Are We Designing the Workplaces of the Future?

CBRE’s occupier surveys tell a consistent story once you look past the headline: the industry’s defining behaviour this cycle is not contraction, it is flight to quality. Two years ago 53 per cent of occupiers expected to shrink their footprint. Now it is 37 per cent, while the share expecting to grow has nearly doubled, from 20 to 38 per cent, contraction and expansion sentiment now within a single point of each other. Where portfolios are shrinking, the money is moving upmarket: better buildings, better locations, more expensive space, not simply less of it. Portfolio optimisation, cited by 80 per cent of CRE teams as their top goal, is the accurate description of what is happening. Blanket right-sizing is not.

Ask the same leaders what is reshaping their business fastest, and you get a different, equally consistent answer: AI. Not as a talking point. As capital. CBRE, JLL and Cushman & Wakefield have each committed real money to it over the past three years, and I will get specific about exactly how later in this piece. Whatever else is true of the industry’s AI strategy, indifference is not one of them.

Those two answers rarely get reconciled in the same room. The portfolio strategy answers the first question: better buildings, tighter footprints, higher-value space, sized for the way people work. It has almost nothing to say about the second: what happens to that carefully optimised floor plan once the way people actually work inside it changes. That is the gap this piece is about, are we designing space for how work happens today, or for how it is going to happen next?

I do not have a crystal ball, so let me be honest about what follows: it is an opinion, not a settled fact. My opinion is that growing or shrinking the footprint is not the question that matters. What it is being optimised for is, and I think most of the industry has not started asking it.

The paradigm we are optimising for is today’s, not tomorrow’s

Two things are true in workplace strategy right now, and put together they describe an industry solving for today with real conviction, and no plan for tomorrow.

Collaboration space is king. But who will we be collaborating with?

Open the current architect’s brief and you will find the same organising idea nearly everywhere: collaboration space is king. Strip out the rows of owned desks, add breakout, add huddle, add the coffee-and-sofas zone where the magic supposedly happens. It rests on one belief: that the collaboration worth designing for is human to human.

That belief is out of date. Collaboration is becoming hybrid, human and agent, and the interface to the agent is voice. Not because voice is fashionable, but because typing is the bottleneck: a keyboard is built for one person entering one stream of characters into one task, and it is no way to direct several agents at once. If a growing share of tomorrow’s collaboration is talking to agents by voice, continuously, is an open collaborative floor the right answer, or the wrong one? We built for a conversation that is being replaced, and did not build for the one replacing it.

We are optimising for the world of today, not tomorrow

Call it optimisation rather than contraction: fewer occupiers planning to shrink, more planning to grow, money moving toward better buildings rather than simply less space. That is a real, defensible response to how work patterns have shifted since 2020.

But optimise against what? The portfolio being reshaped is still sized for how people worked yesterday: typed, deskbound, mostly silent. The way people will direct AI is voice, and that is not a small interface change. It is a different physical activity happening at the desk, all day, and nobody has priced it into the plan.

AI for the CRE function, in service of what space?

CBRE, JLL and Cushman & Wakefield are genuinely changing the CRE function with AI. CBRE’s Nexus platform runs facilities operations across a billion square feet. JLL’s Falcon abstracts leases. Cushman & Wakefield’s AI+ platform and Microsoft copilots support brokers. Real investment, real change, and none of it is wrong on its own terms.

But a function exists to serve something, and here that something is space. CBRE’s own flagship 2026 workplace report mentions AI exactly once: as a tool to “optimise space, monitor utilisation and make informed decisions.” Optimise which space, though? The one built for yesterday’s way of working. Change the function without asking what the space needs to become, and all you have done is make yesterday’s floor plan cheaper to run.

If the space itself needs to change, and I think it does, the services the function provides need to change with it: not faster lease abstraction, but different criteria for what a lease should secure. Not smarter facilities management, but facilities management built for a different kind of collaboration. Nobody has made that connection yet. And the irony is that the infrastructure to make it, the geometry of the space itself, is exactly what these AI tools already run on.

The missing question: more space, or less?

Not how much space to remove. Whether work is about to need a different amount of space altogether, and in which direction.

