Should you buy CRE portfolio SaaS or build custom AI?

AdvancedUNO
10 Aug, 2026
Should you buy CRE portfolio SaaS or build custom AI?

7 min read

The Allocation Dilemma

  • For the Managing Director: Deciding whether to allocate capital to verticalized SaaS platforms or direct-to-foundation-model custom software.
  • The Hidden Friction: Building custom models introduces massive, recurring data-cleaning overhead that software vendors historically absorbed.
  • The Immediate Directive: Freeze new multi-year enterprise SaaS commitments exceeding $250,000 until you audit your internal data-ingestion pipelines.

The Quiet Fracturing of the PropTech Wrapper

The light in the midtown conference room is always flat, reflecting off glass tables and the blue screens of Excel models that have not changed in a decade. We sit here because we want to believe that the right software will finally make the real estate behave. For years, the promise of commercial real estate (CRE) portfolio SaaS was simple: pay a recurring fee, upload your messy PDFs, and watch your net operating income (NOI) clarify. Deciding between off-the-shelf CRE portfolio SaaS and bespoke enterprise AI models is fundamentally a choice of where you want to store your operational risk.

Now, that promise is fracturing. The rise of custom foundation models has introduced a cold, competitive tension between buying a standardized platform or building a proprietary intelligence layer directly on top of raw data. This choice lands on the roadmap this quarter because the economic margins of traditional property management are tightening. The pressure to compress operating expenses and defend cap rates leaves no room for redundant software licenses that merely act as expensive databases.

The market is moving quickly. Wall Street titans like Blackstone and Brookfield are actively betting on custom AI, pushing the industry toward a future where major firms bypass traditional proptech intermediaries altogether [1]. Anthropic’s $1.5 billion joint venture with financial leaders including Blackstone and Goldman Sachs signals a structural shift [1]. Foundation model developers are moving upstream, targeting the core underwriting and portfolio management workflows that software vendors once claimed as their exclusive territory [1]. If custom AI becomes cheap and ubiquitous, the traditional SaaS wrapper begins to look like an expensive, redundant layer.

The OCR Mirage and the Cost of Unstructured Leases

The friction is never in the vision; it is in the plumbing. Property data is notoriously stubborn, resistant to clean schemas, and scattered across decades of inconsistent property management systems. When JLL and Slate Asset Management announced their joint venture to commercialize JLL Asset Beacon, they acknowledged this exact pain point [2]. Asset Beacon integrates financial, operational, and leasing data to create a single source of truth, relying on JLL Falcon’s generative AI for lease abstraction and entity resolution [2]. But when you try to replicate this level of integration in-house, the illusion of cheap, custom AI quickly dissolves.

To understand the depth of this challenge, consider a representative secondary-market office portfolio. The acquisition team decides to bypass third-party platforms, opting instead to build a custom LLM pipeline using raw API connections to a foundation model. The goal is simple: automate the extraction of tenant recovery clauses to calculate net effective rent. The project stalls almost immediately because the historical leases, scanned over a fifteen-year period, contain poor OCR quality. Tenant names vary wildly across documents—one lease lists "ACME Corp," another "ACME Holdings," and a third simply "ACME."

The Danger of Algorithmic Hallucinations in Cash-Flow Projections

The custom model, lacking a sophisticated entity resolution engine, treats these variations as separate entities. It misses a critical co-tenancy clause in a major anchor lease, leading to an unexpected drop in occupancy when a satellite tenant exercises their right to terminate. The mistake quietly bleeds cash flow before anyone realizes the algorithm misread the unstructured text. This is where the build-your-own thesis breaks down: the cost is not in the API tokens, but in the human labor required to clean the data before the model ever sees it.

"We spent $400,000 building a custom underwriting interface only to realize our analysts were still manually copying rent rolls from PDFs because the model couldn't parse local municipal tax clauses."

This operational reality is why the historical pitch decks of early proptech startups often underestimated the sheer drag of real-world implementation [4]. While companies like Matterport and OpenDoor successfully productized specific slices of the transaction cycle, the daily grind of running a multi-family or commercial portfolio remains a fragmented exercise [4, 5]. G2 reviews of property management tools consistently reveal that no single platform fits every portfolio; some teams require deep accounting integration, while others prioritize tenant communication [6]. The dream of a single, unified software solution frequently collides with the messy reality of physical assets.

