AI 101 Series, Part 2
Last month, we kicked off this series by answering the most basic question: what actually is AI? The short version is that AI is software that learns from examples, recognizes patterns, and automates work that used to require human judgment. If you missed it, start there.
Now we go one level deeper. When an AI system reads your lease, pulls data from your rent roll, or flags a budget variance, what is actually happening? Most vendors skip this part. They show you a demo, point to time savings, and move on. But if you understand how these systems actually work, you ask better questions, make better buying decisions, and use the tools more effectively. That is the point of this series: to provide a clear explanation of how AI learns, processes information, and produces outputs you can act on.

Breaking Down Machine Learning
Here’s the simplest way to think about AI. AI enables machines to learn so they can make useful decisions without human intervention. The two primary methods of machine learning are Classical Machine Learning (ML) and Deep Learning (Deep Neural Networks).
Classical Machine Learning
Classical ML dominates when the problem is deterministic, and the data is structured (think: is a transaction fraudulent—yes or no?). It doesn’t reason, hallucinate, or drift. It’s deterministic and produces a specified output every time.
What it does, very well and very fast, is spot patterns and use them to make predictions about data it has never seen before. Feed a well-trained AI system thousands of commercial leases, and it learns what a lease looks like. It learns where rent escalation clauses typically appear, and how to distinguish a base rent figure from a CAM charge. It does not understand a lease commencement clause, but it recognizes them because it has seen similar patterns numerous times.
Deep Learning and Large Language Models
Deep Learning has a different approach and shines when the problem is probabilistic, and the data is unstructured (think: interpreting scanned documents or classifying images). This is the method used to train Large Language Models (LLMS).
LLMs are trained using transformers, which are massive Deep Neural Networks trained on a large collection of data. They’re trained to predict the next word (or token) in a sequence. Due to their comprehensive multi-layer training, LLMs produce remarkable results across multiple domains. Since they’re trained to be conversational, you can teach them new things with only a few examples—also known as zero-shot prompting. You can train LLMs on industry-specific terminology to better predict next steps and ways of “reasoning.” Think of it as giving the model a chain of thought to follow and power “reasoning.” It can explore a topic broadly, backtrack, and refine its answer to better align with your desired outcome.

The Loop: Input, Processing, Output, Feedback
Every AI interaction follows the same basic cycle, starting with an input. In real estate, that might be a PDF lease, a scanned rent roll, an email from a tenant or resident, or a data export from your property management system. The system then compares that input to the patterns it learned during training. Then, it produces an output: a summarized lease abstract, a flagged anomaly in your financials, a predicted renewal probability, or a drafted response to a maintenance request.
Here is the part most vendors skip over: for most AI systems, that output is a statistical prediction, not necessarily a guaranteed truth. When an AI system extracts a lease commencement date or calculates a CAM charge, it is giving you its best prediction based on patterns it has seen before. A well-built system will be right the overwhelming majority of the time by setting up guardrails and verifications. But most outputs deserve a human eye on anything that carries financial or legal weight. Think of it the same way you think about a sharp junior analyst—accurate, fast, and worth reviewing before you sign anything. For certain important workflows, human review and approval (or human-in-the-loop) becomes a dedicated step built into the process. This way, workflows have integrated approvals at the pivotal points—such as data extraction or output generation.
The fourth and final stage of the loop is feedback. Over time, corrections and new data improve the system’s accuracy. Good AI products are not static tools. They get better as they process more of your data and as teams flag errors and corrections.
AI Learns by Examples, Not Rules
Traditional software follows rules that programmers write by hand. If a lease includes this clause, do this. If the payment is more than 30 days late, flag it. These rule-based systems work well for predictable situations and fall apart when reality gets complicated, which, in real estate, it always does.
AI takes a different approach. Instead of rules, it learns from labeled examples. In supervised learning, developers feed the system thousands of documents with correct answers already filled in. Here is a lease, commencement date, base rent, and termination clause. The system figures out, on its own, how to find those answers in new leases it has never seen before. In unsupervised learning, there are no pre-labeled answers. The system groups similar things together and surfaces patterns that look unusual. This is how anomaly detection works in financial analysis: the system flags a transaction that does not look like any of the normal patterns in your portfolio. In reinforcement learning, the system tries different approaches, receives feedback on what works, and gradually improves its own responses over time.
For real estate applications, the key insight is this: the quality of what the AI produces depends directly on the quality of the data it trained on. A system trained on commercial leases across office, multifamily, retail, and industrial assets will outperform a general-purpose tool every time. Real estate is full of specialized terminology, document structures, and financial frameworks that generic AI simply does not know. Domain-specific training is not a marketing claim—it separates useful from unreliable.
Understanding Training and Models
Two words you will constantly hear in AI conversations are “training” and “model.” They are straightforward concepts.
Training is the process of feeding data into a system repeatedly so it refines its pattern recognition.
A model is the result of that training: a tuned algorithm that applies learned knowledge to new inputs. Think of it like building expertise through repetition. The more quality examples a system processes, the more reliable its outputs become.
Once developers train a model, they validate it by testing it against data it has never seen before. If it performs well on unfamiliar documents, it is ready to deploy. If it does not, more training is required. The best AI companies treat this as a continuous process, not a one-time effort. Real estate does not stop changing, and a model that performed well two years ago may need updates to stay accurate today as new lease structures emerge, market conditions shift, and document formats evolve. Ongoing model improvement is the difference between a tool you can trust and one that drifts out of reliability.
Why This Matters for Your Team
Understanding how AI actually works changes how you evaluate the tools available to you. A system not trained on real estate data will produce inconsistent results because it has never seen your document types, your terminology, or your financial frameworks. A system with no validation controls will surface outputs with no way to know how confident it actually is. A system that cannot handle unstructured data from multiple sources will create more work than it saves.
At Outcome, we built our platform specifically to address these gaps. AI is not magic. It is a powerful, learnable technology. The more your team understands how it works, the better positioned you are to use it well and ask the right questions of any vendor seeking your business.
Stay tuned for the next article in our AI 101 series.

You must be logged in to post a comment.