The Integration Gap: Why a Decade of PropTech Hasn't Closed CRE's Space-Utilization Problem — and What an Agent Platform Does Differently

📄 Formatted PDF editions (English · 繁體中文) are free for subscribers — subscribe to get them →

Executive Summary

Commercial real estate has a paradox it has not been able to spend its way out of. In 2025, investors put US$16.7 billion into property technology — a 67.9% jump year over year — and the pace accelerated into 2026, with January alone drawing US$1.7 billion, up 176% on the prior January (figures aggregated by JLL/CBRE research, reported in AISB's corpus, June 2026). Roughly 70% of recent PropTech deals now carry an AI component (CBRE/JLL deal data, per AISB corpus aggregation, June 2026). By any measure of capital and ambition, the industry has committed to fixing how buildings are run.

The largest, best-documented inefficiency in the asset class has barely moved. Global office utilization sits at 53–56% — JLL's 2026 Global Occupancy Planning Benchmark puts it at 56%, CBRE's at 53% — against space that, in most portfolios, is leased near full capacity. Corporate offices specifically run 40–50% utilization on independent sensor data (Occuspace 2024; Density Q1 2025). The gap between what is leased and what is used is where the money sits, and it has stayed open through the entire PropTech build-out.

The reason is not bad software. It is architecture. JLL's own research finds 92% of CRE organizations have piloted AI tools, but only 5% report achieving their stated objectives — and the corpus has a name for the pattern it produces: the "90-day adoption cliff," where firms buy a license, announce a deployment, and watch adoption stall within three months. Purchasing is not operationalizing. Every tool in the stack optimizes a local decision against a local data stream: the occupancy sensors know how full the floor is, the building management system knows the HVAC state, the booking platform knows the reservations, HR knows headcount, the lease platform knows the expiries — and none of them are connected. The insight that would change a decision never forms.

The same disconnect shows up even where the technology demonstrably works. US Department of Energy / Lawrence Berkeley National Laboratory field data across 60,000+ building units found roughly 40% of air-handling units and 30% of terminal units carry a fault on any given day, yet median realized savings from fault-detection tools is only ~9% against the 15–30% of HVAC energy typically recoverable. Detection is solved. Closure is broken. That is a coordination failure, not a sensing failure — and it is the same shape of problem as the utilization gap.

There is, encouragingly, a named destination. In Emerging Trends in Real Estate 2026, PwC and the Urban Land Institute described what they call a "property operating system" (propOS) — "AI agents, digital twins, and data integration layers hovering above the legacy platforms they aim to supersede," drawing data through APIs and shifting "from batch processing to real-time streaming analytics." Two of the field's most authoritative institutions have, in effect, sketched an agent platform. The strategic question for an owner-operator is no longer whether to move toward that architecture, but how to reach it without hand-assembling seventeen vendors that will not connect — something the point-solution market, by its structure, does not deliver on its own.

The bottom line: the gap is not a shortage of tools — the industry has bought roughly seventeen of them. It is the absence of a trustworthy, auditable, integrated operating layer that turns fragmented data and one-off AI pilots into decisions that compound. Building that layer tool-by-tool is the slow path. A purpose-built CRE agent platform — pre-integrated across the five data silos, sharing one provenance-tracked knowledge layer, and built so occupancy intelligence can be acted on rather than merely displayed — is the faster one. This report quantifies the pain, dissects why it persists silo by silo, maps the agent-platform answer to the specific decisions it has to improve, and prices the opportunity with an honest accounting of what we can and cannot yet prove.

1. The Pain — the leased-versus-used gap

1.1 The market by the numbers

Headline office vacancy is at multi-decade highs, and the major data providers disagree on the exact figure — which is itself instructive:

• CBRE reported 18.6% overall US office vacancy in Q1 2026, separating out 12.7% for prime/Class A stock.
• JLL reported 22.5% national vacancy in its Q3 2025 US Office Outlook.
• Moody's CRE Analytics put Q2 2025 vacancy at 20.6%, the highest since the late 1980s.

The 17.6%–22.5% spread reflects timing and methodology (prime versus total stock, sublease treatment), not measurement error. The direction is unanimous, and the recovery is genuine but partial: CBRE counted six straight quarters of positive net absorption (~127 million sq ft) through Q1 2026.

Vacancy, though, is the landlord's problem. The deeper and more universal problem — the one that touches every occupier in any market — is utilization:

• Global building utilization: 53–56% (JLL 2026 Benchmark: 56%, up from 54% the prior year; CBRE measure: 53%). The pre-pandemic baseline was ~61% (JLL/CBRE pre-pandemic series).
• Corporate offices specifically: 40–50% (Occuspace 2024 Workplace Utilization Report; Density Q1 2025 reported average peak occupancy of 47%).

A caution worth stating plainly, because a sophisticated reader will raise it: these utilization figures are sensor- and badge-sampled, and definitions of "occupied" differ across providers (a badge swipe, a desk-sensor heat signature, and a calendar entry are three different things). Read them as a consistent directional picture — a large share of leased, conditioned, cleaned, and insured space is empty on a typical day — not as a precise census.

1.2 The hybrid pattern broke static planning

The averages hide the structural change that makes this a new problem rather than an old one. Attendance has concentrated:

• Employees attending 3–4 days per week rose from 36% to 55% in a single year (+19 points; JLL 2026 Benchmark).
• That demand is not evenly distributed: Tuesday occupancy reaches 73% of capacity while Monday and Friday average below 40% (JLL data, AISB corpus, May 2026).
• 87% of organizations now set explicit utilization targets, and nearly half aim for the 76–85% band (JLL 2026).

