Work / Case
How do you build an AI advisor that customers can trust?
By Nicola Sirchia. Published . 3 minute read.
The short answer
I proposed an AI product advisor on the day I watched a customer walk up and down a shelf without knowing which product would do his job. It went live with a first version that was narrow on purpose. People came to trust it because the company's own experts checked its answers before launch, and because we had set in advance when it must hand over to a person.
Where did the idea come from?
I was checking a shelf in a shop. Next to me a man kept walking left and right along it, picking up one product, putting it back, picking up another. He clearly had a job to do at home, and he had no idea which product would do it. The shelf had all the answers, and none of them was talking to him.
It is a problem across consumer goods and retail. There is so much choice that people cannot tell whether a product fits their case, whether it does far more than they need, or whether the price is right for the job.
How do you prepare an organisation for an idea like this?
I tend to see the finished solution and the route to it at the same moment, so I proposed it immediately. I shared it with my manager first and then let it travel through the organisation, well before anyone was asked for a decision. An idea of this kind needs time to be digested.
What was the first objection?
Technical feasibility. An advisor is only as good as its knowledge base, and the questions were fair: did we have the right data sources, and how would we make sure the advice was correct?
What settled it was a demonstration before launch, with every expert who could act as quality control in the room. The accuracy surprised all of us. Nobody had expected it to be that high.
What were the risks?
There were two, and I weighed them equally. The first was a wrong answer, with what that means for people's safety and for the brand's reputation. The second was delivering on time and on budget. I see them as two sides of the same coin.
So every answer passes safety guardrails. When a question involves safety equipment, the advice is to replace the item and never to repair it. We also defined in advance the cases in which the advisor stops and passes the request to a technical adviser, and what happens outside working hours.
What was the hardest part?
The technology. It took a great deal of work, weekends included, before it served the project and could draw on everything the company knows. The knowledge was already there. Making it usable by a machine was the real task.
A chatbot is quick to build. An advisor people can trust with a real repair is a very different undertaking.
What went into the first version, and what waited?
Version one
The right product for the job, and how to use it
We left voice conversation and where to buy out of the first release, on purpose.
Phase two
Show it a photo, or simply talk to it
Image recognition, voice on the website and on the phone line, where to buy, and a look at the background of the photo for other jobs it can help with.
How do we know it works?
We watch three signals together: the rating customers give each conversation, the share of requests resolved without a person, and overall satisfaction.
The conversations have also become a continuous source of knowledge about what customers are trying to do, and about where product information is unclear.
What is my advice to a company starting now?
Work with credible partners and choose secure solutions. Building software by describing it to an AI is the new trend, but you still need experts who know how to do it properly.
And keep the purpose in view. Customers stay loyal to a brand for as long as it keeps solving their problems. An advisor like this lets the brand do that at the shelf, at home, or on the phone at eleven at night.