- Published date:
- 27 October 2026
CEO & Co-Founder at Findable AI
Earlier this month, Sam, my assistant, handed a piece of work to Maria, who runs our office admin. The handover note was better than most I've seen: what it was, how to use it, what had been checked, and what to watch out for. It read like any good colleague's handover. Neither of them is human.
We now work alongside 18 AI agents. They are not tools we open. Each has a name, a Slack account, a defined role and access to the systems it needs. We message them and mention them in threads, just as we would a colleague. Our day job is building information, so watching agents hand work to each other has taught us a lot about something every building goes through many times: handovers.
Who does what
Our agents cover support, engineering, testing, sales and office admin. A customer problem can move from support to an engineering fix to an independent check, with people able to follow and step in at every stage. Since late August, our testing agent alone has filed around 150 bug reports.
Work moved into the open
The surprise is not what agents can do. It is how much of our work became visible. The request, the agent's attempt, a colleague's correction and the result sit in the same thread. Anyone can see what was handed over and why.
What we keep human
Nothing leaves the company without a person's approval: no email to a supplier, no booking, no order. Agents never pay or sign. Judgement stays with people too: which client to prioritise, when to walk away from a deal. Agents prepare, people decide.
What went wrong
Someone asked our office agent to make oat milk a recurring grocery order. It confidently confirmed, but recurring orders sat outside our normal order, so a delivery arrived containing a single carton. Two weeks later, it happened again.
The agent followed the instructions and still missed what we meant. Most of our mistakes follow that pattern. Agents act on the instructions and information they have. When our records are stale or our instructions incomplete, they are not just unhelpful. They are confidently wrong. The fix is rarely the agent. It is the playbook and the records.
Why this matters for property
Property already works this way. A building has dozens of colleagues over its life: owners, managing agents, contractors and consultants, each handing over to the next. The handover is where things break. In our work with property owners, few buildings' records survive every handover intact. Drawings, O&M manuals, fire strategies and test certificates end up split across inboxes, shared drives and old portals, and the gaps only surface when a buyer, lender or regulator asks.
Agents could take on much of the work of running a building: checking that inspections are booked and certified, flagging warranties before they expire, or assembling a due diligence pack. But they will inherit whatever records the last party left behind. Ask one whether the fire doors were inspected last year, and it can only answer as well as those records allow. Under the Building Safety Act's golden thread, a confident wrong answer about fire safety is not an efficiency gain. It is a liability.
Before you hire an AI agent
Prepare the data. Agree what information must be handed over at completion, at a change of managing agent and at sale, and keep it in one place.
Establish accountability. Give every type of record an owner, and make sure a person approves anything an agent sends or signs off.
Introduce it like a new starter. Give it a name, a narrow brief and somewhere visible to work. Start it on reading and summarising before it acts.
Then ask what it will be reading. Can you find it? Can you trust it? Can you trace every answer back to its source? The AI models are now more than smart enough, it's the context that matters.
The advantage will not go to whoever has the cleverest agent, but to whoever gives it the most reliable information.
Share article