This is the crux. As AISB's corpus put it in May 2026: "average utilization is no longer the binding metric; peak-day capacity versus mandated-day demand is." A portfolio planned to a static headcount is simultaneously over-built for Friday and under-built for Tuesday. A 1,000-desk floor at 50% average utilization sounds half-empty — but if Tuesday runs 73% (730 desks) and Friday 34% (340), the floor is nearly full on the day that matters and a ghost town on the day that doesn't. Plan to the average and you under-provision the peak; plan to the peak and you pay for Friday's emptiness all year. Neither static answer is right, which is why the decision has to be continuous and data-driven — and why it requires the demand signal (booking + badge), the supply signal (lease + space), and the comfort signal (BMS) to be read together, which in most portfolios they never are.

1.3 The same problem, three regions, three different levers

"Office" is not a single market, and the utilization problem wears a different face — and demands a different lever — by region.

United States — the give-back market. With vacancy at 18–22% and roughly half of prime space underutilized (JLL), the dominant lever is recovering space: giving back floors, subletting, and consolidating. The hybrid Tuesday-peak pattern above is drawn primarily from US/UK data, where it is most pronounced.

APAC / Singapore — the over-leasing-avoidance market. Singapore CBD attendance has held closer to 70–75% of pre-pandemic full-week levels in Grade A stock (JLL APAC commentary, 2025) — smaller dwellings and dense MRT access mute the Tuesday-trough. But Singapore CBD Grade A vacancy was just 4.1% in Q1 2026 (Cushman & Wakefield), Marina Bay 6.0% (JLL Q4 2025), and new office construction is projected to hit a multi-decade low (Knight Frank / JLL Singapore pipeline commentary, 2026). In a supply-squeezed market an occupier cannot easily give back space, so the value of utilization intelligence shifts to not over-committing — right-sizing renewals and expansions so the firm doesn't lock in premium Grade A rent for capacity peak demand never touches.

Europe — the compliance-coupled market. In much of Europe the utilization decision is increasingly entangled with energy-performance regulation (EPBD recast and national MEES-style minimum standards), so the same occupancy data that drives a give-back also feeds a decarbonization-compliance case. The lever is dual-purpose: every square foot rationalized is both a rent saving and a carbon-and-compliance saving.

The unifying point: the measurement is identical everywhere — peak-day demand versus leased supply — but the action (give-back vs. avoided over-leasing vs. compliance-coupled rationalization) is regional. A single global dashboard that reports one number cannot serve all three; the optimization has to be decision-aware and locally grounded.

1.4 What the gap costs — by building type

At NYC prime rents of US$70–100/sq ft per year (CBRE US Office Figures Q1 2026; CoStar NYC), and with JLL estimating roughly half of office space underutilized in prime markets, the arithmetic is stark. A 100,000 sq ft occupier paying $80/sq ft spends US$8 million a year on its lease, a meaningful fraction of it buying empty space. The arithmetic of recovering a tenth of that footprint is straightforward — 100,000 sq ft × $70–100/sq ft × 10% = US$700,000–1,000,000 per year in rent efficiency (an illustrative, independent calculation from the rent inputs; the realizable portion depends on lease structure and the ability to act on freed space), before energy, services, and churn. Little wonder 55% of organizations are actively cutting their physical footprint (JLL, 2025–2026).

But "office" is not one thing, and the cost — and the right intervention — varies sharply by archetype.

Asset archetype	Illustrative rent	100k sq ft lease	Value of 10% util. gain/yr	Dominant lever
CBD prime tower (e.g. Midtown, Marina Bay)	$80–100/sq ft	$8.0–10.0M	$800K–1.0M	Give-back / sublet
Suburban / campus	$30–45/sq ft	$3.0–4.5M	$300–450K	Consolidation / densification
Supply-tight APAC prime (SG)	~$90/sq ft (illustrative)	~$9.0M	~$900K	Avoided over-leasing on renewal
Life-science / lab-office hybrid	$60–110/sq ft	$6.0–11.0M	$600K–1.1M	Re-zoning bench vs. desk

(Rents are illustrative market ranges for sizing, not quotes for a specific building; the APAC figure is an explicit assumption, not a sourced rate.) An average cannot make this point: the intervention differs by archetype, which is precisely why a one-size dashboard underperforms and why the optimization has to be decision-aware.

1.5 The cost of inaction compounds

The numbers above are annual, and that is how they are usually discussed — which understates the problem, because a lease is not an annual decision. Office leases run five, ten, sometimes fifteen years, and the integration gap quietly bills the occupier for every one of them.

Take the 100,000 sq ft prime occupier again. Carrying a 10-point utilization gap costs on the order of $700,000–1,000,000 a year in rent efficiency (Section 1.4). Across a ten-year lease term, that is $7–10 million of cumulative spend on space that, on a typical day, no one uses — a number that rounds, for a CFO, to a full year of the lease purchased and discarded. Layer in the energy and operations lines (Section 6) and the figure climbs further. None of it appears as a line item, which is exactly why it persists: there is no invoice that says "empty space," only a lease that renews.