Speech is the interface, and speech is not private

Voice wins the interface argument on pure efficiency: speech runs roughly three times faster than typing, in a controlled study across English and Mandarin. Anyone directing more than one AI agent through a keyboard already knows why that matters. A 2024 study on multi-agent interfaces found users straining just to manage which agent they were talking to. A slow input method on top of that does not help.

That is not just a lab result. Wispr Flow, a dictation app that pivoted from hardware to pure voice-to-text in 2024, is growing at more than 50 per cent a month and has raised $81 million to date. Founders do not chase a market that is not moving. And the market is already loud: most of us already do synchronous work by voice, over Teams and Zoom calls, all day. Add agents into that mix and the volume does not fall. It compounds.

But speech has a property typing never had to worry about: it cannot be private in a shared room. You can type a confidential instruction with your screen angled away, and nobody else on an open-plan floor is any the wiser. You cannot dictate one. The moment work becomes something you say out loud, it becomes something everyone near you can hear, whether they are trying to listen or not.

That would matter less if open-plan offices already handled speech privacy well. They do not. It is, by a wide margin, the thing open-plan gets worst. A 2013 study found up to 59 per cent of open-plan occupants dissatisfied with sound privacy, the worst-rated attribute measured, worse than temperature, light, or general noise. A 2025 follow-up, surveying 349 occupants across 28 offices, found that lack of privacy was a stronger predictor of acoustic dissatisfaction than the noise itself, by roughly 25 per cent. Leesman, the industry’s largest workplace-experience benchmark, finds the same pattern at far greater scale: noise and acoustic conditions are essential to 71 per cent of employees, yet satisfaction with them ranks among the lowest of any workplace attribute measured, and dissatisfaction with noise correlates more strongly with employees saying their workplace does not support their own productivity than any other factor. We built the open floor to host a conversation, and it has never actually been private enough to host one well.

From comfort problem to compliance problem

Layer continuous, spoken AI dialogue onto a floor that already fails at protecting an ordinary phone call, and the problem changes category: from comfort to compliance. GDPR treats speaker diarisation, working out who said what, as biometric data processing requiring explicit consent. The EU AI Act separately restricts biometric and emotion-recognition processing of voice at work. Client details, deal terms, personal data, said out loud to an AI agent at the next desk: that is not a noise problem, it is a data-loss-prevention problem. DLP is a term every regulated business already takes seriously for email and file transfer. It has never had to mean the air in the room.

I want to be precise about what I am claiming and what I am not. I am not saying every desk needs to become a phone booth. I am saying the acoustic-privacy failure of open-plan, already the single worst-rated attribute of the format on the evidence above, gets materially worse the moment speech becomes the primary way people work with AI, and almost no space plan I have seen has that variable in it at all.

The line items nobody has priced

Start with what already varies. Office space per person is not a fixed number: JLL and CBRE benchmarks put it at roughly 175 to 225 square feet historically in the Americas, closer to 210 in continental Europe, and as low as 140 in Japan and 50 in China. A North American office can already be 50 per cent larger than a European one, and more than 75 per cent larger than a typical Asia-Pacific office, for the same job. If space per person already swings by four times across markets for reasons of cost and culture, there is no principled reason it cannot move again for a reason of substance. The planning lever already exists. Nobody has pointed it at this question.

Second, the compute has to live somewhere. Conversational AI that responds at the pace of speech needs low-latency inference, and the industry is already moving that processing out of distant cloud data centres and closer to where it is used: inference now accounts for roughly two-thirds of all AI compute, up from a third in 2023, and a growing share of that runs at the edge. Not a return to the on-premise server room of old, but something like it: hub rooms, local inference capacity, technology infrastructure that has to sit inside the building because its occupants need answers in under a second. That is space too, and it has never been a line item in a workplace strategy before.

Third, the humans doing all this listening, talking and reading need somewhere to stop. Heavy AI use is already showing up as a measurable cost: researchers surveying nearly 1,500 workers found 14 per cent reporting a form of mental fatigue they call “AI brain fry,” and separate research found 45 per cent of frequent AI users reporting high burnout, against 35 per cent of non-users. Add continuous spoken interaction with agents on top of that, and recovery space, quiet, low-stimulation, off, stops being a wellness amenity and starts being an operating requirement. I do not know the ratio. I am fairly confident there needs to be one.

