractor hardware, and visitor equipment on one enterprise wireless network produce a posture that is complicated to run and worse to audit. Add regulated data and the complexity becomes liability.
Shared spectrum buckles under density. Wi-Fi runs on unlicensed spectrum by design, at 2.4 and 5 GHz, and at 6 GHz with Wi-Fi 6E and Wi-Fi 7. Unlicensed means contested. On a full hospital floor, a packed concourse, or a busy plant, performance falls off exactly when the AI workload is heaviest and matters most. High-density testing shows the curve is not gentle. A high-density study from Alethea, a wireless test and measurement firm whose customers are the people who build access points, measured video quality holding steady to 75 clients and then collapsing from 100 percent to 30 percent by 100. Lab testing rather than peer-reviewed work, so read it as directional, and note which way it cuts. The degradation is non-linear, which is another way of saying the network looks fine right up until it does not.
The Building With a Dozen Networks and No Signal Where It Breaks
Why multi-tenant properties fail all three populations at once
Multi-tenant properties break the assumption that every argument about in-building Wi-Fi quietly depends on: that the building has one wireless network, with one owner, covering the whole property. That holds on a single-tenant corporate campus. It is close to false everywhere else.
Walk a mixed-use development and count the networks. The retail tenants run their own. The residential owners run theirs. The office tenants run theirs, and the hotel runs another. The property's own maintenance team belongs to none of them. They pass a dozen networks they have no standing to join, in a building their employer owns.
Owners do fund wireless in their own buildings. Corporate networks for their office employees, and common area Wi-Fi for everybody else. Both are real, both get budgeted, and neither was ever meant to reach the plant room. Coverage stops where the payroll stops and where the public gathers, which are the two places a business case can see. That is not negligence and it is not a fault anybody can file a ticket against. It is the ownership structure working exactly as designed, and it does not resolve just because the owner and the occupant are the same company.
Then consider where those people actually work. Basements, stairwells, IT closets, mechanical rooms, back-of-house corridors, the roof. No tenant has a reason to cover any of it. The owner usually has not either, because the business case for Wi-Fi was written around the people who pay rent rather than the people who keep the building standing. So the maintenance team troubleshoots the failure in the one room in the building with no signal at all, which is also the room where things break.
None of this makes Wi-Fi a bad technology. It makes Wi-Fi a technology built for a different job: managed devices, belonging to one organization, in reasonably friendly space. That job still exists. It is no longer the whole job.
Bring Cellular Inside
How distributed antenna systems, private cellular, and neutral host serve every device
The answer is not more access points. The answer is putting cellular inside the building, through a distributed antenna system, a private cellular network, or a neutral host that carries multiple operators, and designing the in-building network on what cellular already does well.
Those three are not interchangeable, and the distinction decides which populations get served.
Neutral host carries the public carriers, which is what the people who work on your building and the people it works for arrive holding. Private cellular on CBRS gives an operator dedicated control over the devices it provisions, and it reaches anybody else's phone only through roaming agreements or a second profile on the device. A building that intends to serve all three populations generally ends up running both.
It serves the devices people already carry. No membership required. Anything with a cellular radio and an active SIM or eSIM connects: phones, tablets, connected tools, sensors. The contractor's field application works in the plant room. The patient's health application works in the basement radiology suite. No credentialing and no guest network. Devices with no cellular radio still need Wi-Fi, which is another reason most buildings run both.
Continuity is a design goal rather than an extension. Cellular is built to hand a session between cells, floor to floor, wing to wing, patient room to corridor to procedure suite. Cellular handoffs fail too. The difference is architectural: moving a live session between cells is what the system was designed to do, where in Wi-Fi it is an extension added to a protocol that re-associates. Stateful AI depends on that continuity.
Performance becomes contractible. 3GPP, the body that writes the cellular standards, defined network slicing in TS 23.501, the stage 2 system architecture for 5G, and has continued to build out ultra-reliable low-latency communication, or URLLC, across subsequent releases. What any given building can commit to depends on the architecture, the spectrum, and the operator, and the envelope is a negotiated number rather than a property of the technology. That it is negotiable at all is the point. It makes the target acceptance testing under a different name: a written tolerance the system either meets or does not. For latency-sensitive work like autonomous control, safety monitoring, and live decision support, the gap between best effort and guaranteed is the gap between a demo and a system people rely on.
