AI for building operations falls into three levels: conversational AI that answers questions from documents, analytical AI that detects anomalies and recommends actions, and operational AI (also called agentic AI) that executes interventions — changing setpoints, creating work orders, and verifying results — under defined permissions. Most AI for building operations deployed in Nordic real estate today stops at the first level.
This blog post explains what separates the three levels, why most AI stops at the answer, what infrastructure operational AI requires, what operational AI actually does day to day, and how buyers should evaluate vendors that claim to offer it.
Key takeaways:
– Only 5% of Nordic organisations report high expertise in agentic AI, and 49% expect agentic AI transformation to be more than three years away (Deloitte, State of AI in the Nordics 2026).
– The barrier to operational AI is fragmented building data and the lack of write-back access to operational systems — the AI model is rarely the constraint.
– An open semantic data model — RealEstateCore — solves the fragmentation problem and was developed in the Nordic market.
– Eufemia, a KLP property in Oslo, reduced energy use by 31% — equivalent to 292,000 kWh per year — with ProptechOS energy management.
What are the three levels of AI in building operations?
AI for building operations operates at three levels of capability: conversational (answers questions), analytical (recommends actions), and operational (executes and verifies actions). The difference between levels is the infrastructure underneath the language model, not the quality of the model itself.
- Level 1 — Conversational: summarises manuals, logs, and leases; answers natural language questions. Requires documents only, no integration with operational systems. Where most Nordic deployments are today.
- Level 2 — Analytical: detects anomalies, recommends actions, surfaces insights. Requires read access to sensor and system data. Some deployments.
- Level 3 — Operational: diagnoses, intervenes within permissions, creates work orders, verifies results, logs every action. Requires a structured semantic data model, write-back integrations, governance and audit trails. Few deployments.
Level 1 is genuinely useful — and it is also the easiest to build, because it requires no integration with operational systems. Level 2 is better than a static dashboard, but the response still depends on a human deciding what to do. At level 3 — operational AI — the loop closes without a specialist orchestrating every step. The rest of this blog post is about that third level: what it requires, what it does, and how to recognise it.
Tietoevry’s Nordic AI Survey 2026 found that organisation-wide AI use in production grew from 7% to 31% in a single year — but only 4% of organisations describe AI as a critical part of their core infrastructure and business operations. The activity is real. The integration into operations is not yet.
How big is the Nordic AI gap in real estate?
Nordic real estate is roughly two years behind the US conversation on agentic AI: at Realcomm 2026, the debate had already moved to governing AI agents and measuring their outcomes at scale, while most Nordic organisations have not yet reached the starting line of that conversation.
The data supports this. Deloitte’s State of AI in the Nordics 2026 — a survey of 170 senior executives across Denmark, Finland, Norway, and Sweden — found:
– Only 5% of Nordic organisations report high expertise in agentic AI.
– 49% expect agentic AI transformation to be more than three years away.
– Strategic preparedness dropped from 61% last year to 43% today.
AI investment is happening in Nordic real estate, but it is largely disconnected from building operations. Property companies have assistants that can summarise maintenance logs or answer questions about a lease. What they do not yet have is operational AI that can act on a building. That gap is not permanent — but it is widening.
Why does most AI for buildings stop at answering questions?
Most AI for buildings cannot act because building data is fragmented: systems from different vendors use incompatible protocols, naming conventions, and data structures, so the AI cannot reliably identify the assets it would need to control. The barrier is the infrastructure, not the AI model.
The fragmentation is particularly acute in Nordic portfolios, where systems have been procured independently over decades. A chiller in one system may be identified by a floor number. In another, by a serial code. In a third, it may not be surfaced at all. An AI that cannot reliably identify what it is looking at cannot safely change operational states — it can still answer questions from documents, but that is where it stops.
Three separate problems keep AI at level 1:
- Data fragmentation
Without consistent, structured knowledge of the building, its assets, its spaces, and the relationships between them, the agent works from incomplete and contradictory information. - Execution access
Reading sensor data is different from writing to a building management system. Identifying that a zone is overcooled is different from adjusting the setpoint. Solving the first does not solve the second. - Governance
Autonomous action without traceability, permissions, and audit trails is not something most property owners will deploy at scale. The hesitation is good judgment, not technophobia — and elsewhere that governance problem is being solved now.
JLL’s 2025 Global Real Estate Technology Survey — covering more than 1,500 senior CRE decision-makers across 16 markets — found that 88% of investors and 92% of occupiers are running AI pilots, but only 5% have achieved all their AI programme goals, and more than 60% of investors remain strategically, organisationally, and technically unprepared for scaled AI implementation. JLL also warns that existing technology maturity gaps widen with AI rather than close. Organisations that defer the infrastructure work do not catch up faster later. They fall further behind.