One more possibility

This one is an opinion, not a finding. If AI accelerates decisions but still needs a human in the loop for judgement, and if that judgement is increasingly the valuable part of the process, some of the work currently offshored for cost reasons might need to move back to where the decision-makers actually are. There is early, real signal in this direction: some organisations are already reshoring AI-oversight, orchestration and quality-assurance roles, though the honest finding is that they come back as small, high-skill teams working alongside AI, not a one-for-one return of the offshored headcount. If that pattern holds and spreads, it does not mean more space everywhere. It means more space in fewer, higher-cost locations, for a smaller number of people doing more consequential work, a different planning problem to the one most portfolios are solving for today. I have not seen this connected to space planning anywhere else. Take it as worth testing, not as a conclusion.

So which metrics matter now?

This is where I think the industry’s current toolkit runs out. Utilisation, the metric behind almost every right-sizing decision right now, answers one question: how many desks were occupied. It says nothing about how many of those occupied desks were hosting a spoken, confidential conversation with an AI agent, or whether the space around that desk could actually contain it. It is a post-COVID metric, built to answer a post-COVID question, attendance, and it is being asked to carry a workload it was never designed for.

If voice-first, human-plus-agent collaboration is coming, the metrics that matter change: enclosed-versus-open seat ratio, some measure of speech-privacy provision per desk, eventually perhaps something like agent-conversation minutes per floor. None of those exist as standard reporting today. That is not a criticism of the teams running utilisation surveys. It is a gap nobody has been asked to fill yet.

Where is the leadership?

These are lock-in decisions on a ten-year clock, capital committed once and lived with for a long time. They deserve evidence, not instinct and not last cycle’s benchmark. So: where is the leadership actively deciding whether the AI-era floor needs more enclosure, not just fewer desks? Has your design standard changed? Your space metrics? Has the direction, more or less, even been discussed as a live question, rather than quietly assumed?

I have given you my instinct here, and I have tried to back it with real research rather than just conviction. But an instinct, however well sourced, is not evidence. The evidence to actually test this properly already exists, sitting in a stack most real-estate teams have never thought to point at a question like this one. Testing it, not trusting my instinct, is next.

The conclusion

What I am fairly sure of: whether your footprint is growing or shrinking is not the question that matters, CBRE’s own occupier data now has those two groups within a single point of each other, 37 per cent against 38 per cent. What matters is what that space is being optimised for. The industry is confidently optimising for a paradigm, human-to-human collaboration on an open, densified floor, that is already being replaced, using a metric, utilisation, that was built to answer a different question. My instinct is that more space, not less, will be required overall, at least of the enclosed, acoustically private kind voice-first collaboration needs. But more space only helps if it is the right space: getting bigger without getting the design right just repeats the same mistake at a larger scale.

What I am not sure of: exactly how much more space, what ratio of enclosure to open floor, whether this is a five per cent correction or a reversal of a decade of densification. I have an instinct, not a number, and I am wary of anyone, including myself, who claims more precision than the evidence currently supports.

What I am confident of is the shape of the mistake if we get this wrong twice. We already optimised once, fast, on good evidence, for the world that was leaving. Doing it again, just as fast, without asking what the world that is arriving actually needs, would be the same mistake made with better data behind it.

The way through is not more conviction, mine or anyone else’s. It is data-informed design: using evidence of how people actually work to decide how much space the AI-era workplace needs, and what kind, rather than instinct or last cycle’s benchmark. That discipline, not this opinion, is what should actually size the next floor plan.

I want to hear from people already thinking about this properly: acousticians, workplace strategists, anyone who has run a spoken-AI pilot on a real floor and found out what it does to the room. If that is you, reach out.

Frequently Asked Questions

Mark Cunningham is the founder of Insights². He has spent more than a decade building data and analytics products at the intersection of data and corporate real estate. He writes at insights-2.com about how to make the meaningful measurable.