AI operations: fewer maintenance handoffs and automatic payment recognition
Dwelly's AI agents now carry maintenance tasks across tenant communication, property context, contractor coordination, and follow-up. During August, over 1,000 tasks ran through this system. Average human work fell from roughly three hours per task to 15–20 minutes.
We are also automating financial administration. More than 80% of payments are now recognized without any finance-administrator time spent on matching and reconciliation, saving the team another two to three hours each day. And with a few upcoming changes, we expect this metric to reach 98%.
Maintenance workflows
In a letting agency, a maintenance task can start with a leaking pipe, loss of heating, an electrical fault, a failed appliance, or a damaged lock. Resolving the case involves multiple steps: interpreting a tenant's description and images, checking the property record and landlord's instructions, assessing urgency, finding the right contractor, arranging access, handling changing availability, securing approval, keeping several parties informed, and finally, confirming that the work was completed.
Each case can span several people, systems, and days. Information may be missing, contractors may be unavailable, and some repairs need a landlord's decision before work can continue. At the previous operating baseline of around 30 human actions per task, 1,000 maintenance cases represent roughly 30,000 manual actions.
Our AI agents move each task forward by reading the request, checking the property record and landlord's instructions, coordinating with contractors, arranging access, following up, and keeping the relevant people informed. When the system cannot safely determine the next step, it asks our human agent for help and continues with the new decision or context.
In August, we rolled out major improvements to AI agents:
1,000+ maintenance tasks ran through Dwelly's AI agentic systems.
Five human escalations occurred per task on average.
Three minutes of human input were needed per escalation on average.
That adds up to roughly 15 minutes of human work per task, with a practical range of 15–20 minutes.
A conventional manual workflow requires around 30 human actions. Each takes six to seven minutes once reading, checking, coordination, system updates, and follow-up are included. The total comes to roughly three hours of human work.
Dwelly's agentic workflow reduces that human time by around 90%.
Human escalation
An escalation occurs when the AI reaches a point where it cannot choose the next action with sufficient confidence. It presents the case to a person, receives a decision or missing context, and resumes the workflow.
We use these moments to improve the product. For every escalation, we can examine the information the agent lacked, the tool or policy that blocked it, the judgment the case required, the time spent by the person, and whether the task continued successfully afterward.
Repeated, low-judgment escalations point to capabilities we can add. High-stakes and relationship-sensitive escalations help us define where people should remain involved. We optimize for the right escalation at the right moment, with enough context for a person to decide quickly.
Automatic payment recognition
Finance administrators previously handled every payment manually: reviewing the incoming payment, identifying the correct account or obligation, and reconciling the relevant records.
More than 80% of payments are now recognized automatically, so they require no time from the finance administrator to match and reconcile.
This saves them two to three hours of work each day. The finance team can direct its attention to payments with exceptions, ambiguities, or risk management considerations.
With an upcoming partnership, we plan to improve this to 98%.
Measuring completed work
We evaluate our agents through complete operating workflows. Human minutes per task and escalation frequency show how much work the system can handle on its own. Accuracy, reopenings, speed, and customer outcomes tell us whether it succeeded.
Because Dwelly deploys its platform inside the agencies we operate, we can see the full path from an initial request or payment to the final operating result. Each task gives us another opportunity to improve the system and deploy that improvement across the network.
We will keep reducing avoidable escalations, expanding the workflows our agents can complete, and reporting on how those improvements change practice.