Skip to main content
Patterns to be learnt in tenancy agreements
Product Updates

How AI Recognises Patterns

By Marc Trup, Director of Hobson AI
November 11, 2025

Think of an AI model as a boxer and its training data as the trainer holding up pads.
Over time, the boxer learns to recognise patterns: when the left pad rises, throw a left jab; when the right pad moves, throw a cross. The boxer isn’t thinking about why — they’ve simply learned from repetition that certain movements mean certain actions.

Now imagine the trainer suddenly changes things.
They lift the pad higher, lower, or at a different angle.
The boxer hesitates, throws the wrong punch, or misses entirely.

It’s not that the boxer has forgotten how to punch — it’s that the pattern no longer matches what they’ve seen before.

AI works the same way.
When it’s trained on many examples, it learns to spot the patterns that link input (the text it reads) to output (the answers it gives).
But if the wording, structure, or style of new data is very different from what it’s seen — like an unexpected pad position — it may respond less accurately or pause while it “figures out” what’s happening.

With time (or retraining), the boxer learns to adapt to the new positions — and the AI learns to recognise new patterns too.
Both are improving their ability to generalise — to respond correctly even when the pattern changes a bit.


Applying this to complex leases

When an AI extracts information from leases, it’s using the same kind of pattern recognition.
If every lease followed the same layout — the same “pad positions” — it could identify rent reviews, renewal clauses, and break dates with ease.

But in reality, leases vary in style, age, and wording.

One might say:

“Rent shall be reviewed every fifth anniversary,”

while another buries that same idea deep inside a schedule or a definition.

To the AI, that’s like a trainer suddenly moving the pad to a new position — the pattern looks unfamiliar.
The result may be a missed or delayed “punch” — a clause that isn’t recognised or is interpreted incorrectly.

As the model encounters more of these variations, it learns to adjust, improving its accuracy in handling even the most complex and unconventional leases.

This article was also published by the Property Investors Bureau.