Colin Stokes, Managing Director adiuvo talks about development of AI products.
Everyone is talking about AI, multiple companies are selling AI products but a significant number of agents we speak to are considering building their own. Having developed our own both for internal use and sale we thought we would share our experience.
I don’t think its controversial to say that building an AI tool is relatively straight forward if those doing so understand the issues clients and the sector have and the resulting product solves one of those issue or improves one or many multiple processes.
For agents considering their own, of course they understand (better than anyone) what they are trying to achieve so no case of the Proptech “solution” failing at that hurdle, its then just (he says) a case of hiring or sourcing the right talent and then that talent having access to the right data.
The data is the key here and it’s the most difficult part of the process. Your Large Language Model or LLM needs to learn from the largest possible data set so the question you need to address is do you have access to that data? If agents are doing this in-house then that battle is half won because sitting in your CRM, or your staff’s knowledge bank is the information you need, you just need to extract it. In our case, developing a maintenance assistant, we had over 900,000 historic tickets and 15 years’ worth of scripts. This allowed us to have closed learning (my preference) and control over the data the AI could learn from and meant not having to risk open learning in the big bad world where apparently things on the internet can sometimes turn out to be not true.
So, you have a product which has learned to an initial suitable level from trustworthy data, you even have the correct guard-rails in place so it cannot be manipulated like a DPD bot and in testing it hallucinates at less than 2% you have a winner right? Well in our experience the process we have discussed here is, as we said, the easier part.
What’s the difficulty then? It is simply put, adoption and not from a costs point of view which perhaps in the past has been the main objection as if you have built an AI product properly with true company fit, it should absolutely allow cost or resource savings an elevation of service (preferably both) therefore what objections have we encountered?
When we introduced products t generated an almost immediate split into three; one third were wary of it and although remain open minded wanted to see it in use elsewhere or with other parties first, one third dismissed it and said, amongst other reasons, they didn’t foresee it being the norm or useful to them and the final third jumped straight in and volunteered straight away to test it.
Unsurprisingly(?) these differences of reactions were based on demographics; younger participants were overwhelmingly positive so pick your initial targets carefully as successful internal testing and in-house advocates are the key to success. Once you find those partners, again in our experience, the process becomes easier and takes on a life of its own and a virtuous circle of feedback and improvement begins.
For more from adiuvo, click here.