There is also an opportunity cost that does not show up in any spreadsheet. The operator who cannot see peak-day demand cannot confidently consolidate, cannot negotiate a renewal from data, and cannot tell a board whether a return-to-office mandate needs more space or less. Decisions get made on instinct and politics because the data to make them on evidence was never assembled. In a market where 55% of organizations are already cutting footprint (JLL), the firms that move on evidence will rationalize the right floors; the firms that move on instinct will cut the wrong ones — and discover the error only at the next renewal, when it is expensive to reverse.

The case for action is therefore not merely "save money this year." It is "stop compounding a structural cost across the lease term, and earn the ability to make the next space decision from evidence rather than instinct." That is a CFO-grade argument, and it is the one the integration gap has been hiding.

2. Why it persists — the five-silo data problem

2.1 The architecture of fragmentation

To make one intelligent occupancy decision — should we give back a floor, re-stack the building, or change the booking policy? — an operator needs five data streams to agree on reality. In most portfolios they live in five separate worlds, each with a different owner, format, and refresh cadence.

Each silo has, over the last decade, acquired its own analytics layer and now its own AI. But an AI reasoning over one silo reasons from a partial picture. The occupancy sensor's model can say a floor is 45% full; it cannot know the lease on that floor expires in fourteen months, that HR is hiring 200 people into that business unit, and that the 45% is a Friday artifact masking a 78% Tuesday. The decision that needs all five never assembles — not because any one tool is bad, but because no tool owns the join.

2.2 The five silos, one by one

The fragmentation is concrete and worth walking through, because each silo fails to share for a different structural reason — which is why no single integration trick fixes all five.

1 · BMS / BAS (the building's nervous system). Owned by engineering; holds HVAC, lighting, and energy-metering data. The dominant platforms — Siemens Desigo CC, Honeywell Forge, Johnson Controls Metasys — increasingly publish open APIs. The integration challenge is less any one vendor than the data layer itself: building-systems data often speaks BACnet or device-specific dialects, and even where BACnet is nominally "open," object models and point-naming are configured per building, so two towers — even on the same product line — rarely expose data identically. Every integration is therefore partly bespoke. The industry's historical default treated building data as a by-product of operations rather than a shared asset; that posture is shifting, but slowly.

2 · Badge / access control (the richest signal, and the untouchable one). Owned by security/IT. A badge swipe is a high-confidence signal that a person was physically present, where a desk sensor infers presence indirectly and a calendar entry records only intent. It is also the most legally fraught: badge data is personal, sometimes biometric, and walled off from facilities by policy and privacy concern. The richest occupancy truth is the one most organizations are not allowed to use.

3 · Booking / calendar (intent, not attendance). Owned by workplace/IT (Condeco, Robin, Microsoft/Google calendars). It captures what people meant to do — and people book rooms they don't use and show up to desks they never booked. On its own, booking data systematically overstates demand; only when reconciled against badge and sensor data does it become trustworthy.

4 · HR / headcount (the who and how many). Owned by HR (Workday, SAP). Knows team sizes, hiring plans, and org structure — the forward-looking demand signal — with no native line to where those people sit. The lease decision needs HR's growth plan; HR's system has no concept of a floor plate.

5 · Lease / financial (cost and the clock). Owned by finance and asset management (MRI, Yardi, CoStar, and increasingly VTS). Knows the rent, the expiry, and the break clauses — the timing that makes a utilization insight actionable (you can only give back space at a break or a renewal) — and is the silo most disconnected from day-to-day operations.

The table below makes the structural point: the silos differ not just in content but in why they can't share — protocol, privacy, reliability, schema, and timing respectively. A point tool that masters one of these does nothing for the other four.

Silo	Owner	Dominant systems	Why it can't share
BMS / BAS	Engineering	Desigo CC, Forge, Metasys	BACnet/device-specific schema, configured per building
Badge / access	Security / IT	HID, Lenel, Genetec	Personal/biometric — legal wall
Booking	Workplace / IT	Condeco, Robin, calendars	Intent ≠ attendance (unreliable alone)
HR / headcount	HR	Workday, SAP	No spatial concept (who, not where)
Lease / financial	Finance / AM	MRI, Yardi, CoStar, VTS	Disconnected from operations; timing-bound

2.3 The point-solution market made this worse, not better

This is not an immaturity that scale will cure; it is the market structure. The PropTech landscape is a field of well-funded single-stream specialists, each an excellent business built around owning one slice of the data — which is exactly why they do not, and structurally cannot, converge.

Category	Representative vendors	Scale / funding	Stream owned
Occupancy sensing	VergeSense, Density, Disruptive Technologies	≈$82.6M / ≈$92M+ raised (Crunchbase/PitchBook)	Sensor
Workplace / IWMS (enterprise)	Tango, Planon, Spacewell (Nemetschek)	Enterprise / institutional	Space
Workplace / IWMS (mid-market)	OfficeSpace, Saltmine	Mid-market	Space
Lease & asset intelligence	CoStar, VTS	Public / category-leading	Lease
Booking	Condeco, Robin, calendar incumbents	Workplace	Booking
BMS / controls	Siemens, Honeywell, Johnson Controls	Incumbent triumvirate	Building systems

The tiering matters: an occupier with 80,000 sq ft will correctly conclude the enterprise IWMS platforms are not sized for it and exit before reaching an integrated answer. And note VTS — the dominant commercial leasing/asset platform, now moving aggressively into AI and occupancy analytics — a player most "occupancy tech" framings omit, but one that controls the lease silo.

