/ Project description
Situation
A new core, and the question whether it holds.
MANAGBL.AI's phone assistant answers calls for property managers. It identifies the caller, classifies the request and creates a ticket for the team. The first version ran on several threads and had four known problems: concurrency, interruptions, testability and response time.
The team had therefore built a new event-driven runtime with an agentic core. Two questions were open. Does the rebuild really solve the problems? And how does the bot behave when real people call with real requests?
July 2026: Architecture review
Checked line by line.
- Derived 32 verifiable claims from the architecture docs and checked each against the code, with file and line references
- Measured turn-taking and interruption timings against findings from conversation science
- Compared three telephony platforms on whether they replace the standard layer without bending the domain logic
Result. Three of four core problems are solved in the delivered code. Response time remained open. We endorsed the migration, with a clear measurement gate before the first customer.
32
claims from the docs checked against code
3/4
core problems solved in the code
October 2026: Quality test
The way callers experience it.
- 169 test calls in four modules: simulated everyday requests, edge cases such as noise and spelling, security attacks and a human sample
- Matched every call against the ticket that reaches the property manager's team
- Measured response time across 645 responses and compared it with voicebots in production
Result. An interactive report with audio examples, 19 error classes and recommendations ordered by urgency. MANAGBL.AI accepted it and is optimizing its system on that basis.
170
test calls in four modules
19
error classes with recommendations
640
responses measured for latency
What we take away
A conversation test alone is not enough.
Some of the serious weaknesses never showed up in the conversation. They only appeared in the ticket that reaches the team. So we matched every test call against the result in the system.
The same goes for architecture. We judge the code, not the docs. If a number only exists as a projection, it gets measured before go-live.
Visit MANAGBL.AI

