AI Transparency & Data Trust Notice

1. Why We Wrote This

SCORE offers optional AI-powered quality-control features (we call them AIQC) — including optical character recognition (OCR), automated quality checking, and structured field extraction from survey reports. These features are only available on certain subscription tiers, and you can choose to enable or disable them per account.

When you use AIQC, your survey content is sent to our AI processors. Because that processing involves your clients' personal data, you deserve a plain-English explanation of exactly what happens to it, who sees it, and what we do — and don't do — with it.

This notice sits alongside our Privacy Policy and our Data Processing Agreement. Where any of them disagree, the DPA wins for contractual purposes.

2. Our Commitment in Plain English

When you use AIQC:

• Your data is never used to train AI models. Not ours, not our AI processors', not anyone else's.
• Your data stays in the UK and EU. Storage is in the UK (London); AI processing happens only inside our two trusted EU/UK AI LLM processors — AWS Bedrock (UK London / EU Ireland) and Mistral AI (EU France). Nothing leaves UK/EEA jurisdiction.
• Your data is encrypted at every step. AES-256 at rest, TLS 1.2+ in transit — including the round-trip to our AI processors.
• A human reviews every AI output. AIQC is a quality-check assistant, not an automated decision-maker. Nothing acts on AI output without surveyor sign-off.
• You can control AIQC. It is an opt-in feature on your account that you can enable or disable. If you disable it, no survey content is ever sent to an AI processor.
• We delete on request. If you delete a job or close your account, the underlying survey content is removed from our systems on the timelines in our Data Retention Policy.
• We improve AIQC based on your feedback (optional). You can choose to share feedback on how well individual AIQC checks are working for your account. This feedback helps us understand which checks are reliable and which need improvement. We collect only which checks you flagged and your assessment — never your client data or survey content. This is performance feedback, not model training.

The remainder of this notice explains how each of those commitments is enforced.

3. What AIQC Actually Does

When a surveyor uploads a survey report PDF to SCORE, AIQC performs three jobs:

• OCR — extracts the text from the PDF so it can be searched and validated.
• Quality checking — runs the extracted text against rules (consistency of figures, presence of required RICS fields, missing inspection elements) and flags issues for the surveyor.
• Field extraction — pulls structured values (e.g. property address, RICS condition ratings, valuation figures) into the SCORE database so they can be reused, exported and audited.

Each task is routed to the AI LLM that gives the best result — we may use AWS Bedrock, Mistral AI, or a combination depending on the task and current model performance.

The output is a list of suggestions on a dashboard. The surveyor accepts, edits or rejects each one. AIQC never sends a report, generates a final document, or alters a record on its own.

4. What Data Goes to Our AI Processors — and What Doesn't

SCORE uses two EU/UK-resident AI LLM sub-processors — AWS Bedrock (UK London / EU Ireland) and Mistral AI (EU France) — deliberately chosen so we can pick the best model for each AIQC task without ever taking customer data to a new vendor. We route OCR, quality checking, and field extraction to whichever AI processor gives the best result.

Full details for both are in our Sub-Processor Register.

4.1 What is Sent

• The complete survey report PDF in full — all text, images, and embedded content
• Extracted field values returned for confidence-scoring

4.2 What is Never Sent

• SCORE account credentials — passwords, 2FA secrets, recovery codes
• Billing or financial data — bank details, sort codes, invoices, card data
• Other clients' surveys — each request contains only the document being processed
• Marketing, analytics or product-usage data
• Audit logs, login history or device telemetry

4.3 How Long Our AI Processors Keep It

Under the commercial agreements that govern both AI processors, survey content is processed transiently to produce the OCR / QC / extraction output and is not retained for model training. Short-term operational logs (used by either processor for abuse prevention and service integrity) are governed by their own DPA and the EU GDPR. We re-check both processors' published data-processing terms at each scheduled review of this notice.

4.4 Optional Feedback on Check Accuracy (Performance Improvement Only)

You can submit feedback on how well individual AIQC checks work for your account. This is entirely optional and can be enabled or disabled per-account.

What feedback data we collect:

• Which AIQC check was evaluated (e.g. "missing RICS condition field", "property address consistency")
• Whether the check was helpful or unhelpful in your assessment
• (Optionally) a brief note on why the check worked or didn't work
• Timestamp and account identifier (anonymised before analysis)

What we never collect:

• The survey content itself or any client data
• Names, addresses, or other personal data from reports
• Report text, field values, or extracted data
• Financial or sensitive information from jobs

How we use feedback:

• Identify which checks have high accuracy and which need refinement
• Spot patterns (e.g. "checks fail on certain property types")
• Prioritise improvements to underperforming checks
• Update check rules and model prompts

This is performance feedback for product improvement, not model training. Feedback data is never used to train the underlying AI models (AWS Bedrock or Mistral). Your participation helps us make AIQC more reliable for your specific use cases.