Each category, however capable, hits the same structural ceiling — and the ceiling is instructive because it is different for each, which is why no single acquisition or partnership closes the gap:

• Occupancy sensing can tell you, precisely, how full a floor is right now. It cannot tell you whether the emptiness is a Friday artifact, a comfort problem, or a lease that expires next year — because the calendar, the BMS, and the lease platform are not its data. The richer its AI, the more confidently it answers a question that is only one-fifth of the decision.
• IWMS / space management owns the floor plan and the move process, but it ingests occupancy as a periodic import, not a live signal, and it has no native read on building-systems comfort. It plans space against a snapshot, in a world that now moves by the day.
• BMS / controls holds the comfort and energy truth and is closest to real-time, but it has no concept of a lease, a headcount plan, or a booking — its world ends at the building envelope.
• Lease & asset platforms know the timing that makes any insight actionable, but sit furthest from operations; they learn a floor was underused at renewal, when it is too late to have acted.
• Booking captures intent, which systematically overstates demand until it is reconciled against attendance it does not hold.

The pattern is exact: each tool is sovereign over one silo and blind to the other four. An AI bolted onto any one of them inherits that blindness. This is why the market's natural response — "we'll add an analytics layer" — does not converge: a smarter answer to a one-fifth question is still a one-fifth answer. Convergence requires an actor whose job is the join, not the slice.

As a widely-cited industry estimate puts it, the average workplace now runs roughly seventeen disconnected software tools (attributed to Gartner/Okta in trade press; an industry estimate, not a primary measurement, and directional). Over the same period that tooling multiplied, employee satisfaction with workplace tools fell from 40% to 29% (Leesman/Gartner, trade press). More tools, less satisfaction — the signature of fragmentation.

The independent CRE commentator Antony Slumbers framed the economics sharply in February 2026: the market is splitting between AI-first platforms that remove whole job categories and point solutions that merely automate a task within existing headcount — for which, he argues, "you paid twice. Once for the software. Once for the salary of the person operating it." (Industry-commentator perspective; see antonyslumbers.com.)

2.4 Pilot purgatory: the evidence, stated separately

The result is a well-documented failure-to-scale. The figures below come from different methodologies and populations and should not be conflated — they are presented individually, with scope.

• JLL (CRE-specific): 92% of organizations have piloted AI tools; 5% report achieving stated objectives. The most directly relevant figure because it is CRE-domain and from a primary industry source. (Worth noting JLL both runs this survey and sells services to close the gap; read it as a directional indictment, not neutral physics.)
• MIT (Sloan/Digital, 2025): ~95% of enterprise GenAI pilots never scale to production.
• RAND (2024): 80.3% of enterprise AI projects produce no measurable value.
• Gartner (2024–2025): ~88% of AI implementations never reach production (its frame mixes research projects with products).

Definitions of "pilot," "failure," and "full deployment" vary by source; the figures are directional, not precisely comparable. But they rhyme: buying AI is easy; operationalizing it is where the industry stalls.

2.5 The 90-day adoption cliff

The corpus labels the recurring pattern the "90-day adoption cliff": a firm purchases licenses, announces a deployment, and watches adoption stall within roughly three months. The mechanism is predictable. A point tool lands in one department, produces a dashboard that department already half-knew, requires manual effort to connect to anything else, and — because it cannot drive a cross-system decision — quietly becomes another tab nobody opens. The license renews on autopilot; the value never arrives. Purchasing is not operationalizing, and the cliff is the single clearest symptom that the problem is integration, not capability.

2.6 Even where the sensors work, action doesn't follow

The fragmentation problem has a twin that deserves to be read as a headline finding, not a footnote. US DOE / Lawrence Berkeley National Laboratory field data across 60,000+ building units found ~40% of air-handling units and ~30% of terminal units carry a fault on any given day, while median realized savings from fault-detection tools is only ~9% against the 15–30% of HVAC energy typically recoverable. (The LBNL data establishes fault prevalence and realized FDD savings; the persistence of uncorrected faults is corroborated by commissioning/retro-commissioning field practice, where cross-building data on closure rates remains sparse — itself evidence of the workflow gap.) The diagnosis: the binding constraint has shifted from detection to fault-closure persistence.

The cost of that gap is concrete. On a 100,000 sq ft office at $3/sq ft total energy with HVAC at ~50% of load, HVAC energy is about $150,000/year; the ~9% that FDD realizes in the field is roughly $13,500/year, against the $22,500–45,000 (15–30%) available if faults were actually closed. The tools detect the problem accurately. The organization does not close the loop. The occupancy story is the same shape: the sensors are good, the data exists, and what's missing is the layer that integrates it, reasons across it with provenance, and drives it to a closed decision.

3. The AISB Agent-Platform Analysis

3.1 The category is named; the path is the question

The most important shift of 2026 is that the field's own authorities have named the architectural answer. PwC and ULI's Emerging Trends in Real Estate 2026 defines what it calls a property operating system (propOS): a coordinated layer of specialized agents, digital twins, and real-time data integration that sits above the legacy platforms, draws their data through APIs, and continuously optimizes operations — with the shift "from batch processing to real-time streaming" cited as what "unlocks new insights."

That describes an agent platform. To be precise about the logic: PwC and ULI named a category and a trajectory — they did not certify a vendor, set a standard, or endorse any product, AISB included. The valid inference is narrow and we hold to it: propOS is an agent architecture; AISB is an agent architecture; therefore AISB sits in the category the authorities have validated. The strategic consequence is what matters — the question for an owner-operator is no longer whether to move toward propOS, but how to reach it without building it tool-by-tool, which Section 2 argued the point-solution market cannot do for you.