Can AI agents run on an existing BMS?
Yes — operational AI can act on existing building management systems when a semantic data model standardises asset, space, and system data across vendors, and verified integrations provide write-back access under defined permissions. No BMS replacement is required.
ProptechOS addresses the fragmentation problem through RealEstateCore, an open semantic data model for real estate developed in the Nordic market and now running on Microsoft Azure globally. Rather than a custom schema, RealEstateCore provides a shared vocabulary for buildings, systems, spaces, assets, sensors, and the relationships between them — consistent across vendors and properties.
This structured model is what allows an AI agent to reason. When the platform knows that a specific air handling unit serves floors four through seven, is connected to a particular BMS, and has had three filter replacements in the past eighteen months, an agent diagnosing a temperature anomaly on floor six has context. Without that structured understanding, the agent is pattern-matching against noise.
Write-back capability through verified integrations is what turns analysis into operational AI. When the agent determines a setpoint adjustment is appropriate and within its defined permissions, it executes the adjustment. The action is logged, attributed to the agent, linked to the triggering condition, and the resulting sensor data is monitored to confirm the intended response.
This is a structured semantic layer that gives AI agents the specific, relational knowledge they need to act safely and verifiably across a portfolio — and it is a standard Nordic real estate helped create. The question is whether Nordic real estate will be the last to adopt it.
How do expert agents change the work of building operations?
Expert agents encode the diagnostic playbooks of experienced specialists and execute them continuously across hundreds of assets, shifting the engineer’s role from watching dashboards to managing exceptions.
The clearest way to understand one is as a new digital colleague. It takes on the repetitive diagnostic work that fills a specialist’s day — the constant scanning, cross-referencing, and first-line triage of recurring faults — and frees that person to spend their time on higher-value work, where human judgment actually changes the outcome. The digital colleague applies the same expert reasoning to every asset, every minute, without fatigue or context-switching. The specialist stays in charge: setting the policies the colleague works within, handling the exceptions it escalates, and bringing judgment to the problems that genuinely need it.
The dashboard era gave property companies more data and more visualisations. It did not significantly reduce dependence on experienced specialists to interpret the data and decide what to do. Most AI in property today is generative — useful for administrative tasks and research, but still reliant on a person to turn its output into an operational decision. That is where the majority of Nordic property companies are today.
Maintenance teams, energy managers, and building engineers remain the bottleneck: there are not enough of them, their expertise is difficult to transfer, and their attention is consumed by repetitive diagnostic work that follows recognisable patterns. With a digital colleague in place:
- The specialist’s expertise scales across the portfolio instead of one building at a time.
- Repetitive diagnostic analysis is automated and runs continuously, without fatigue or context-switching.
- The engineer’s attention shifts to exceptions, edge-case supervision, and problems that genuinely require human judgment.
In a Nordic market facing both a shortage of technical specialists and increasing pressure on operational costs, that leverage matters.
What can an operational agent actually do?
Operational AI is easiest to understand through the tasks it takes off a technical property manager’s plate. Each one maps directly to the work that fills a shift.
For the Technical Property Manager, an Operational Agent can:
- Monitor consumption and climate anomalies. Each morning, report deviations in district heating, cooling, electricity, water, indoor climate, and presence against a calculated overnight baseline — surfacing unnecessary consumption, potential water leaks, and comfort drift before anyone files a complaint.
- Watch district heating delta-T. Track the temperature difference between supply and return against outdoor temperature and an expected average, and explain the cause behind each deviation — reducing penalty charges, avoiding over-purchasing energy, lowering pump wear, and keeping indoor climate even.
- Triage alarms. Instead of a flat, unmanageable alarm feed, analyse how often alarms recur, identify probable causes, and escalate only what genuinely needs attention.
- Catch unnecessary ventilation. Flag abnormal operating times — fans running when no one is present, or at the wrong time of day.
- Analyse fault tickets. Produce a weekly summary of fault tickets per property and tenant, so the team can meet complaints proactively.
- Unify access. Bring all the information and systems needed to operate the portfolio into one place, behind one login.
For the Operations Technician, an Operational Agent can:
- Surface lift and escalator deviations. Report predefined operating deviations in lifts and escalators that a technician can resolve directly, saving a callout.
The pattern across all of these is a deliberate progression in authority. The agent begins by notifying — surfacing the deviation and its likely cause. It moves to proposing — recommending the specific adjustment. And, where permissions allow, it advances to acting — making bounded setpoint changes in consultation with the operations technician, with every action logged and attributable. Operational AI is a graded capability, with authority that expands only as trust is established.
