What to check before you trust an AI powered property lead

The "AI powered" label problem

Almost every property sourcing tool now describes itself as AI powered. The label has no fixed technical meaning, so on its own it tells a buyer nothing about whether the underlying claims are checkable. LandTech's blog post "AI Land Sourcing Tools Compared: What Actually Matters," published 31 July 2026, sets out a genuinely useful way to cut through this, worth crediting directly. LandTech splits "AI powered" claims into three categories: an embedded chatbot, a general purpose model wrapped in a chat window with no built in way to show where an answer came from; a connected model, where a platform's data is exposed to an external frontier model via a connector, powerful but reasoning happens outside the platform's own guardrails with a real risk of drift; and a purpose built engine, AI logic built natively on the platform's own data, which LandTech argues is the only one of the three that can realistically guarantee every claim traces to a specific, checkable source. Their core point, paraphrased: the useful question is not whether a tool has AI, but which of these approaches it uses and whether you can check its working. LandTech positions its own assistant as the third category, scoped to speeding up early stage triage rather than replacing professional judgment, an honest positioning worth acknowledging.

A fourth category: prediction and scoring models

LandTech's framework was built for a document and chatbot use case: did the AI accurately find or summarise this document. Property sourcing increasingly also uses a different kind of AI, prediction and scoring models that estimate the likelihood an owner will sell within a future window. That is a genuinely different problem with a different verification checklist, because there is no source document to point to. The claim is a probability about the future, not a citation to something already written down, so "can I check the source" has to be replaced with a different set of questions entirely.

A practical checklist for any AI sourcing tool

• Data sources and recency. What underlying data feeds the model, and how current is it, refreshed how often.
• Disclosed methodology. Does the vendor explain what kind of model this is, chatbot wrapper, connected LLM, or a model trained on their own data, and roughly how it works, at a level a technical buyer could sanity check.
• Out-of-sample validation. Is there evidence the model was tested on data not used to train it. Accuracy alone can mislead on imbalanced datasets: a 99% accurate classifier can still be badly wrong on the 1% of cases that matter most.
• Retraining cadence. How often is the model refreshed with new outcomes, and is there drift monitoring between refreshes.
• Explainability. Can the vendor show, for an individual prediction, which underlying signals drove the score, or is it an unexplained number.
• Genuine post-hoc human review. Does a human review the model's output before it is acted on, after the score is produced and tied to the actual outcome, rather than a "human in the loop" claim that just means someone entered data upstream.
• UK GDPR and profiling transparency. Is there a documented lawful basis for building a behavioural profile, and can an individual find out how their profile was built and object to it.

Why "no single headline accuracy number" is not automatically a red flag

Vendors disclosing no methodology at all is a genuine red flag. Vendors who disclose their data sources, model type, features and validation approach, but deliberately decline to publish one headline accuracy percentage, with a stated reason, sit in a different and legitimate category. This is a common, defensible data science position: headline accuracy alone is often a misleading or vanity metric on real world, imbalanced data, and a single flattering number can obscure exactly the cases a buyer cares most about. The distinguishing question is not "did they give me a percentage," it is "did they explain what they measure instead, and why."

AI washing is a real, regulator recognised problem

AI washing is the deliberate or reckless exaggeration of a company's AI capabilities, data assets, or AI driven productivity gains. Recognisable red flags include vague claims with no metrics, methodology or outcomes attached; claims that the AI "does everything" with no supporting evidence; a total absence of any risk or limitation discussion; undisclosed data sourcing, licensing or quality; and "proprietary AI" that turns out to be an undisclosed wrapper around a general LLM API with no fine tuning. Legitimate claims tend to come with some number attached, even if it is not one flattering headline stat, plus visible methodology behind it.

This is not hypothetical. In March 2024 the SEC brought its first ever AI washing enforcement actions, against Delphia (USA) Inc. and Global Predictions Inc., fining them $225,000 and $175,000 respectively. Delphia had falsely claimed to use client data to train an algorithm predicting thousands of publicly traded companies up to two years ahead; Global Predictions had falsely claimed to be the first regulated AI financial advisor. Neither firm had the capabilities it claimed. Separately, Gartner's 2026 research found that of thousands of vendors claiming "agentic AI," only around 130 offer genuinely autonomous agentic features, and predicts a third of companies will damage customer trust in 2026 by deploying AI prematurely. None of this points at any specific property sourcing vendor; it illustrates that regulators and analysts now actively scrutinise unsubstantiated AI claims, which is the stakes this checklist exists for.