This is where a purpose-built CRE agent platform changes the economics. AISB is not a sixth sensor or an eighteenth dashboard. It is a fleet of domain-specialized agent squads, pre-integrated across the five silos and sharing one audited knowledge layer. The squads operate in parallel and in feedback loops, not as a linear handoff chain: a fault the Technical Services squad finds recalibrates the Space Planning squad's baseline; a Privacy Broker constraint shapes the data granularity every other squad may use.

3.2 The squad workflow, traced through one real decision

To make this concrete, follow a single decision — a 100,000 sq ft occupier deciding whether to give back its worst-utilized floor at a renewal fourteen months out — through the platform.

Space Planning — the demand truth. The Hybrid-Work Calibrator translates the client's actual attendance policy (say, a 3-day mandate) into a peak-simultaneous-headcount demand curve. On our worked example it finds the floor in question runs 78% on Tuesday but 34% Monday/Friday — so the "45% average" that a single sensor feed reported is the wrong number to plan against. Its headline output, Policy-Space Mismatch Detection, flags that the lease is sized to the average, not the peak. The Utilization Analyst fuses sensor + badge + reservation data — three of the five silos — under Bayesian priors, and explicitly detects sensor decay, "hot-desk shadow," and staleness, refusing to act on data older than a set threshold (a 90-day minimum on clean data is the squad's hard gate). It is the direct counter to the data-quality failure mode: the platform knows when its own inputs have rotted.

Technical Services — the building truth and the closed loop. The BMS/IoT agent brings the building-systems and FDD layer into the same picture and assesses digital-twin readiness; on our example it surfaces that two of the floor's AHUs carry uncorrected faults — meaning part of the "low occupancy" is comfort-driven avoidance, not genuine demand absence. The Facility Operations agent owns work-order triage and fault closure — the precise gap LBNL identified — so the faults become tickets that close, not alerts that age.

Soft Services — the honesty check. The Tenant Experience agent runs KPI-Theater Detection, watching for divergence between a vendor's reported KPIs and occupants' actual experience (NPS). When the dashboard says "optimized" and the people on the floor disagree, the platform surfaces the contradiction instead of burying it.

Privacy Broker — the reason this is even legal. Fusing badge and sensor data to understand occupancy is exactly the move enterprise legal teams block, because it touches personal and potentially biometric data. AISB's Privacy Broker is designed to apply differential privacy (per-zone, per-day epsilon budgets), a k-anonymity floor, and regional consent enforcement spanning GDPR, Singapore PDPA, CCPA, BIPA, and the EU AI Act — the capability that lets a legal team approve a badge-fused occupancy product, and the piece single-stream point solutions by construction do not offer. (These capabilities describe AISB's design architecture; specific implementations are validated per deployment and per jurisdiction, not represented as a uniform contractual commitment.)

Knowledge Engine — provenance against the trust deficit. Every output carries its lineage. The Claim Classifier admits a fact only if it passes a five-signal test — source authority, standards anchor, numeric specificity, cross-source corroboration, and a contradiction check — and each squad writes to an audited brain. A recommendation therefore arrives with which squad produced it, from which source, against which standard (IPMVP, ASHRAE), at what confidence. In candor: this report is the inaugural systematic deployment of that gate at publication scale, and documented instances of the classifier catching and correcting AISB's own output will be published in subsequent issues as the evidence base accumulates. We would rather state that honestly than imply a fully-proven record we have not yet published.

On the worked decision, the platform's integrated answer is not "give back the floor." It is: close the two AHU faults first (comfort is suppressing real demand), re-measure for one quarter, and if Tuesday peak still leaves 30% headroom, re-stack two teams onto it and give back a different, genuinely idle floor at renewal — a conclusion no single silo could have reached, and one that carries its provenance with it.

3.3 A second trace: the APAC over-leasing decision

The same machinery, a different lever. Consider a Singapore occupier with 500,000 sq ft approaching a renewal in a 4.1%-vacancy market where give-back is not realistic. Here the Space Planning Hybrid-Work Calibrator, reading badge + booking against the firm's actual (high, ~72%) attendance, finds genuine peak demand maps to roughly 420,000 sq ft — meaning a like-for-like renewal would lock in 80,000 sq ft the firm never peaks into. The Lease/financial read (the fifth silo, often owned by a different team in a different country) supplies the break and renewal timing; the Privacy Broker clears the badge fusion under Singapore PDPA; the Knowledge Engine attaches provenance so the CFO can audit the recommendation before committing nine figures of lease. The platform's output is not "give back" (impossible here) but "renew at 420,000, not 500,000" — the avoided-over-leasing lever Section 1.3 described, worth materially more than a give-back precisely because supply is tight. No occupancy point tool reaches this conclusion, because the deciding data lives in the lease silo it never touches.

3.4 Why the platform shape matters more than any single agent

No single one of these agents is unique — sensor vendors do occupancy analytics, IWMS players do space management, privacy vendors do differential privacy. What is hard to replicate is that they hand off to each other across the very silos that fragment the point-solution market, over one shared, provenance-tracked knowledge layer. The utilization decision that "never assembles" in Section 2 assembles here, because the demand truth, the building truth, the honesty check, the privacy clearance, and the provenance live in one coordinated system. The exhibit below maps which squad reads which silo — the coverage that is the whole point.

