Artificial Intelligence is often presented as one single thing. In reality, it is a family of technologies that developed in stages, each building on the one before it. Understanding these layers helps separate what AI can genuinely do from what is marketing noise — a useful distinction for anyone working in property, where decisions must remain defensible.

Think of it as a family tree with six branches: Classical AI, Machine Learning, Neural Networks, Deep Learning, Generative AI, and Agentic AI. One caution before we start: this is not a clean historical staircase. The layers overlap. Transformers, for example, are a deep learning technology and also the engine behind the language models most people now use daily.

Stage 1 — Classical AI: telling the computer the rules

The earliest form of AI was rule-based. A human wrote the logic, and the computer followed it.

In valuation terms, this is the old expert-system approach:

• If a property is older than 40 years and in poor condition, apply higher depreciation.
• If it has a sea view, add a premium.

This is useful and transparent, but limited. The computer does not learn anything. Every rule has to be written by hand, and real markets rarely fit neat "if X, then Y" logic.

Stage 2 — Machine Learning: letting the computer learn from data

Instead of writing every rule, we give the computer past data and let it find the patterns itself.

Feed a model thousands of past sales — area, age, district, plot size, price — and it learns relationships on its own: larger area usually lifts price, location matters, older buildings often sell for less, coastal properties tend to command a premium. This family includes regression, classification, random forests, gradient boosting, and clustering.

For most property price prediction work, including automated valuation models, this is the relevant layer. It is essentially a regression problem.

Stage 3 — Neural Networks: a brain-inspired structure

A neural network is a set of connected nodes: inputs flow in, hidden calculations happen in the middle, and an output comes out the other end.

For a valuation, the inputs might be area, location, age, and bedrooms, and the output an estimated price. The model repeatedly adjusts its internal "weights" until its predictions get closer to reality. The mechanics have names — perceptron, cost function, backpropagation, activation function, hidden layers — but the core idea is simple: the model learns by measuring how wrong it is and correcting itself.

Stage 4 — Deep Learning: neural networks with many layers

Deep learning is simply a neural network with many hidden layers rather than one. The word "deep" refers to that depth.

Each layer learns something more abstract than the last. With property photos, an early layer might pick up edges and colours, a middle layer might recognise windows and roofs, and a later layer might conclude "this looks like a renovated kitchen" or "this looks like a house in poor condition."

Different structures suit different data: CNNs for images and computer vision, RNNs and LSTMs for sequences and time series, transformers for text and language, autoencoders for compression and spotting anomalies. In practice, a CNN could scan property photographs and flag a sea view or a luxury finish, while a transformer could read valuation text and understand phrases like "no separate title deed," "share," or "right of way issue."

Why deep learning became powerful

The concepts existed for decades but only became practical when three things came together: far more data (digital records, images, GPS, online listings, transaction databases), far more computing power (particularly GPUs), and better algorithms. The breakthrough moment came in the 2010s, first in image recognition and later in language.

Stage 5 — Generative AI: AI that creates

Generative AI, built mostly on deep learning, produces new content rather than just predicting a number — text, images, code, summaries, reports. This is the layer behind large language models, image generators, and multimodal systems that handle text and images together.

In valuation, you could hand it a set of comparables, subject property details, and location notes, and it could draft market commentary, a SWOT, or an explanation of assumptions. The critical caveat: it can produce confident but incorrect statements, so every output must be checked before it goes anywhere near a report.

Stage 6 — Agentic AI: AI that acts

The newest layer does not only answer questions — it can plan steps, use tools, search files, read databases, run code, and carry out a task from start to finish.

The difference is practical. A generative system might tell you, "You should analyse the sales comparables." An agentic system could open the database, filter the comparables, remove outliers, run the model, generate charts, draft the wording, and flag weak assumptions — as a sequence, with memory of the context. That is the meaning behind terms like planning, tool use, and autonomous execution.

Bringing it together with one property

Picture estimating a house price at each layer:

• Classical AI applies fixed rules: old house, reduce 20%; sea view, add 10%.
• Machine Learning studies past sales and learns the relationship between area, age, location, plot size, condition, and price.
• Deep Learning may capture more complex interactions — recognising that an old house near the beach is not simply "old," but potentially a redevelopment opportunity.
• Generative AI writes the commentary: "The property benefits from proximity to the coastline; however, age and condition reduce immediate marketability."
• Agentic AI runs the whole workflow: finds comparables, checks for title issues, runs the model, drafts the report, builds a sensitivity table.

The key takeaway

Deep learning is not magic. It is machine learning built on large, multi-layer neural networks, and it is genuinely strong where there is plenty of good data, complex patterns, and a mix of images, text, and maps to work with.

But for valuation, the safest position stands firm: AI can assist with speed and scale, yet professional judgement and defensible assumptions remain essential. The tool does not carry the responsibility — the valuer does.

Trifonas Mamas
Property Valuer (MRICS–ΕΤΕΚ)
Registered Estate Agent

MSc Data Analytics Candidate