Lorenzo Bianco on What AI Can and Can’t Replace in Trust & Safety

Estimated reading time: 6 minutes

The Product Manager behind Immobiliare’s trust and safety architecture. On moving from 60% human review to a system where AI and humans make each other better.

Executive summary (TL;DR)

Most product managers talk about features shipped and metrics moved. Lorenzo Bianco talks about trust.

Lorenzo is Product Manager for the B2C team at Immobiliare.it — Italy’s #1 real estate portal. The platform has over a million active listings and is growing across Southern Europe. His role covers how sellers list, rent, and sell properties. It also covers how the platform catches fraud before it reaches buyers.

That work recently took him to the stage at DMA Conference 2026. He joined platform leaders from across Europe to share what Immobiliare.it has learned building moderation at scale.

His central argument is simple. The platforms getting trust and safety right aren’t the ones chasing full automation. They’re the ones who figured out that automation, AI, and human judgment work best as a stack, not as alternatives.

Step One: Automate the Simple Stuff (Ruthlessly)

The starting point for any moderation system is removing obvious decisions from the human queue. For Immobiliare.it, that meant moving away from hard-coded rules. Before the change, updating filters required coordination across product and engineering, development time, release cycles, and post-release analysis. As Lorenzo puts it:

“We felt slow, not only in making iterations, but also in analysing results. What we do now in minutes took days before.”

At the same time, more than half of all cases were going to human review. The team spent most of its time manually checking listings that followed predictable patterns. It worked but it made scaling improvements slow.

The priority became shifting that balance. Today, more than 95% of cases are handled automatically, mainly the clear, repeatable scenarios that don’t require human judgment. Automating the foundational layer created the space to focus people and systems on more nuanced decisions higher up the stack.

Step Two: Use AI to Build Context, Not Just Enforce Moderation Rules

Automation handles the obvious. AI handles the ambiguous. That distinction is where most moderation systems win or fail. Lorenzo explains it with a simple question: how do you spot a fraudulent property listing in Rome?

“Counterintuitively, a very good-looking listing can be a scam. Scammers don’t necessarily publish low-quality ads. Often they use paid visibility, good photos, and polished descriptions. On the surface, everything looks legitimate.”

The real signal is rarely the listing itself. It comes from the surrounding context, behavioural patterns, account history, naming combinations, or the sequence of actions before a listing went live. Individually, these signals look harmless. Together, they reveal risk patterns that static rules can’t catch.

The same applies to message moderation. Traditional systems ask: does this message contain a link, a phone number, or suspicious wording? But in a marketplace, those things are often completely normal.

“The most important thing is not only the content of the message, it’s the context of the conversation.
You can be very strict and say every message with a link is suspicious. But it’s very common among people to share links or phone numbers. You can’t just use a high-level rule. You have to have the whole picture.”

This is where AI changes the approach: not by replacing human judgment, but by connecting fragmented signals into a more contextual understanding of risk.

Step Three: Make the Human Touch Your Luxury Differentiator

The dominant narrative in trust and safety right now is about pushing automation rates as high as possible. Many teams are racing toward 100% automated review. Lorenzo believes that’s the wrong goal.

“We shouldn’t aim for 100% automation. We still need humans in the loop for the most critical, ambiguous and high-value cases.”

For Immobiliare.it, the remaining 3–4% of manually reviewed cases are not considered system failures. They are intentional escalation points: complex or ambiguous situations where human judgment adds real value.
What changed is not the existence of human review, but the role humans play inside the process.

Previously, reviewers had to reconstruct every case manually, gathering account information, checking behavioural patterns, and piecing together the context needed to make a decision. Today, much of that preparatory work is handled upstream by AI systems, allowing reviewers to start from a more informed position.

“Two years ago, even when a case reached a human team, they still had to do all the research themselves,” Lorenzo says. “Now AI helps frame the case and orchestrates the critical ones with context, so humans can focus on the actual decision-making.”

This also changed what the moderation team does day to day. What was repetitive validation became analytical work like reading signals, spotting emerging patterns, and feeding structured feedback back into the system.

“Before, a lot of the work was manually checking listings and trying to understand whether they were legitimate. Now the focus is much more on reading signals and helping improve the moderation system itself. I don’t see it like AI is replacing humans, AI is empowering humans within our company.”

Human involvement also remains essential for another reason: continuous improvement. There’s a second reason humans stay in the loop that Lorenzo considers equally important: the feedback loop itself.

Without that feedback loop, the system stops learning.

Key Takeaways for Product and Trust & Safety Leaders

• Start by flipping the ratio. If more than half your moderation cases are going to human review, the foundation isn’t right yet. Immobiliare.it moved from 60% human review to 95% automation — and that shift is what made everything else possible.
• Context beats content every time. Moderation built on isolated signals will always be outpaced by sophisticated fraud. The winning architecture synthesises account history, behavioural patterns, and conversation context together before any decision is made.
• AI’s highest value is preparing the ground for human judgment. The goal isn’t AI instead of humans, it’s AI that assembles the full picture instantly, so that the 3–4% of cases that need a human can be resolved fast, confidently, and at scale.
• Human reviewers are a learning system, not a fallback. Moving from four people skimming listings to two people generating feedback signals isn’t a headcount reduction — it’s a capability upgrade. That feedback loop is what
  keeps the whole system improving.

This article is based on Lorenzo Bianco’s session at DMA Conference 2026 (Digital Marketplace Association), where he joined platform leaders from across Europe to discuss the future of AI in trust and safety in online marketplaces.

Questions We Hear a Lot

Share this article