Silo	Read by AISB squad/agent	Turned into
BMS / BAS	Technical Services — BMS/IoT + Facility Ops	Fault closure, comfort truth
Badge / access	Space Planning — Utilization Analyst (via Privacy Broker)	Verified attendance
Booking	Space Planning — Utilization Analyst	Intent, reconciled vs. attendance
HR / headcount	Space Planning — Hybrid-Work Calibrator	Forward demand curve
Lease / financial	(decision context) + Knowledge Engine	Actionable timing + provenance

That is the propOS PwC and ULI described — built for CRE, pre-integrated, and domain-trained, rather than assembled by an owner from seventeen vendors who will not connect.

3.5 How it connects, how you buy it, and what we can prove

Three questions a serious buyer will ask immediately, answered plainly.

How does it connect to the building? Through the legacy platforms' own interfaces — BACnet/IP and MQTT where the BMS exposes them, vendor APIs (Desigo CC, Forge, Metasys) where it does not, and a thin middleware/normalization layer that maps each building's idiosyncratic point names into a common model. AISB is designed to sit above the incumbents and read them, not rip and replace them — which is also what makes the propOS framing more than rhetoric.

The harder half of "connect" is not the wire but the meaning. Two buildings can both speak BACnet and still disagree on what a point is called, what unit it reports, and what "occupied" means — so raw connectivity yields five streams of mutually unintelligible data. The normalization layer therefore does three things before any agent reasons over the data. First, it maps each building's point names to a shared schema (an emerging open vocabulary such as Brick or Haystack-style tagging is the direction of travel), so "AHU-3 supply temp" and "AirHandler_03_SAT" become the same concept. Second, it reconciles the streams against each other — booking intent against badge attendance, sensor heat against headcount — and records which source it trusted and why, so a downstream recommendation can be traced to its evidence. Third, it scores readiness: not every building is twin-ready, and the platform is explicit about which tier a given asset is in — from "metered and API-exposed" (full real-time optimization) down to "legacy controller, manual export only" (periodic batch analysis) — so an owner knows what they will get before they pilot, rather than discovering the data was too thin after.

Security and tenancy isolation are part of the same layer, not an afterthought: building-systems networks are sensitive operational technology, so AISB is designed to read through brokered, least-privilege interfaces and to keep each client's data isolated — the integration is one-way intelligence, not a new control path into the building.

How do you engage it? As a platform-plus-squad service: the agent squads run against the client's integrated data, with a human override at every decision node and the owner's team in the loop on every action. This is an operating partnership, not a self-serve dashboard licence — deliberately, because the failure mode in Section 2 is precisely "bought the licence, never operationalized."

What can we prove today — and what can't we? Honestly: there is, as yet, no published office-utilization ROI figure for agent platforms that we would stand behind. The most-cited agentic-platform results in circulation come from a single, unnamed, residential deployment — an asset class whose economics (unit turnover, maintenance dispatch) do not map onto office space utilization — so we deliberately do not anchor on them, however tempting the headline numbers are. What we can show you instead is the mechanism (Sections 3.2–3.4) and an ROI model built entirely from independent, attributable market inputs (Section 6). AISB's own office-deployment outcomes will be published, with provenance, as they accumulate. We would rather show you the architecture and the math than borrow a number that does not transfer.

4. Recommendations

Five recommendations follow from the analysis. For most owner-operators we suggest this sequence: instrument the demand baseline first (R3), demand provenance and clear privacy in parallel (R2, R5), then act decision-by-decision (R4), and only then rationalize the tool stack (R1) once the integrated layer makes redundant tools visible.

• R3 — Instrument a single source of truth before layering AI. Integrating BMS, badge, booking, HR, and lease data into one provenance-tracked layer is the prerequisite. AI on top of silos just produces faster partial answers. This is step one, not step three.
• R2 — Demand provenance on every AI output. A recommendation you cannot audit is one you cannot act on. Require that every insight arrive with its source, the standard it rests on, and a confidence level. Treat "trust the model" as a disqualifier.
• R5 — Clear the privacy path early. Badge-and-sensor fusion is the highest-value occupancy move and the one legal will block. Solve consent, differential privacy, and k-anonymity at the start — it is the gate to the whole opportunity, not a compliance afterthought.
• R4 — Measure against a decision, not a dashboard. Tie every deployment to a specific decision it must improve (give-back, re-stack, policy change) and verify the outcome with measurement & verification (IPMVP where energy is involved). A dashboard nobody acts on is shelfware by another name.
• R1 — Then stop buying the eighteenth tool. Once the integrated layer exists, redundant point solutions become visible. Before any new purchase, ask whether it integrates the five streams or adds a sixth. If it owns one stream, it is part of the problem you already have.

5. Implementation Playbook

A phased path from fragmented stack toward propOS, designed so each phase pays for the next — with an explicit failure path, because a playbook that only describes success is marketing.

