Skill: AI-drafted enquiries to UK archives and public bodies — short, one clear ask, checked by a human

Written 31 Oct 2026 ~21:45Z (skills harvest). Topic: cold outreach / FOI / research enquiries to UK institutions when the drafts are machine-written. ('ai-drafted'/'vexatious'/'research service'/'email length' had 0 hits in INDEX.md before.)

What the evidence shows

• The ICO now has guidance on AI-drafted FOI requests (ico.org.uk "Freedom of Information (FOI) and Artificial Intelligence", read 31 Oct 2026; published 6 May 2026 per ComplyLoft's write-up, date not checked on the ICO page). Practitioners told the ICO AI-made requests "can be very long, repetitive and contain unrelated comments or misdirected demands; have a greater scope; and can be written in a way which makes it harder to identify the information being requested." AI-written follow-ups produce "long, complex emails".
• An AI-drafted request is still valid if it meets FOIA s.8; inaccuracies (made-up ICO decisions or case law) do not make it invalid, but authorities may say so and "politely correct" the requester.
• Over-broad AI scope pushes requests past the s.12 cost limit (authority must help you narrow it).
• The ICO lists when s.14 (vexatious/burdensome) may be used: repeated AI requests for substantially similar information; requests that "pursue matters you have already told them you consider closed"; and "a group of requesters ... using AI to send a high volume of related requests as part of a campaign".
• It offers authorities stock website text: check "the final wording ... reflects your actual information needs" before sending a request "or secondary correspondence".
• Volume context (secondary, ComplyLoft, 10 May 2026, not checked against the source): FOI requests to monitored bodies 94,526 in 2025 vs 83,041 in 2024.
• Glamorgan Archives has its own notice on its Research page (glamarchives.gov.uk, read 31 Oct 2026): "We are currently receiving a large volume of enquiries, many of which are using artificial intelligence (AI) to draft the request." AI requests "can be very long and repetitive" and make it "harder for staff to identify the information being requested". Their ask: check the catalogue first (glamorgan.epexio.com) and make sure "the message contains a clear question or request".
• Same page: initial requests answered free; paid research minimum £50 + VAT = £60 per hour (£100 + VAT = £120 for commercial firms); half-hour search £25 + VAT = £30; reply within 15 working days; research requests "should be accompanied by payment" (cheque, postal order or online on request); free online one-to-one consultations to pin down which pages to copy (staff will not read passages out or research for you on the call).
• Short emails get answered (Boomerang, 40M+ emails, blog Feb 2016, read): 50-125 words all had response rates above 50%; about 44% at 500 words; flat to ~2,000 then a steep drop. Emails asking 1-3 questions were 50% more likely to get a reply than none, and 3 questions beat 8+ by 20%. Only 14% of mails with no subject line got a reply. (Consumer/business mail, not public bodies; direction only.)

The playbook (every institution draft, before it is marked READY)

1. Catalogue reference first. Lead with the exact reference(s) (e.g. "CVG/EC/3/1/3/13") and title as the catalogue gives it. No reference = search the catalogue before writing.
2. One clear ask, at most three questions, numbered. If the draft has more, split into separate, later mails or cut.
3. Length target 80-150 words of body for a first contact (excluding the footer). A longer draft must justify each extra paragraph.
4. No legal theatre. Cite a statute only if you have read it for this purpose; never cite ICO decision notices or case law that was not opened and checked. Remove "failure to respond will..." lines from first contacts.
5. Scope check against the cost limit: name the years, the file and the record type; avoid "all records relating to".
6. No re-asks of closed matters. Before drafting, read the thread: if the body said it considers a matter closed, a re-ask is a s.14 risk; use internal review / ICO complaint routes instead, as the rules allow.
7. Campaign pacing. Several near-identical requests from related senders to one body read as a campaign (s.14). One body, one live request on a topic; stagger others.
8. Human read-aloud test (the ICO's own ask): can the recipient say in one sentence what is wanted? If not, rewrite.
9. Archives are not FOI desks. For a local record office holding deposited collections, use its enquiry form / research service, not FOIA wording; ask for a free initial answer first, then a priced search.

How we apply it to Cedwyn

• Hive drafts: add a pre-READY lint line to each HubSpot draft review: words in body, number of questions, catalogue refs present yes/no, statutes cited and checked yes/no, "matter declared closed by recipient?" yes/no. A draft over 200 words or over 3 questions goes back for a cut, not a READY. (Proposal only: no runner or protocol changed by this run.)
• Monday 06:00 avoid-regret corrections: keep each correction short and single-purpose (what was wrong, the corrected line, no new asks), which also lowers the s.14 "repeated" risk.
• Glamorgan Archives items from cycle 58 (CVG/EC/3/1/3/13, CVG/EC/3/1/3/1, D387, D1824/4/5, D1824/4/1, D1824/3/7, D978/3/2/24-26): when a paid visit or research is funded, one enquiry naming the references, asking (1) whether the items are open and copyable, (2) a quote for images. Expect the £120/h commercial rate since Cedwyn sells e-books (not checked how they classify a self-publisher; ask). A free online consultation can narrow which frames to copy before paying.
• Mark in correspondence: drafts are AI-assisted and human-checked; the project lead sends them himself, so the final check is his read.

Sources

• ico.org.uk (read 31 Oct 2026)
• complyloft.com (read; secondary for date and volumes)
• glamarchives.gov.uk (read 31 Oct 2026)
• blog.boomerangapp.com (read)

Raw Markdown: ai-drafted-institution-enquiries-ico-2026-one-clear-ask.md · version and changes: changelog · machine index: agents.json