AI Disclosure

The EU AI Act (Regulation (EU) 2024/1689), national AI guidance from the U.S. NIST AI RMF, the UK's AI Regulation White Paper, the Turkish Personal Data Protection Authority's automated-processing guidance, and modern transparency expectations require us to disclose how AI works in the Service. This page is that disclosure.

Given a point or polygon on a map, the Service combines third-party point-of-interest data with geospatial computation, then asks a large language model to interpret the combination as a narrative analysis: cultural “DNA”, foot-traffic estimates, competitor mapping, sector opportunity scoring, negative signals, and spillover potential. The model takes structured numeric and categorical inputs and emits a structured, structured-then-textual output.

• Google Gemini 2.5(Flash and Pro variants) — accessed through Google Cloud's Vertex AI under contract. Used for narrative synthesis and structured extraction.
• Deterministic geospatial code (not AI) — H3 hexagon binning, distance and density computations, scoring formulas. These are auditable code paths, not statistical models.

We may change AI providers from time to time (notice in this page's version history). We currently do not use any open-source local model.

We do not train, fine-tune, or in-context-cache machine learning models on Customer Data, Customer prompts, or generated Reports. Our contract with the AI provider prohibits the provider from using our prompts and outputs to train its models. We do not retain prompts in the provider beyond what is technically necessary for response, and the provider commits to a short retention window for abuse monitoring as published in its own documentation.

The underlying foundation model was trained by its provider on publicly available text and on data the provider sourced under its own legal basis. We do not control that training and we make no representation about its lawfulness.

When you start an analysis, the Service sends the AI provider the coordinates or polygon you chose, the parameters you set, and a list of points-of-interest retrieved from third-party data sources. We do not send your email, billing identity, IP address, or any other Account-identifying data with that prompt.

The transmission to the AI provider is over TLS and is subject to the provider's DPA listed on our Sub-processor page.

The Service is well-suited to:

• exploring areas you don't know;
• structuring qualitative impressions of neighbourhoods;
• generating shortlists of sector hypotheses;
• preparing dossiers for human review.

It is not suited to:

• any decision about a specific individual's rights, opportunities, or money;
• investment, real-estate, lending, or insurance underwriting;
• policing, public-safety scoring, or any law-enforcement use;
• tasks requiring guaranteed factual accuracy or numerical precision;
• tasks involving regulated demographic, ethnic, or sensitive categories.

• Hallucination. The model may invent a venue, a statistic, or a relationship that does not exist. Even structured outputs can be wrong.
• Stale data. Point-of-interest sources are often months out of date; shops open and close faster than they can be indexed.
• Coverage bias. OpenStreetMap and similar sources are denser in wealthy/Western/urban areas, sparser elsewhere — outputs can mirror that bias.
• Reasoning bias. Large language models inherit biases from their training data and may reproduce stereotypes about neighbourhoods.
• Non-determinism. The same query can return different narratives.
• Prompt sensitivity. Small changes to inputs can produce disproportionate output changes.

You must apply meaningful human oversight to any decision you make on the basis of an AI Output. The Service is decision-support, not a substitute for human judgement. AI Outputs are not automated individual decisions producing legal effects or similarly significant effects within the meaning of GDPR Article 22; you remain the decision-maker.

You may not deploy, use, or repackage the Service in any way that would constitute a prohibited practice under Article 5 of the EU AI Act, or a high-risk AI system under Article 6 / Annex III, unless you independently fulfil the conformity assessment, fundamental-rights impact assessment, registration, transparency, human oversight, and post-market monitoring obligations the AI Act imposes on deployers. Specifically, the Service must not be used for:

• biometric categorisation by race, political opinion, trade-union membership, religion, philosophical belief, sex life, or sexual orientation;
• social scoring of natural persons by public or private actors;
• predictive policing of individuals;
• emotion recognition in workplaces or educational institutions;
• access to or enjoyment of essential private and public services (housing, credit, insurance, public benefits) where the analysis would materially influence outcomes about specific individuals; or
• employment decisions about specific candidates or workers.

See Terms §14 and AUP §4 for the full list and the consequences of breach.

We monitor outputs for systematic skew and patch promptly when discovered. We publish changes here in subsequent versions of this Disclosure. If you believe the Service has produced biased or harmful content about a place, a community, or a business, please email abuse@poigeo.app.

PDF Reports include the label “AI-assisted” on each page and a footer identifying the model family and the date of generation. Metadata records the prompt template version. Do not remove these markers — doing so violates the AUP and may, in the EU, breach Article 50 of the AI Act.

If you are an EU/EEA, UK, Swiss, Brazilian, or Turkish data subject and an AI Output significantly affects you, you have the right to obtain human review and to contest the output. Send your request to privacy@poigeo.app. We respond within 30 days (or as required by the applicable law).