5. The Controls Behind the Commitments

6. Why Two AI LLM Processors — and What That Means for You

AI models evolve rapidly. A model that leads today can be overtaken in months. Using a single AI vendor would force us to choose between falling behind on accuracy and cost, or sending customer data to a new vendor every time a better model emerges.

We have engineered out that trade-off. SCORE uses a fixed pair of EU/UK-resident AI LLM processors — AWS Bedrock and Mistral AI — and we route each AIQC task (OCR, quality checking, field extraction) to whichever LLM gives the best result inside that pair. You get the benefit of continuous model improvements; your data never goes to a third vendor and never leaves UK/EU jurisdiction.

6.1 The Two Processors

Both AWS Bedrock and Mistral AI can perform all three AIQC tasks. We choose which to use based on current model performance:

• AWS Bedrock — managed LLM hosting in our existing AWS tenancy (UK London / EU Ireland). Gives us access to a curated catalogue of models (e.g. Anthropic Claude, Meta Llama, Mistral-on-Bedrock, Amazon Titan) without data ever leaving the AWS account where your S3 storage already lives.
• Mistral AI (Paris, France) — direct API to Mistral's LLM service for survey document processing.

6.2 What "We Optimise Models" Means in Practice

We test different Bedrock-hosted and Mistral models against real survey workload — accuracy on RICS fields, OCR performance on different report formats, latency, and cost per request. When we switch which LLM handles a given AIQC task, the change is always within Bedrock or Mistral — never to a new vendor and never outside UK/EU jurisdiction.

Model upgrades are routine engineering work for us; they do not change your sub-processor.

6.3 What Would Still Trigger Customer Notice

A change within Bedrock or Mistral (e.g. switching from one Bedrock-hosted model to another) is an internal optimisation and does not require advance notice — the processor, the data flow and the legal basis are unchanged.

Adding a new AI sub-processor outside this pair, or sending personal data to a region outside UK/EU, is a sub-processor change. We will not do that without:

• (a) a refreshed DPIA covering the new flow;
• (b) at least 30 days' written notice to customers under our Sub-Processor Register §7 process; and
• (c) the opportunity for customers to object on reasonable data-protection grounds.

7. What This Means if You Are a Surveying Firm

You are the data controller for your clients' survey data; SCORE is your data processor.

• You decide whether AIQC is on for your account.
• You decide what survey content is uploaded.
• We process that content only on your documented instructions, set out in our Data Processing Agreement.
• If you receive a data-subject request from a homeowner, vendor or purchaser whose data appears in a survey, we will help you respond — see our Data Subject Rights Procedure.

8. What This Means if Your Survey Is on SCORE

If you are a homeowner, purchaser or vendor whose details appear in a survey processed through AIQC: your surveyor is the data controller, not SCORE. Contact them first. If you don't hear back, you can escalate to compliance@scorehq.io.

9. Asking Us Anything

You also have the right to complain to the UK Information Commissioner's Office (ICO) at any time — see §14 of our Privacy Policy.

10. Review

This notice is reviewed at least every six months, and immediately whenever we add, change or remove an AI sub-processor. The next scheduled review is October 2026.

---

Document Information

• Document Reference: SCO-POL-AI
• Version: 0.2
• Status: Published
• Approved Date: 4 May 2026
• Last Review: 28 April 2026
• Next Review: 28 October 2026

Version History

• v0.2 (approved) (4 May 2026): Revised for clarity — §1 now explains AIQC is opt-in and tier-dependent; consolidated AWS Bedrock and Mistral AI references as "AI LLMs" throughout; clarified both LLMs perform all three AIQC tasks (OCR, QC, extraction); added emphasis that surveyors control what data is in reports; added optional feedback mechanism with detailed §4.4 on feedback data collection and §5 controls table row; rewrote §6 to consolidate processor strategy. Approved and published.
• v0.2 (29 April 2026): Reframed Bedrock from future evaluation to second active EU/UK AI sub-processor. AIQC now described as a fixed pair of processors (AWS Bedrock UK/EU + Mistral EU) inside which models can be optimised without changing vendor or jurisdiction. Updated §2, §4, §5 controls table and rewrote §6 around model flexibility within a fixed processor pair.
• v0.1 (28 April 2026): Initial draft. Public-facing AI transparency notice covering AIQC pipeline, Mistral sub-processor, and the six commitments (no training / EU only / encrypted / human-reviewed / opt-in / deletable).

220 Yards Limited (trading as SCORE) — registered in England and Wales.