Security sits in the architecture rather than on top of it. SIM-based device authentication under the 3GPP authentication framework, an encrypted air interface, licensed or coordinated spectrum, and isolated network boundaries. That baseline applies to every device the network admits, not only the ones IT issued.
Spectrum is no longer the barrier it was. In the United States, the Citizens Broadband Radio Service opened the 3.5 GHz band under FCC Part 96. An enterprise can now run its own cellular network in coordinated spectrum without owning a national license.
Side by Side
Wi-Fi versus in-building cellular compared across six dimensions
Dimension | Wi-Fi | In-building cellular |
Who can connect | Anyone the building provisions | Any device the network admits. On neutral host, that is every carrier subscriber who walks in |
Hard spaces | Coverage depends on access point placement in space nobody planned to cover | Coverage is engineered against the structure, room by room |
Movement | Handoff behavior depends on the Wi-Fi design | Mobility between cells is native to the architecture |
Performance | Best effort on contested unlicensed spectrum | Negotiated envelope via slicing and quality of service, subject to architecture and operator |
Security baseline | Per-network credentials, varying by device class | SIM authentication and encrypted air interface, uniform across admitted devices |
Administration | Per-user provisioning, guest management, segmentation | Device authentication handled by the cellular network |
This is not an argument for ripping out Wi-Fi. Most buildings will run both. It is an argument about which layer carries the traffic that has to work for everyone.
What It Costs, Honestly
The design, coordination, timeline, and capital tradeoffs of in-building cellular
In-building cellular is not free and it is not fast. It requires a radio design against the actual building rather than a heat map, coordination with the carriers whose subscribers you want covered, a spectrum decision if you are going private, sign-off from the authority having jurisdiction, and a deployment timeline measured in months. That design work is the whole promise, and it is worth being plain about the limit: no wireless architecture eliminates dead zones by category. Coverage gets engineered into specific rooms or it does not exist in them. Adding access points is cheaper on Tuesday and available on Friday.
The capital question is separable from the engineering question. A network-as-a-service structure moves the system to operating expense, with the owner-operator carrying the equipment, the monitoring, the upgrades, and the recertification, and a building that already owns aging equipment can sell it back into that structure. None of that makes the design work smaller. It does mean the up-front number is a decision rather than a gate.
Here is the part that decides it. The access points you add on Friday do not solve the credential problem, and the credential problem is the one that locks out two of your three populations. You can spend the cheaper money repeatedly and never close the gap, because the gap is architectural.
There is precedent for treating a signal as something you build rather than something you assume. Emergency responder radio coverage systems, known in the codes as ERRCS, are mandated across the country, coordinated with local authorities, and subject to periodic compliance testing. If your building has one, you have already accepted the premise. You engineered a signal into the structure because the consequence of its absence was unacceptable, and you agreed to prove on a schedule that it still works. Nobody argued that one on payback. The question now is which other signals have crossed that line.
The Stakes Went Up
Why in-building connectivity security became a governance and safety concern
In-building connectivity security gets filed as an IT concern. In the AI era it is a business, safety, and governance concern, because the network is no longer carrying email for employees. It is carrying operational decisions for everybody in the building.
AI systems act now. A system that recommends a treatment, watches a production line, walks a technician through a repair, or approves a transaction is not a passive tool. It holds decision rights somebody transferred to it, usually without a meeting, and the network underneath it is part of its risk profile. So the question is whether the in-building network can be trusted to carry that work reliably and securely for every person in the building, including the people who will never hold credentials on it.
What This Changes
What in-building cellular means for owners, occupants, trades, and the people a building serves
Stop treating in-building connectivity as an IT utility provisioned for a list of corporate users. Start treating it as shared infrastructure that decides who gets to use AI inside your walls. That changes how buildings are designed, valued, and run.
Owners and operators. In-building cellular becomes core building infrastructure, as basic to the asset's value as power, water, and physical security. In the same Coleman Parkes research, building and real estate professionals estimated that dedicated indoor wireless coverage could raise a property's value by 28 percent on average.