Phase	What you do	Who owns it	Output / gate
0 — Baseline	Quantify the gap: leased vs. used by floor, peak-day vs. mandated-day demand, $/sq ft idle	Asset manager + FM	A defensible utilization baseline and a dollar figure
1 — Map the silos	Inventory the five data streams, owners, and gaps; assess BMS API posture (BACnet/MQTT/vendor) and digital-twin readiness	IT + engineering	Source inventory + integration-cost map
2 — Clear privacy	Stand up differential-privacy / k-anonymity / consent for badge-sensor fusion in the operating jurisdiction	Legal + DPO	Legal sign-off to fuse occupancy data
3 — Integrate + pilot ONE decision	Deploy the squads against one high-value decision (e.g., a give-back on the worst-utilized floor) with human override at every node	AISB + owner	A verified decision with an M&V'd outcome
4 — Close the loop	Wire fault-closure and policy feedback so detection drives action, not just alerts	FM + ops	Fault-closure persistence, not just detection
5 — Scale across the portfolio	Extend the integrated layer building-by-building; each verified decision lowers the cost of the next	Owner	A compounding operating loop

The failure path (read this before Phase 3). If effective-utilization improvement versus baseline is below 5% at 90 days, the protocol is not "push harder" — it is a structured reassessment: (1) a sensor-coverage and data-quality audit, (2) a booking-system data review, (3) a change-management check with the facilities team. Persisting below 5% at ~140 days triggers a pilot-suspension with full data export and baseline-reversion documentation. Knowing the exit in advance is what separates an operating partnership from shelfware.

The sequencing is the whole point: baseline → map → clear privacy → act on one decision → verify → scale. Most failed PropTech deployments skip to a tool purchase before Phase 0 and never reach Phase 3.

6. ROI & the metrics that matter

The honest framing: the headline industry case study is real but single-sourced and residential (Section 3.5), so the model below is built from independently-sourced Tier-A inputs, shown as arithmetic, and presented as illustrative ranges, not assured outcomes. Actual results depend on lease structure, jurisdiction, building systems, and — above all — whether the organization acts on what the platform surfaces.

Scenario A — 100,000 sq ft occupier, NYC prime.

• Space efficiency. A 10% effective-utilization gain: 100,000 sq ft × $70–100/sq ft × 10% = $700K–1.0M/year in rent efficiency (realizable portion depends on ability to give back or sublet).
• Energy. Published studies report HVAC-energy reductions in the 20–30% range from AI/RL optimization, and HVAC is 40–60% of commercial energy. Applied to the HVAC portion of a $3–5/sq ft total-energy bill: 100,000 × $3–5 × (40–60%) × (20–30%) = $24K–90K/year (independent calculation; note this corrects a common overstatement that applies the savings rate to total energy and ignores the HVAC fraction).
• Fault closure. Separately, closing the faults FDD already detects moves realized HVAC savings from ~9% toward the 15–30% achievable — on this footprint, roughly $13.5K today versus $22.5–45K available (Section 2.6).
• Demand flexibility. Grid-interactive controls can curtail up to 30% of commercial peak load (DOE/LBNL); programs such as the Edo–National Grid demonstration in New York pay buildings for that flexibility, turning HVAC from a pure cost toward a potential revenue line (value is tariff-dependent and not modelled here).

Scenario B — 500,000 sq ft occupier, Singapore prime (the APAC "avoided over-leasing" case). In a 4.1%-vacancy market, give-back is hard, so the value is in not over-committing. If integrated peak-demand data shows a firm renewing 500,000 sq ft genuinely needs ~420,000 at peak, avoiding the 80,000 sq ft of unnecessary renewal at an illustrative ~$90/sq ft (an explicit assumption, not a sourced rate) = ~$7.2M/year of avoided commitment — a larger prize than give-back precisely because supply is tight. (Energy savings stack on top and depend on Singapore Green Mark / tariff context.)

Sensitivity. Because the space-efficiency line dominates, the model is most sensitive to two inputs — rent and the realizable utilization gain:

Realizable util. gain	@ $40/sq ft (suburban)	@ $80/sq ft (CBD)	@ $100/sq ft (prime)
5%	$200K	$400K	$500K
10%	$400K	$800K	$1.0M
15%	$600K	$1.2M	$1.5M

(Per 100,000 sq ft, rent × gain; illustrative.) The lesson for a buyer: the energy and fault-closure lines are real but secondary; the decision economics live in the space line, which is why R4 (measure against a give-back/re-stack decision) is the recommendation that matters most.

The metrics to track — and to hold any platform, AISB included, accountable to:

1. Effective utilization (peak-day, not average) versus leased area.
2. Share of decisions carrying provenance — the auditability rate (target: 100% of acted-on recommendations).
3. Verified savings (M&V) — IPMVP-anchored, not modelled.
4. Data-coverage % across the five silos (a leading indicator; <100% means partial-picture risk).
5. Fault-closure persistence — not just detection rate.
6. Tool-count reduction — fewer silos is the clearest leading indicator of integration.

The strategic ROI is not any single number; it is decision-cycle speed. As one practitioner put it in 2026, "intelligence orchestration is becoming just as critical as market knowledge." Every operator at scale has sensors. The edge belongs to whoever integrates and acts first.

7. Risks, caveats & honest limits

This report is AI-assisted research under human editorial review, reviewed by AISB's CRE squads, C-level + Expert Council, and a CLO legal-risk gate before publication. It is not professional, engineering, legal, or investment advice. Specific honest limits:

• Source variance is real. Vacancy and utilization figures differ across CBRE, JLL, and Moody's by methodology; we present providers side by side rather than picking one as authoritative. Utilization figures are sensor/badge-sampled with differing definitions of "occupied."
• We deliberately excluded a widely-circulated agentic-platform ROI case study because it is single-sourced, unnamed, and residential — its figures do not transfer to office utilization, and we would rather omit a headline number than anchor on one that does not apply. The ROI in Section 6 rests only on independent, attributable inputs.
• Several circulating figures were deliberately excluded as unverifiable against a primary source (notably a widely-quoted "$430–550B unlockable value" estimate). We would rather omit a number than publish one we cannot stand behind.
• Some figures are industry estimates or analyst forecasts (the "~17 tools," market-size CAGR projections) and are labelled as such, with wide error bars.
• AISB capabilities are described as design architecture. Specific implementations — the Privacy Broker's privacy protections in particular — are validated per deployment and per jurisdiction, not represented as a uniform contractual commitment.
• An agent platform is not magic. Its value is bounded by the organization's willingness to act on what it surfaces; the "90-day adoption cliff" is an organizational failure as much as a technical one, and no platform removes the need for human decision-makers and override.