What does a closed-loop operational AI intervention look like?
A closed-loop intervention means the agent detects a deviation, diagnoses the cause, acts within its authorised range, verifies the result against live sensor data, and logs the full chain — escalating to a human with diagnostic context only if the action does not produce the expected outcome.
A concrete sequence:
- 06:42 — The agent detects a deviation in supply air temperature on floor three.
It cross-references the asset model and identifies the most probable cause from the pattern of sensor readings.
It determines the applicable response is a setpoint adjustment within its authorised range.
- 06:43 — The adjustment is made.
Sensor monitoring confirms the floor returns to target range within eleven minutes.
Had the adjustment failed within a defined window, the agent would have escalated and created a work order with full diagnostic context attached.
The agent owns the routine. The specialist still owns the exception. Compare this with the conventional process at most Nordic property companies: something fails, someone notices, a ticket is created, a specialist investigates, a work order is issued, and someone eventually verifies the outcome — with hours or days between anomaly and resolution.
This is already moving from illustration to deployment. Locum, Region Stockholm’s healthcare property company, has proven the full loop end to end on ProptechOS: sensors in the platform trigger work orders directly in Locum’s facility management system, with detection, action, and feedback closing without manual handoff. For the emergency care building at Danderyds Sjukhus, Locum is preparing a structured organisation of agents — analyst agents monitoring heating, cooling, ventilation, and electricity against defined KPIs, a supervisor agent quality-assuring their output, and an executor agent handling the work orders that need no site visit. Because the environment is a working hospital, control stays human-supervised: the agents analyse, diagnose, and propose, and AI-generated work orders are reviewed and approved before they reach an operator. The loop closes; the judgment stays human. Read the full story in the Locum case study.
Energy results follow the same principle. At Eufemia in Oslo, a KLP property, ProptechOS energy management delivered a 31% energy reduction — equivalent to 292,000 kWh per year. Results of this scale require more than better analysis. They require the ability to act on it at a pace and frequency human teams cannot sustain manually.
How does operational AI change the commercial model?
Traceable, attributed, measurable action makes outcome-based commercial models possible: when every intervention can be connected to a specific agent action and its effect on operational and financial performance, integrators can be compensated for results rather than hours.
Some system integrators are already exploring models that move from selling consulting hours toward acting as strategic advisers compensated for measurable operational outcomes, such as improved net operating income or sustainability performance. This model is already emerging in more advanced markets, where demonstrating and attributing operational improvement is becoming a baseline expectation. In the Nordic market, most integrators are still compensated for hours and deliverables. A conversational AI that answers questions in a separate interface, disconnected from the building’s operational systems, cannot support an outcome model — the cause-and-effect chain is not traceable. Operational AI can, because every action is attributable.
How should buyers evaluate operational AI?
The test for operational AI is what the agent can do after producing an answer — not whether the vendor uses the word “agent.” Eight questions separate operational platforms from conversational ones:
- Can the agent write back to operational systems, or only read from them?
- Can it create and route a work order automatically?
- Can it adjust a setpoint under defined permissions?
- Can it verify the result of its own intervention against live sensor data?
- Is every action logged and attributable to the agent that took it?
- Does the system understand relationships between assets, spaces, systems, and operational data — or does it reason from documents alone?
- Can it operate consistently across a portfolio with multiple buildings, systems, and vendors, or only within a single application?
- Are permissions, policies, and human oversight built into the execution model?
A growing number of established platform vendors and challengers are introducing agents, autonomous building capabilities, and AI-first propositions. The direction is legitimate and accelerating, and it is reaching the Nordic market now.
JLL’s research is instructive: companies with successful technology programmes are widening the competitive gap over those still running disconnected pilots — and the gap widens faster with AI, not slower. Waiting is not a neutral position. Nordic organisations that start asking these eight questions now will make better investment decisions and close the capability gap faster.
What comes next for AI in real estate?
The next phase of AI in real estate will be defined by which platforms can safely turn understanding into action and prove the outcome — not by which assistant gives the most polished answer. That phase belongs to operational AI.
According to PwC’s 29th Global CEO Survey, only 30% of CEOs globally report revenue increases attributable to AI. The majority have not yet captured measurable value from their AI investments. That is the baseline. The organisations that move beyond pilots and into operational AI in the next two years will define what the industry considers normal by 2028.
The Nordic market has the infrastructure, the expertise, and the open standards to move quickly — including a semantic data model it helped build. What it needs now is the willingness to move from AI that answers to AI that acts.
ProptechOS was built to provide that foundation — the operational AI layer underneath the chatbot, not another interface on top of it.
Book a demo – https://www.proptechos.com/demo
