Vera AI Fact Sheet

This fact sheet describes the AI features of PropDev OS’s assistant, Vera, in the structured form a chartered surveyor can drop into their own due-diligence file and risk register. It is provided to support RICS members in meeting their obligations under the RICS professional standard Responsible use of AI in surveying practice (1st edition, effective 9 March 2026). It is not a statement of RICS accreditation, compliance or endorsement — that standard binds the member, not this software. The member remains responsible for any output they rely on.

1. What kind of AI this is

• Type: a Large Language Model (LLM) — a general-purpose, probabilistic text model. It is not a bespoke valuation model, an automated valuation model (AVM), or a deterministic calculator.
• Provider: Anthropic (the Claude model family). See the sub-processors page.
• Models used: the agentic “Vera Agent” (Appraisal Autopilot) uses Claude Opus 4.8; the everyday read-only assistant uses smaller Claude models (Haiku / Sonnet). The specific model is recorded with every agentic build (see §4).

2. How Vera works (way of working)

Vera reads the project you have open (budgets, appraisal inputs, schedules) to answer questions in context. In Agent mode (Appraisal Autopilot), Vera can propose a complete development appraisal from a plain-English brief. Critically:

• The LLM does not compute the returns. Vera proposes inputs (unit mix, rents, build cost, financing, timeline). The IRR, NOI, cap value, residual land value, cashflow and profit are then calculated by PropDev OS’s own deterministic, auditable financial engines — the same engines used throughout the dashboard — not by the language model.
• Nothing is applied without your approval. Vera presents the whole proposal on a single review card showing every assumption, its basis, a confidence rating, and any values the system clamped. You approve once; the change is applied as a single step and can be undone in one action.
• Assumptions are sourced and flagged. Default figures are drawn from cited industry benchmarks and are fully editable; Vera labels anything it estimated.

3. Limitations and failure modes

• Hallucination / inaccuracy. Like all LLMs, Vera can produce confident but wrong or incomplete output, including plausible-looking but fabricated figures. Treat every number as a proposal to be checked against source data and market evidence.
• Assistive, not authoritative. Vera does not exercise professional judgement and its output is not a valuation, not investment advice, and not a Red Book opinion. It is a drafting and analysis aid.
• Assumption sensitivity. Appraisal outputs are highly sensitive to a few inputs (rents/sale values, build cost, yield/exit, finance). A small input error can move viability materially — review the flagged assumptions before relying on the result.
• Currency of data. Benchmark defaults reflect a point in time and are not a live market feed; verify against current evidence for the specific scheme and location.

4. Human accountability & explainability (RICS §4.2, §4.4)

• Named reliability sign-off. Each agentic build is confirmed via a “Reviewed — build & accept responsibility” action that records the signed-in reviewer’s name and the time of review.
• Provenance record. A per-build provenance entry is stored on the project — the model used, when it was built, who reviewed it, and the list of assumptions with their confidence and sources.
• Activity log. An AI-assisted entry is written to the project activity log on commit, so the build is visible in the audit trail.
• Disclosure on export. Reports exported from an AI-assisted appraisal carry a disclosure line identifying that AI-assisted figures are present, the model, and the named reviewer — ready to lift into your terms of engagement.

5. Data handling

• No training on your data. Your prompts and project data are not used to train the providers’ base models; API calls pass the no-training flags each provider supports.
• Tenancy isolation. Vera only reads data visible to your organisation under our row-level-security rules; there is no cross-tenant access.
• Retention. Chat history is retained for 1 year then automatically deleted. See the AI Transparency Notice and Privacy Policy for detail and deletion routes.

6. Bias

LLMs can reflect biases present in their training data, and benchmark defaults can embed assumptions that do not fit a particular site, tenure or community. Vera surfaces its assumptions for your scrutiny rather than presenting a single “answer”; apply your own judgement, especially on affordable housing, tenure mix and local market nuance.

7. Environmental note

Running large AI models consumes energy. We keep everyday assistance on smaller, lower-cost models and reserve the larger Opus model for the agentic build path, which limits unnecessary compute.

8. Liability and professional indemnity

PropDev OS is a software tool. It does not provide regulated valuation or surveying services and accepts no responsibility for professional decisions made using its output. Where a RICS member carries a Vera-assisted figure into regulated work, the member’s own duties (accountability, disclosure, explainability) and their firm’s professional indemnity arrangements apply. Confirm your PI cover treats AI-assisted work as you intend.

9. Reporting issues

Hallucinations, biased output, or other AI issues can be reported to privacy@propdev-os.com — we review reports as part of ongoing model monitoring. This page is printable for inclusion in a due-diligence file or risk register.