The UK specific check: profiling, not just prediction

Under UK GDPR, profiling is any form of automated processing of personal data used to evaluate or predict aspects of a person's behaviour. A model predicting when a property owner is likely to sell is profiling under that definition, regardless of whether Article 22's stricter regime applies on top of it. Article 22 only adds extra restrictions where a decision is solely automated, with no meaningful human involvement, and has a legal or similarly significant effect, the ICO's own examples being things like automatic credit refusal or automated recruitment decisions. Sending a marketing or outreach letter is unlikely to meet that "similarly significant effect" bar the way a credit or recruitment decision does, though this is a judgment call, better phrased as "likely does not meet the threshold" than "definitely does not."

Even where Article 22 does not apply, general UK GDPR principles still apply: a documented lawful basis is required, and individuals must be able to learn how their profile was built and object to it, including for marketing. The genuinely non obvious nuance, stated directly by the ICO, is that mere human involvement somewhere in an AI's lifecycle does not automatically count as meaningful human review. If a human only supplies input data upstream and the model's output then directly determines what happens next, Article 22 can still apply even though a human technically touched the process. For review to be meaningful, it generally needs to happen after the automated output and relate to the actual outcome, not be data entry earlier in the chain. A vendor claiming "a human is in the loop" has not automatically addressed the concern; ask when and what the human actually reviews.

The practical test for a reader is not "is this Article 22 compliant," which an outreach only use case probably does not trigger, but "is this profiling under the broader UK GDPR, and if so, does the vendor have a lawful basis, transparency, and a genuine post-hoc, outcome relevant human checkpoint." Useful questions to put to any vendor: is there a documented lawful basis for the profiling; can individuals learn what data built their profile and object; are there additional safeguards for vulnerable groups; and what is the retention policy limiting how long profiles are kept.

Applying this checklist to GalimAI itself

The point of a checklist like this is that it should not spare its author. GalimAI's own published posts on why machine learning beats rule based filtering and on how the sell-timing model works disclose the data sources feeding the model, Companies House, HM Land Registry and Gazette filings across 30 or more features, that it is a gradient boosted tree model rather than a deep learning or large language model system, that it retrains quarterly with continuous drift monitoring, and that it produces per prediction feature attributions rather than an unexplained score.

On headline accuracy specifically: GalimAI deliberately does not publish one accuracy percentage. It discloses directional lift instead, that the top decile of predicted sellers transacts at multiples of the base rate in the underlying universe, and states why: overfitting risk, and the fact the right metric varies by use case. Judged against the checklist above, that is disclosed methodology with a stated reason, not an absence of methodology, though a reader should judge for themselves whether directional lift framing is sufficient for their purposes.

On human review: GalimAI's published position is that owners flagged with formal Gazette distress signals are scored by the model but also reviewed by a human before outreach, and letter copy is written by people rather than generated, closer to genuine post-hoc review than upstream data entry, though limited to sensitive segments rather than every prediction. On limitations, GalimAI publishes specific caveats: the model cannot capture idiosyncratic causes such as a death in the family, macro shifts can outpace recalibration, and property level scoring is limited to England and Wales. Publishing limitations at all is itself one of the checklist items.

The point is not to pass or fail any one vendor, GalimAI included, but to give an investor concrete questions to ask before treating any "AI powered" lead source as more reliable than demonstrated. For response rate evidence once a lead is sourced, see 160,000 letters and what the response model learned.

Common questions

What does 'AI powered' actually mean for a property sourcing tool?

On its own, very little. LandTech's framework splits AI property tools into an embedded chatbot, a connected model linked to an external frontier model, and a purpose built engine trained on the platform's own data, plus a fourth category for prediction and scoring tools, each needing a different check.

Why don't some AI sourcing tools publish a headline accuracy number?

A single accuracy figure can be misleading on imbalanced real world data. Vendors who disclose their data, methodology and validation approach but decline to publish one number, with a stated reason, differ from vendors disclosing no methodology at all.

Does UK GDPR restrict using AI to predict when someone will sell a property?

Predicting behaviour from personal data is profiling under UK GDPR regardless of use case. Article 22's stricter rules add restrictions only where a decision is solely automated with a similarly significant effect, which an outreach letter is unlikely to meet. General profiling obligations still apply regardless.

What is 'AI washing' and has it actually been enforced?

AI washing is the deliberate or reckless exaggeration of a company's AI capabilities. The SEC's first AI washing enforcement actions, March 2024, fined Delphia (USA) Inc. and Global Predictions Inc. $225,000 and $175,000 for claims they could not support.

This is general information, not legal or compliance advice. Assess any AI sourcing tool, including GalimAI, against your own use case and current UK GDPR guidance before relying on it.