Weighing the Friction of Custom AI Against Standardized SaaS

The decision to build or buy is not a matter of choosing the superior technology. It is an operational trade-off between two distinct types of organizational friction. Each path has its own balance sheet, and each breaks under different operational pressures.

Building custom AI requires your organization to become a software company. You must hire and retain machine learning engineers, manage API latency, and establish rigorous data-cleaning protocols. The benefit is absolute control. Your proprietary underwriting algorithms remain entirely yours, protected from competitors who use the same commercial platforms. You can tailor the model to your specific niche, whether that is navigating regional price volatility in Australia’s tight rental markets or optimizing yield under the National Housing Accord's five-year delivery targets [3].

Rule of Thumb: If your portfolio is under 10 million square feet of gross leasable area, building custom AI is an expensive vanity project; stick to standardized SaaS platforms that force operational discipline onto your leasing teams.

Buying an enterprise platform like JLL Asset Beacon shifts the development burden to the vendor [2]. The vendor handles the security patches, the OCR updates, and the API integrations. The trade-off is conformity. You must adapt your workflows to the software's pre-built data structures. If your investment thesis relies on a highly non-standard leasing structure, you may find yourself fighting the software’s rigid fields, or paying six-figure customization fees to get the reports your investment committee demands.

Major Capital Commitments to Real Estate and AI Initiatives
Housing Australia Future Fund10 $BAnthropic-Blackstone JV1.5 $B

Figures compiled from the sources cited below.

The Three-Phase Data Integration Playbook

To avoid over-allocating capital to vanity tech projects, operators must follow a strict, sequenced playbook that prioritizes data readiness over algorithmic sophistication.

  1. Audit your unstructured data inventory: Before writing a single line of code or signing a SaaS contract, run an audit on your lease agreements, utility bills, and historical T12 statements. If more than 15% of your files are non-searchable PDFs, your immediate priority is OCR normalization, not advanced AI. You cannot run a custom model over data that a human analyst cannot easily read.
  2. Establish the system of record: Decide where the cleaned data lives. If utilizing a platform like JLL Asset Beacon, ensure all financial and operational metrics flow into a unified database before attempting to layer natural language query interfaces over them [2]. The database must serve as the single source of truth, with clear ownership assigned to the property management team.
  3. Implement scoped automation pilots: Deploy targeted AI tools—such as JLL Falcon's lease abstraction or specific underwriting scripts—on a single, isolated fund [2]. Run parallel manual checks for 90 days to verify entity resolution accuracy. Do not scale the technology across the entire portfolio until the error rate on critical lease terms drops below 1%.

Frequently Asked Questions

What happens to our proprietary underwriting models if we upload our historical deal data to a third-party CRE portfolio SaaS?

Most commercial SaaS vendors include terms in their enterprise service agreements that allow them to use anonymized, aggregated data to train their internal machine learning models. If your underwriting formulas represent a genuine, mathematically proven alpha, uploading your historical deal sheets to a shared platform risks diluting your competitive advantage. In this scenario, you must negotiate strict data-isolation clauses or opt for a private cloud deployment of the software.

How do we handle the tenant data privacy requirements under GDPR or local equivalents when using generative AI for lease abstraction?

Generative AI tools cannot be deployed in a vacuum. When abstracting leases that contain personally identifiable information (PII)—such as guarantor names, home addresses, or personal financial statements—you must ensure the AI pipeline uses zero-data-retention APIs. If you are using a platform like JLL Asset Beacon, verify that the underlying JLL Falcon infrastructure does not store or use tenant PII to train public models, and ensure your data processing agreements align with local regulatory frameworks [2].

If we choose the custom AI build path, what is the realistic run-rate cost to maintain model accuracy as leasing structures change?

Building the model is only 20% of the total cost of ownership. The real expense lies in model drift and pipeline maintenance. As municipal tax laws change, utility billing structures shift, and new lease templates are introduced, your custom pipeline will require constant retraining. A typical institutional operator should budget at least $150,000 annually for dedicated data engineering support to maintain a custom AI pipeline, even if the underlying foundation model APIs remain cheap.

The Strategic Verdict: Choose custom AI only if your proprietary underwriting represents a genuine, mathematically proven alpha that off-the-shelf platforms cannot replicate. For 90% of institutional owners, the path to yield lies in clean data plumbing, not custom models. If your data layer is a mess, walk away from both options and spend the capital on basic data hygiene first.

Related from this blog

Sources

Next Post Previous Post
No Comment
Add Comment
comment url