Colin Stokes, MD of adiuvo explains their AI development and why they decided to build their own.
I began my journey in property management around 20 years ago, then several years later founded adiuvo. In that short space of time the tech at our fingertips has changed remarkably from lugging a heavy brick of a mobile phone in one pocket and a folded-up spreadsheet in the other containing all necessary information in the other.
Next came improved CRM’s and integrations, a multitude of add-ons and the birth of platforms for maintenance and many other parts of the management process, smartphones and apps and now we are at the threshold of the biggest advance, change and possibly minefield of our working lives as AI promises the solution to everything.
As an out of hours property management service to over 350 clients and one million units throughout the UK ultimately the motto “build what you use” resonated with us. So we set about building an AI Maintenance assistant (Adi) to triage and data capture incoming calls
That building was no picnic, going from prototype to live production (I.e. real people using it) is never straightforward. With Adi it wasn’t just about getting the tech to work a couple of times, it had to be something we could trust to hold thousands of conversations with residents.
We take customer personal data seriously, so setting up a secure, multi-tenanted platform was the first step. Every client needed to be properly isolated, with no risk of data crossing between accounts. You might have seen OpenAI just announced EU data sovereignty, which makes life easier now, but we had to host on our own servers to ensure no data went to the US.
The biggest challenge though was around getting confident with the AI’s responses. The great thing with AI is it scales – you train it once and it can have as many conversations as you like, concurrently and indefinitely. The tricky thing about AI is it scales any bugs or issues will probably have a big impact.
We initially tested with humans but honestly it was hugely laborious. We had literally thousands of conversation branches to test (Adi was trained on 15 years of scripts procedures and historic tickets to make it an expert in reactive maintenance) and we couldn’t just test them just once, as the nature of GenAI is that each time it responds slightly or very differently (depending on the settings).
We decided to spend some time building another AI agent to hold the conversations, and then another AI agent to assess the quality of the responses. This enables us to hold hundreds of conversations and verify we were comfortable before going live. We actually used different LLMs for this. The irony of AI testing our AI is not lost on us.
Overall, it’s been a fascinating journey; Adi now takes about 1500 interactions a month and rising. It is an option for callers as forced adoption never works in our opinion and incoming calls are, at callers choice, transferred into an immediate AI chat, of which it can do hundreds concurrently. As the assistant speaks every language it also caters for this and any other demographic who prefers this form of communication. We are currently testing the voice version which we believe will bring uptake to a new level and that will be released once we are happy with latency and interruption recognition.
Buoyed by this we have also been working on a soon to be released Lease extraction tool named Eve which reads leases and compiles standard information along with the ability to question the document to quickly and efficiently compile required data. Early tests show it is accurate to 90% which is approximate to human rates.
If anyone would like a demo of Adi, or Eve or some advice on what’s really involved in building your own automations please do email me; [email protected]
