An AI operations layer for a property company
A UK residential property management company
A multi-agent operations layer woven into daily property work — resident messaging, contractor coordination, voice-based staff check-ins and photo-led inspection triage — with read-only analytics and human approval throughout.
24,676
resident messages handled
in production, as of 20 Apr 2026
1,313
jobs created from 20,000+ inspection photos
as of 20 Apr 2026
848
action items from 89 analysed voice check-ins
as of 20 Apr 2026
The challenge
The work this system had to absorb.
Property operations are full of repetitive, time-sensitive work — resident questions out of hours, faults to triage, inspections to turn into jobs, and staff check-ins that vanish into chat. The company needed to absorb that load without losing human control of the decisions that matter.
What we built
A system in production, not a slide deck.
A resident-facing operations agent over messaging that classifies enquiries, retrieves approved policy answers, collects photo evidence and creates jobs.
An internal voice agent that runs structured staff check-ins, then turns the transcripts into director-ready reports.
An inspection intelligence layer that validates room-check photos and turns them into structured maintenance or customer-service jobs.
A constrained analytics copilot that answers operational questions over a safe, read-only query layer.
Governance, in the code
The controls are written into the architecture.
This is the part most firms can’t show you. Governance here isn’t a policy document. It’s constraints the system enforces on itself, every time it runs.
Analytics is read-only by construction — a Postgres function forces read-only mode, so the database itself rejects any write the model attempts.
A cost-approval gate holds any landlord spend above £200 for a human decision; tenants never see cost figures.
Inspection prompts are written to avoid over-diagnosis — no mistaking shadows for damp, no treating untidiness as damage, no accusations about residents.
An action ledger records every tool action with explicit terminal states, including intentionally aborted stale messages; voice sessions store full transcripts.
Outcomes
What it has done.
Honest and labelled: dated where the figures are point-in-time, and described as controlled tests or early-stage where that is the truth.
Handled 24,676 resident messages and 5,164 logged AI actions in production (as of 20 April 2026).
Turned more than 20,000 inspection photos into 1,313 maintenance and operations jobs.
Produced 128 management reports and extracted 848 action items from voice-based staff check-ins.
How it’s engineered
The approach under the hood.
A multi-agent system using fast vision models for photo verification, stronger multimodal models for inspection triage, a real-time audio model for voice sessions, and larger reasoning models for analysis — each output schema-validated.
Frameworks this build aligns to
Standards we design around and prepare clients for. Not certifications we hold.
- UK GDPR
Keep reading
More systems we’ve built and governed.
Find out where your AI exposure sits.
We'll tell you plainly what's worth doing, what isn't, and what a board or regulator will expect to see. No pitch deck.
No obligation · no pitch.