From contract headache to strategic asset: How Edsvard makes sense of unstructured real estate data with IBM Cloud and watsonx.data
Learn how this Swedish technology provider uses AI to extract key data from complex contracts, flag risks and protect revenue.
Real estate contracts are prime candidates for automation. They’re dense, inconsistent across owners and jurisdictions and packed with small clauses that carry significant financial consequences. These clauses include renewal windows, indexation formulas, service inclusions, tenant obligations and exceptions that never make it cleanly into downstream systems.
Sweden’s Edsvard Hallbarhet AB recognized this challenge and created a system to transform the broad array of documents associated with real estate transactions and property management into structured, machine-checkable data at ingestion time. This approach means that customers can stop manual indexing of documents, detect mismatches in invoices early and surface risk-bearing clauses before they become disputes or lost revenue. All in a simple-to-use, AI-powered interface.
The AI company built a solution atop IBM Cloud®, chosen for its flexible hybrid capabilities, and watsonx.data®, whose open lakehouse capabilities helped streamline and simplify Edsvard’s architecture. This approach simplified the process by removing the need to move data for enrichment and processing.
Edsvard’s solution—Contract Intelligence—uses AI to extract key terms and data points directly from contracts. This analysis is then cross-checked against other systems of record such as invoices, ledgers, energy declarations and notifications to find discrepancies and trigger action.
Beyond numeric reconciliation, Contract Intelligence also flags risk-bearing clauses that can materially change a contract’s exposure even when the headline numbers look accurate. Their solution needs to work for both small and large deployments. Everything from 100 documents for specific values to larger-scale rollouts across hundreds of thousands of documents, as is common among the company’s US customers.
Technically, the platform is built as a two-tier system. A source layer contains a range of unstructured data—PDFs, Word documents, spreadsheets, slide decks and even photos from mobile devices. The second tier consists of an AI layer that runs Contract Intelligence (and a parallel Invoice Intelligence capability) exposed through an API.
An architectural view of the Contract Intelligence solution
The solution runs as a service on IBM Cloud and is also designed to run in watsonx.data, on other cloud providers, and on-premises when required. That hybrid requirement emerged in the field: larger customers increasingly insisted on private-cloud or on-premises deployments due to regulatory and privacy concerns. IBM’s “hybrid by design” strategy was a key reason that it appealed to the Edsvard team.
Inside the AI systems layer, the pipeline is deliberately straightforward and production-minded: text extraction for native digital documents; optical character recognition (OCR) for image-based input; text preparation; and named-entity recognition as part of the normalization process before model inference.
Depending on the step, a single component might be assigned to run the task, or there might be several, from which the most appropriate is automatically selected. For example, the OCR step might employ the open source Tesseract, PaddlePaddle or another OCR engine.
The differentiated work happens in the modeling and adaptation loop. Edsvard employs an annotation system and a tailored language model strategy: the team elected to build their own custom model optimized for specific target languages, including English, Swedish, Spanish, Italian and several others. The work included more domain training for contract types and client-specific fine-tuning driven by annotations.
All these actions are aimed at extracting the exact fields that matter (contract numbers, validity periods, property area, building sections, rent and renewal or termination timing, for example).
Edsvard opted to build its own model so they could optimize its size, architecture and training, leading to both faster training and inference. This action also resulted in a smaller memory footprint. The team found this approach to be more deterministic, ensuring the best possible data quality.
The extracted fields are assembled into an “accurate state of the contract,” persisted to a time-series store (watsonx.data) and then surfaced through APIs and dashboards. For contract intelligence, Edsvard ultimately built a custom dashboard for the output—the team initially used Elasticsearch Kibana but couldn’t reach the usability level required. For invoice intelligence, sometimes customers prefer the extracted values to be pushed directly back into their own systems.
An example of the Contract Intelligence interface
A key architectural decision—relevant for both scalability and TCO—is that Contract Intelligence is delivered as a REST service and is not bound to any single document portal.
The Samporten source system (a project-tracking software offering developed by Edsvard) was a strong fit for the original property and construction use cases. However, the new system can easily integrate with an existing DMS/CRM/ERP environment and it can, for example, crawl SharePoint to process documents where they already live.
Watsonx.data is a key enabler here: it allows the platform to bring in the data and process it in place, minimizing data movement and bespoke extract, transform and load (ETL) pipelines while enriching data where it sits. This process reduces operational overhead, latency and potential failures.
The business impact centers on measurable operational outcomes that engineers and product managers tend to care about: a more than 90% reduction in manual handling, better data quality (less information lost between systems), improved forecasting capability with common calculation models and better financial stability. Often, Edsvard’s customers also report an improved ability to negotiate with banks.
On the operational side, the platform makes it easier to scale document workflows from hundreds of documents to thousands and to onboard new properties by processing contracts rather than rebuilding processes. For property teams, more accurate invoicing terms and clearer renewal windows can also reduce tenant disputes and shorten billing reconciliation cycles.
By building on IBM Cloud and watsonx.data—and designing for hybrid deployment from early on—Edsvard turned a contract-reconciliation challenge into a modular document-AI platform: one that can deploy as a service, integrate with existing enterprise systems (including SharePoint), and scale from small Nordic AI assistant deployments to multi-location rollouts without rebuilding the ingestion and data layer.
Looking forward, the team plans to build with the IBM Power11™ to offer a “single box,” end-to-end, on-premises solution. The team is also exploring how agentic AI might automate follow-on actions once contracts are understood. The net result is a foundation that can expand to new document types, new environments and new workflows—while keeping the engineering surface area stable and the user experience focused on accuracy and usability.