/ AI in construction and property

A verbal site instruction, turned into a priced variation.

Small variations get absorbed because nobody could evidence them. Not because the work was not done, and not because the client disputed it, but because the instruction arrived verbally on a Tuesday and was reconstructed from memory three weeks later.

A specialist fit out contractor

Under three million pounds in turnover, running six to ten concurrent commercial refurbishments.

The problem

Instructions arrived on site by conversation, message or email. They were recorded inconsistently and priced late, so small variations that were genuinely carried out went unrecorded and unrecovered.

After

An instruction is captured at the moment it arrives, with its source, site, date and who gave it. The commercial lead prices it and decides what becomes a variation. The record exists before the argument does.

Recovery of variations that went unrecordedRead the full engagement

A delivered Orchyn build. Clients are anonymised at their request, so the sector, scale and measured outcome are real and the name is withheld.

Where the time usually goes

Instructions given verbally on site

The most expensive words on a project are often spoken, unrecorded, in a corridor. Reconstruction weeks later is a memory exercise with money attached.

Small variations quietly absorbed

Each one is too small to argue about without evidence, and collectively they are the margin.

Evidence assembled only when disputed

By the time a variation is challenged, the people who were there have moved to another site.

Maintenance and defect reports duplicated

For a managing agent overseeing around 480 units, triage that escalates genuine danger immediately and deduplicates the rest returned around eleven hours a week.

Where the system stops

It does not price the work, decide what constitutes a variation, agree anything with a client or issue an instruction. It captures and structures what was said, when and by whom. A system cannot retrospectively create evidence nobody recorded, which is precisely why capture at the moment of instruction is the whole intervention.

This is the house position across every Orchyn build: AI assisted, never AI executed. A model is used where judgement over messy input is needed. Every figure is computed by deterministic code against a named source, and a person holds authority over anything with commercial or legal consequence.

Common questions

01What is the honest benefit here, time or money?+

Money, and we say so in the case study rather than claiming an hours figure. The primary benefit is commercial capture rather than time saved: small variations that were absorbed because nobody could evidence them become recoverable when the record exists at the time of the instruction. That is harder to put a single number on than hours, and we would rather describe it accurately than invent a tidy figure.

02Will it work if our site teams do not use it consistently?+

Partially, and that is a real risk worth naming. A system cannot retrospectively create evidence nobody recorded. The capture step has to be quick enough that a site manager actually uses it in the moment, which is a design constraint we take seriously rather than a training problem to push onto your team afterwards.

03Does it price the variation?+

No. It captures the instruction with its context and surfaces it. Your commercial lead prices it and decides whether it is a variation. Anything with contractual consequence stays with a person.

04Can it handle safety critical reports?+

In the property management build, escalation is one directional by design: the model can raise urgency, never lower it. An emergency category can be added by the system and removed only by a person. That accepts more false escalations in exchange for never suppressing a genuine gas, fire, flood or safeguarding report.