That is a survey response rather than a measured transaction premium. It records what people who buy, lease, and manage buildings believe coverage is worth. I spent years at Gartner watching what that kind of belief does here, across hundreds of organizations rather than one, and the pattern held every time: perception moves leasing markets years before it shows up in a comparable, because a tenant does not need a validated premium to walk away from a building. A number that describes what buyers believe is the earlier kind of evidence, not the weaker kind. And the value of a building feature does not have to show up as a premium. Often it shows up as a discount avoided, which is the same money arriving in a form no comparable records.
And the tenant does not phrase it as a network problem.
They phrase it as "our people cannot work in your building."
Enterprise occupants. Coverage extends the return on the AI you already bought, out to the contractors who keep you running and the people you serve.
Trades and field services. Signal in the plant room is becoming a condition of the job. AI-assisted diagnosis, documentation, and delivery are changing skilled trades work, and none of it runs on a dead zone.
The people you serve. A patient managing a chronic condition, a guest using a concierge app, a student pulling up a tutor between classes. Every one of them notices where your coverage stops.
Here is the claim, and I am putting a date on it so it can be held against me. By 2030, in-building cellular coverage will appear in commercial lease negotiations and institutional acquisition diligence as routinely as HVAC capacity and life safety systems do today. Buildings that cannot answer the question will trade at a discount to buildings that can, and nobody will call it a technology discount.
Things arrive on that list in one direction. Something becomes a condition of doing business for one large tenant, then it becomes a question on a form, then it becomes a line item nobody remembers adding. Nothing comes off the list. Connectivity is somewhere between the first step and the second right now.
Five Questions
How to assess whether your building has an in-building connectivity gap
Run these against your own building before anyone shows you a proposal.
Where does coverage fail? Name the four worst rooms in the building. Plant room, sub-basement, loading dock, stairwell core. If you cannot name them, ask the facilities team, because they can.
What does a contractor do to get online? Count the steps. If the answer involves a front desk, a form, or a rotating password, you have friction that scales with every AI tool your vendors adopt.
Does a session survive a walk? Start something stateful on floor three, walk to floor seven, and see whether it is still running. That single test predicts most of what follows.
Which AI tools already assume a connection? Inventory what is deployed, what is piloting, and what your vendors are bringing with them. That is your actual demand curve, not the roadmap.
Who is on your network at peak, and who is not? Count all three populations on the busiest hour of the busiest day. The gap between that number and your provisioned users is the size of the problem.
The Building Is the Network
Why in-building cellular decides who can use AI
Every platform that mattered ended up serving everyone. Electrification did not stop at certain rooms. The telephone network did not serve only certain companies.
AI is on the same road, and the part decided inside your building is simpler than the models. Everyone who walks in is an AI user now. They carry AI-connected devices, they rely on AI-assisted services, and they notice immediately where your building stops supporting both. Wi-Fi answers that for the people the enterprise employs. It was never asked to answer it for the other two populations.
My basement solved itself when I walked up the stairs. That option exists in a house. It does not exist in a plant room, a sub-basement, or a radiology suite, because the person with the problem is the person who has to be standing in the room, and the work does not follow her to where the signal is.
Which leaves the part that makes this genuinely hard, and it is not the radio design.
The owner controls the building.
The tenant controls the network.
The contractor operates the equipment.
The patient, the guest, and the student generate the demand.
The AI application belongs to a vendor none of them have met.
The person paying for the connectivity is not the person who needs it most, which is exactly why the business case for Wi-Fi was written around the people who pay rent rather than the people who keep the building standing.
Nobody in that chain is behaving unreasonably. Each of them is optimizing the thing they are accountable for, and the gap opens in the space between them, in the rooms none of them own. It stays open until somebody decides connectivity is a property of the building rather than a service of a tenant.
That decision is not an IT decision, and the people who make it do not report to a CIO.
The AI opportunity is universal. The network inside the building has to be too.
A shorter version of this argument, written for facility managers, ran in IFMA's FMJ as In, On & For.
There are four seats in the name. The fifth one is yours.
Comments