Scoping an AI pilot for property video with a manual baseline and stop rules

The safest first AI application for an England property-video team is a reversible draft task, not an autonomous property or customer decision. A transcript draft can be compared with the current manual process while a human editor and property reviewer retain authority.

This is an editorial pilot specification dated 5 September 2026. No product was installed, no account was opened and no capability was tested. It does not claim that AI saves time, improves accuracy or is widely adopted.

What to take away

Separate the supplier statement from the proposed task

Supplier announcement: Blackmagic Design's current DaVinci Resolve page describes version 20 functions including AI-assisted script, subtitle and audio features. That first-party claim does not prove England workflow performance.

Editorial scenario: test transcript drafting from a short, specifically authorised listing walkthrough for a two-bedroom flat in Leeds. The output is never published automatically. The video editor owns comparison with the spoken audio, the property reviewer owns every factual statement, and the managing editor controls release.

No external forecast is used. The trial proceeds because a bounded task exists, not because a publisher predicts adoption.

Fix the manual baseline

Use the same clip and approved claim sheet for both routes. In the baseline, a human creates the transcript, checks names, amounts, addresses and qualifiers, then prepares captions. Record elapsed staff time, corrections by category, inaccessible output, reviewer intervention and final disposition. These are trial definitions, not promised results.

For illustration only, a three-minute listing walkthrough might take 45 minutes of manual transcription and checking, with corrections concentrated in street names and price qualifiers.

The AI route uses the same acceptance criteria. Log exact product, edition, version, device, settings, account, processing locations, operator prompts and any supplier support. A comparison is invalid if the manual route receives cleaner input or a different review standard.

Sample acceptance criteria for both routes:

Sample log template fields: date, operator, product and version, device, settings, clip identifier, elapsed manual time, elapsed AI time, corrections by category, reviewer intervention, final disposition.

Control source and rights

Limit the input to footage whose use, alteration and processing have been approved for this test. Keep live enquiry data, access codes, alarm details, family routines and unnecessary faces out. The Intellectual Property Office's digital-images notice explains why online availability is not enough to establish permission.

Create a source card for each property claim. A model may help locate a mismatch, but it cannot approve a transaction, planning, tenure, condition or investment statement. The property reviewer checks the original competent record.

Worked source card example for the Leeds listing walkthrough:

Privacy, consumer and security gates

The ICO's AI lawfulness guidance says distinct processing operations need their own purposes and lawful bases. It is under review after the Data (Use and Access) Act, so publication-day verification and named privacy review are required.

The CMA's 9 March 2026 AI-agent guidance addresses testing, monitoring and human oversight under UK consumer law. A transcript tool is not necessarily an agent, but its control questions remain useful. The CMA page does not certify the proposed product.

NCSC secure-deployment guidance calls for security evaluation, documented limitations, incident response and clarity about data use and storage. Ask the supplier for evidence tied to the chosen configuration rather than accepting a general badge.

Accessibility and stop conditions

Test caption timing, reading order, speaker identification and a manual correction route with suitable users and specialists. GOV.UK's accessible-format guidance discusses captions and alternative formats but does not prove conformance by a private tool.

Stop and recover:

Before you act

Common questions

What task should an England property-video team test first with AI?

The article proposes transcript drafting from a short, specifically authorised property clip. The output is never published automatically. The video editor compares it with the spoken audio, the property reviewer owns every factual statement and the managing editor controls release.

How should the manual baseline be recorded?

Use the same clip and approved claim sheet for both routes. In the baseline a human creates the transcript, checks names, amounts, addresses and qualifiers, then prepares captions. Record elapsed staff time, corrections by category, inaccessible output, reviewer intervention and final disposition.

What conditions should stop the pilot?

Stop if the system invents or drops a property fact, changes an amount, exposes personal data, uses an unapproved asset, produces an unusable caption file or prevents reproduction of the output. Recovery means withdrawing the draft, preserving logs, returning to the manual route and documenting the incident.