AI admin agent for estate-agency paperwork. Stow Brothers · workflow measurement and pilot blueprint
We measured the admin, designed an agent around the constraint we found, and came back a year later to overturn our own conclusion when the constraint had moved.
172 hrs a month measured 5 workflows mapped Every action gated by a reply
Five offices running their admin through one shared inbox.
Stow Brothers is an independent North London estate agency running five offices off one shared CRM and one shared inbox. Five admin staff push every valuation, sale, exchange and price change through that single Outlook thread. Two earlier AI experiments had gone live and then been switched off, so the appetite for another pitch was low.
They asked for evidence before anything else, so we measured what the team actually does on screen, ran the technology against the CRM they use every day, and came back with a recommendation the admin team could work to on an ordinary week. The deliverables were a strategy report and a pilot specification. No agent was built.
- 01
Volumes and minutes measured per task across all five admin workflows, then iterated with the studio manager until the numbers stopped moving.
- 02
A direct read of the CRM’s published API, which set the design: drive the screens and put a human reply in front of every write.
- 03
A year later, a re-check of the same API found confirmed write endpoints and a published event catalogue, so the answer we had given the agency changed.
172 hours of admin a month, measured across five workflows.
Volumes and minutes came off the admin team’s own screens, and the scope call fell out of what they showed. The largest workflow was not the first candidate, because the hours easiest to reclaim are the deterministic ones, where a fixed letter goes to a known recipient and a single field changes. The heavy workflow carries subjective visual judgement at every step, so it waits until the approval loop has proved itself somewhere lower-risk.
-
Price reductions
The lowest volume of the five and the cleanest to automate: a fixed letter to a short recipient list, with one field to change.
First in scope.
-
Valuations
Confirm the booking, send the pre-visit pack, write the follow-up: a tight loop hanging off a calendar event.
In scope, end to end.
-
Memorandum of sale
Sale agreed triggers a chain of letters to buyers, sellers and both solicitors, off deterministic templates and mechanical lookups, at high volume.
In scope, compliance step held manual.
-
Exchanges
A final round of letters and a status change inside the CRM, which is exactly the irreversible action the approval step exists to gate.
In scope, portal step left manual.
-
Brochures
The time sink, and the one carrying subjective visual judgement in every step from image choice to vendor amends.
Sequenced after the first scope proves itself.
Inside the agent design.
- 01
Approval by reply, in the inbox they already use
The shared inbox is already where the work happens, so approval routes through it. Nobody has to learn a new tool: the agent works inside the inbox’s existing access controls, and every approval is stored as a threaded email carrying a sender and a timestamp. If a reply does not come, the action does not happen.
- 02
The CRM screens as the integration surface
With no write endpoints available at the time, the design drove the CRM through the same screens the admin team uses, and kept the browser and the model provider as settings that can be swapped. That is why the later re-check changed what a build would connect to without changing the design around it.
- 03
Two execution modes, to keep running cost predictable
A fast deterministic mode replays known steps, and a slower model-driven mode takes over only when a screen changes or a step fails, writing a new fast path for next time so the system gets cheaper the longer it runs.
- 04
A hard stop on compliance steps
Mailbox identity scoped to a bounded set, outbound email locked to the agency’s own domain, and no progress past an anti-money-laundering or proof-of-ownership check without a manager-level human action.
What changed when we re-checked the CRM API.
The initial review concluded that the CRM was effectively read-only, which held for the legacy export feed but not for the modern integration API the vendor had since put in front of it. When the agency reopened the conversation a year on, we read the vendor’s current documentation again, found confirmed write endpoints and a published event catalogue, and wrote the correction down for the client in plain terms. That changes the shape of any future build, because filing a letter or attaching a compliance document becomes supported integration work rather than something driven through the screens.
The initial review
With no write access, the agent drives the screens and every action waits on a human reply.
The re-check
The vendor had published write endpoints, so the same actions become supported integration work.
- 01
Four write endpoints confirmed in the vendor’s own documentation, covering file notes, documents and reference checks.
- 02
A connector built to expose those capabilities as tools an assistant can call, with the scaffolding validated end to end.
- 03
Nothing has run against live or sandbox data yet. The write schemas follow the documentation and stay unverified until the agency switches the integration on.
What the blueprint measured and specified.
of admin work measured across a typical month, five workflows, five admin staff, iterated with the studio manager until the numbers settled.
a month reclaimable inside the first scope, on the workflows where every step can sit behind a human reply.
tools the connector registers against the CRM’s modern API. Scaffolding validated; nothing has yet been run against live or sandbox data.
webhook event types the CRM publishes, which is what turns a future build from polling on a timer into reacting to what happened.
Specified in the blueprint.
- 01
A computer-use agent that works the CRM’s own screens
The API audit confirmed Vebra Alto has no write endpoints for the actions an admin actually does, so the blueprint has a Gemini-class agent drive the same screens a person uses, in a cloud browser that can be swapped between providers.
- 02
Its own Microsoft 365 mailbox identity
The agent works from its own mailbox under the agency’s access controls, with outbound email locked to the agency’s domain, so every action is attributable and containable.
- 03
Approval by reply on every action
Nothing sends, files or updates until a person answers yes. The gate is the design’s safety mechanism, and it is what lets the agent inherit the inbox’s existing permissions.
- 04
An image-classification pre-test
A contained live test in which a classifier picked a defensible best six photographs against the agency’s own visual rule, run before any build was recommended.
- 05
Letters first, brochures sequenced after
The pilot scope the Director and Studio Manager reviewed: start on the letters workflow, and take on brochure production once the first workflow has earned it.
The blueprint’s stack
- Computer-use AI agent (Gemini class)
- Cloud browser (provider-replaceable)
- Microsoft 365 mailbox identity
- Vebra Alto CRM (driven via screens)
- Image-classification pre-test
OpenKit certifications
- ISO 27001
- ISO 9001, UKAS-accredited
- Cyber Essentials
Controls on this project
- UK GDPR
- Microsoft 365 access controls
- Outbound domain-locked email
More of the work.
Find your first workflow.
We start with a conversation, audit where AI actually pays back, and build the first automation into how your team already works. We reply within one working day.
Start the conversation Embedded AI Lead