An owner should validate every figure here against their own portfolio data before acting.

Appendix A — sources, methodology, glossary & disclosure

Methodology. The pain-point selection was derived from AISB's CRE-intelligence corpus by ranking themes on weighted signal frequency × recency, with PropTech/AI tool-failure themes up-weighted; "space utilization" emerged as the dominant business pain and "data fragmentation / interoperability" as the dominant tool-failure theme. External figures were then researched and tiered by source authority (A = primary/official, B = reputable secondary, C = trade/commentary), with each figure attributed inline and unverifiable figures excluded.

Glossary.

• Utilization — share of leased/available space actually occupied (sensor/badge/booking measured); distinct from vacancy (whether space is leased at all).
• propOS (property operating system) — PwC/ULI's term for a coordinated layer of AI agents, digital twins, and data integration above legacy platforms.
• IWMS — Integrated Workplace Management System (space/lease/asset management software).
• BMS / BAS — Building Management / Automation System (HVAC, lighting, metering controls).
• FDD — Fault Detection and Diagnostics (automated identification of equipment faults).
• M&V / IPMVP — Measurement & Verification; the International Performance Measurement and Verification Protocol for evidencing savings.
• Differential privacy / k-anonymity — techniques that allow population-level analytics on personal data while protecting individuals.

Selected sources (representative; full attribution in-line): - JLL — Global Occupancy Planning Benchmark 2026; US Office Outlook Q3 2025; APAC commentary 2025; AI-in-CRE research 2025. - CBRE — US Office Figures Q1 2026; NYC rent data. - Moody's — CRE Analytics Q2 2025. - Cushman & Wakefield — Singapore Office MarketBeat Q1 2026; Knight Frank/JLL Singapore pipeline commentary 2026. - PwC & ULI — Emerging Trends in Real Estate 2026 (propOS). - US DOE / Lawrence Berkeley National Laboratory — FDD fault-prevalence field study (60,000+ units). - Occuspace (2024); Density (Q1 2025) — independent utilization data. - MIT Sloan/Digital (2025); RAND (2024); Gartner (2024–25) — enterprise AI pilot-failure research. - Crunchbase / PitchBook — vendor funding figures. - Antony Slumbers (antonyslumbers.com, Feb 2026) — industry commentary. - AISB CRE-intelligence corpus (May–June 2026) — aggregation and signal context.

AI disclosure. This report was researched and drafted with AI assistance, grounded in AISB's primary CRE corpus and the named third-party sources above, under human editorial review and domain supervision; all quantitative claims were checked against cited sources. Before publication it passes AISB CRE-squad review, then C-level + Expert Council review, then a CLO legal-risk gate. AISB maintains editorial responsibility for all analysis and recommendations. Named third-party companies and their products are described factually from public information; they neither endorse, nor are endorsed by, AISB. PwC and ULI named the propOS category and trajectory; nothing here should be read as their endorsement of AISB.

Appendix B — The buyer's checklist: 12 questions before you sign

The fastest way to avoid joining the 92% whose pilots stall is to interrogate the architecture before the demo dazzles you. These twelve questions separate an integration layer from a sixth silo. Score each yes/no; a tool that cannot answer the first four is a point solution regardless of how its AI demos.

On integration (the make-or-break four):

1. Which of the five silos does this read natively — BMS, badge, booking, HR, lease — and which does it leave to us to wire? (If the honest answer is "one," it is a point solution.)
2. Does it reconcile booking intent against badge/sensor attendance, or report them separately? (Separate = it will overstate demand.)
3. What is your BMS integration path — BACnet/IP, MQTT, vendor API, or a middleware layer — and who owns the per-building point-name mapping?
4. Does a recommendation carry its provenance (source, standard, confidence), or do we have to "trust the model"?

On privacy (the gate to the highest-value data):

5. How do you make badge-fused occupancy data legally usable — differential privacy, k-anonymity, consent — in our specific jurisdiction?
6. Will our DPO and security team sign off on the data flow, in writing, before pilot?

On action, not dashboards:

7. Which single decision will this improve in the first 90 days, and how will we measure it (M&V / IPMVP)?
8. Does it close the loop — turn a detected fault or a utilization finding into a tracked action — or only display it?
9. What is the failure path if the pilot underperforms — the reassessment trigger, the exit, the data export?

On economics and lock-in:

10. Does deploying this let us retire any existing tool, or does it add a line to the stack?
11. What is the total first-year cost — licence plus the internal headcount to operate it? (Beware paying twice.)
12. Who owns the integrated data if we leave — is it portable, or captive to your platform?

A platform built for the integration gap answers 1–4 with specifics, treats 5–6 as a first-class capability rather than a compliance afterthought, ties 7–9 to a measurable decision, and is unafraid of 10–12. A point solution deflects the first four and sells you on the AI in the demo. The checklist is the cheapest diligence you will ever run — and the one most likely to keep your next purchase off the 90-day cliff.