A maintenance contractor arrives to fix a boiler that was repaired three days ago. A tenant receives a rent reminder on an account that paid yesterday. In property and facility management, the cost of incomplete context is measured in wasted resource, missed SLAs and broken tenant relationships.
Property and facility management doesn’t always feature in conversations about AI transformation. It lacks the headline glamour of fraud detection or the obvious human stakes of healthcare. But as an operational domain, it has a data problem that’s at least as complex – and the cost of getting it wrong, spread across large commercial portfolios or residential estates, can quickly multiply.
Property and facility management systems generate enormous volumes of real-time events – yet most AI systems can only see a fraction of them in context. A large facilities operation might be managing thousands of assets with millions of data points across dozens of sites: maintenance schedules, contractor assignments, lease agreements, SLA obligations, tenant communications, energy consumption, access controls, inspection records.
The information required to coordinate all of this effectively is vast, distributed, and changes constantly. And increasingly, organisations are deploying AI agents to manage it.
So why do so many of those deployments fall short? And what benefits would truly effective AI unlock in this context?
The answer, again comes down to context – specifically, the combination of real-time operational signals and historical depth that most property management data architectures still can’t provide at once. Not being able to see them together triggers a host of avoidable issues.
When the AI only sees part of the building
Facility management AI needs to operate across two dimensions simultaneously. The real-time dimension covers live maintenance tickets, contractor availability and location, energy readings, access events, fault alerts and tenant communications as they happen.
The historical dimension covers lease terms and renewal timelines, SLA agreements and performance history, repair records and recurring fault patterns, tenant satisfaction trends and payment histories.
But neither dimension on its own can produce reliable decisions.
An AI agent working only from real-time signals might dispatch a contractor to a fault that the historical record would show has recurred six times in the past year, a pattern that points to a systemic issue requiring a different response than another reactive repair. But working only from historical data, the same agent might make decisions based on a building state that’s hours or days out of date, missing a fault that developed this morning or a contractor assignment that changed an hour ago.
In a commercial property portfolio, that disconnect produces costly inefficiency: duplicated contractor visits, missed SLA windows, tenant complaints that could have been anticipated. Across a large residential estate, it produces something closer to a breakdown in the landlord-tenant relationship, as automated communications contradict the tenant’s actual experience, and maintenance responses arrive too late or address the wrong problem.
The coordination challenge at scale
Large property operations add another layer of complexity. Coordinating maintenance across multiple sites means reconciling contractor availability, parts supply chains, access schedules and tenant preferences. All of these change in real time and need to be understood against a background of SLA commitments, lease obligations and historical performance data.
An AI agent trying to optimise that coordination faces the same distributed consistency problem we’ve explored throughout this series. A work order management system might update on a different cycle to the contractor scheduling platform. A tenant communication tool might not yet reflect a repair that was completed this morning. Energy management data might lag behind the access control system by enough that a building the AI thinks is occupied is actually empty.
Such issues are the everyday reality of managing complex built environments at scale. They are the reason AI deployments in this sector so often produce decisions that are technically correct based on the data available, but wrong in practice.
How Streambased brings it together
Streambased addresses both problems through its unified layer across Kafka and Iceberg. Maintenance tickets, tenant inquiries, rent payments, contractor assignments, energy readings and building alerts stream into Kafka as they happen. Those same streams persist to Iceberg, building a continuously updated historical record of asset condition, tenant relationships, SLA performance and repair history – all without a separate ETL process.
Streambased sits across both Kafka and Iceberg as a unified query layer, merging real-time operational signals with historical depth into a single coherent view before the AI agent ever sees it. Property management agents retrieve the precise context they need via MCP – current work order status, contractor availability, tenant payment history, lease terms, recurring fault patterns – consistently timestamped across all source systems, in milliseconds.
The practical effect is an AI that can move from reactive to genuinely proactive. Recurring fault patterns become visible before the next incident rather than after it. Lease renewal risks surface through trends in tenant behaviour rather than calendar reminders. SLA breaches are anticipated rather than reported after the fact. Contractor dispatch is optimised against a complete picture of current availability and historical performance, not just today’s schedule.
Different verticals, same challenge
Across the six use cases in this series – customer support, personalisation, anomaly detection, healthcare scheduling and property management – we see the same underlying challenge.
The problem isn’t actually the AI – it’s that agents are only as good as the context they can see. And in most enterprise environments, that context is divided between systems that don’t talk to each other, don’t agree on time, and can’t be queried together to give the essential full context at the moment a critical decision needs to be made.
Streambased solves this challenge at the data layer, unifying real-time streams and historical depth through Kafka and Iceberg without copying, moving or duplicating data. The agent gets a single, coherent, consistently timestamped view, and so can spend its time reasoning about what to do, rather than reconciling what’s actually happened.
If any of the use cases in this series resonate with challenges you’re facing, we’d love to have a chat about your situation and talk you through what the architecture looks like in practice.
.avif)
.avif)
