AI Use Disclosure

The RICS Professional Standard on the responsible use of artificial intelligence came into force on 9 March 2026. It requires practitioners to be open about where AI has been used and to remain accountable for the output. Wren’s whole model depends on that accountability being real, so this page states the position in full rather than in a sentence at the bottom of a page.

A case-specific version of this disclosure, naming the actual version of national policy in force on the day and the timestamp of the last corpus check, can be generated at any time from your account and from a chartered professional’s portal.

• Drafting planning documents, from the facts supplied at intake, the constraints found on the site, and policy text retrieved from authoritative sources.
• Answering questions in Ask Wren and in the per-council assistant, from passages retrieved before the answer is written.
• Producing pre-application reports and short site summaries.
• Reading files you upload, to describe what they are and what they show, so a draft can refer to them.
• Judging whether a change to a monitored policy source is material enough to act on.
• Testing whether what a draft says about a policy is supported by that policy's own text.

This is the operative part of this disclosure.

• It may not recall a citation from memory. In 11 categories, including statute, national policy paragraphs, listed building entries, case law, flood zones, comparable applications, local plan policy, validation requirements, supplementary documents, conservation area appraisals and design guides, it must retrieve from a verified source or refuse to write.
• It may not characterise a policy it has not read. Confirming a policy number exists is not enough. Every characterisation is either a verbatim quotation from the retrieved text or is tested against that text.
• It may not fill a gap. A fact that is genuinely unknown is flagged for a person, never invented. A figure that cannot be produced is omitted, never replaced by a placeholder.
• It may not report a failed search as an empty one. “We searched and found nothing” and “we could not search” are different statements, and the second one is never written as the first.
• It may not treat your own documents as evidence of compliance. Facts read from files you upload are used to describe the proposal and the site. If one of them asserts policy compliance, that assertion is not repeated as fact.
• It may not sign anything.

A citation that cannot be resolved against its source at the moment it is written is a hard failure. Generation stops, nothing is saved and nobody is charged. There is no fabrication fallback, and if a source is unreachable the citation is refused rather than guessed at.

A named chartered professional, regulated by one of RTPI, RICS, IHBC, CIWEM, CIEEM, CIHT, ICE, CIfA, IEMA, LI, AA, ICF, IOA, IAQM, reads every document in full, revises it as they see fit, and signs it under their own membership number with their own professional indemnity insurance behind it.

Their signature is taken over the exact bytes of the file they uploaded, and signing is refused until they have uploaded their own version. What they attest to is set out at how a signature works.

Nothing is submitted to a local planning authority until you have approved the pack yourself, bound to the exact document versions you were shown.

For every document produced on a paid case:

• Every citation attempted, whether it was verified, the source consulted, the address it was resolved against, the reason if it failed, and the time.
• The passage that was actually retrieved, stored with its own hash, so a citation can be re-verified later against the exact text that was read that day.
• The model used and its version.
• Two confidence figures per section: how sound the reasoning is, and how complete the underlying data is.
• Every query raised for the reviewer, classified by whether it is a judgement, a missing fact or a missing reference.
• The signature record: who signed, their body, their membership number, the checksum of the file, and when.

When a source changes its wording, a new version is written and the previous one is demoted rather than overwritten, so the record of what a policy said on the day your document was drafted survives the policy being rewritten.

Every question you ask and every answer you receive is written to an append-only log on your account, which nothing in the product can edit after the fact. You can export it as a spreadsheet or as JSON at any time from your settings, along with a plain-text AI Use Disclosure naming the policy basis in force on the day.

Wren uses Anthropic’s Claude models, chosen per task rather than per feature, under commercial terms that prohibit the use of your inputs or outputs to train them. Text is turned into vectors for searching by Voyage AI. Neither is a source of facts, and neither appears in our data register for that reason.

The instruction set the drafter works under is not secret in substance: never invent a fact about the site; use a fact if it has been given rather than writing a placeholder for it; flag a genuinely unknown fact; cite by exact reference and never invent a policy number; draft local policy from the retrieved text and not from what such a policy usually says; state, in the same sentence as any policy citation, the requirement it is cited for; plain precise English with no marketing language; and never address the reviewer inside a section a council will read.

These controls make fabricated citations very difficult. They do not make a document correct, and they are not a substitute for the professional judgement of the person who signs it. That is why a person signs it.

Research outputs bought with credits, including Ask Wren answers and research drafts, are produced under the same verification rules but carry nobody’s signature, are marked on every page as not for submission, and are not professional advice.

Where our own data is incomplete, that is published rather than smoothed over. See coverage and gaps.