Coming to your point on the differentiation between doing it manually and doing it through AI, we recently would typically see someone doing it anywhere between 4 and 6 minutes per product for tagging etc. Generally, it is outsourced to someone who may not understand a lot about fashion, about styles, about the occasion and it gets subjective so someone needs to review it and go from there. Generally, we’re seeing that as you increase the scale, there is a level of inaccuracy that seeps in which makes it difficult to replace but when you use automation of AI or any kind of product like the one that Okkular is offering, we did 100,000 products recently for a client within 48 hours and we worked out that if they had to do it manually it would have taken them about 6 to 8 months to do that level of work. Of course, with multiple people thrown at it as well. And really the response from the customer was ‘if we had to do it manually, we just would not do it, we would just put it out there which goes back to that cycle of you saying ‘it is arduous, it is painful’ and if that last leg is not done properly, retailers just put it out there hoping that customers will stumble across the product. You would never do that for your physical store, you would not just put clothes in different categories and different areas and hope that customers will then pick up those clothes and see what they want. So why would it be any different for the online store and why would the result be any different? The conversion rates will subsequently be lower and lower.
In response to your question around the AI part, I think it is good to just reflect on why has AI become so topical? I think there’s a lot of different verbiage about predictive analytics but everything can become noise if your fundamental details aren’t known. If you do not know any level of predictive analytics anything that is slightly more out there in terms of the retailer’s mind is difficult to do because you don’t have that foundation layer done appropriately. So that’s why we believe that needs to be the first layer fixed because there’s been a lot of inaccurate information there, you need to remove that and have the right set of information and apply further enhancements so that analytics can be done.
Ben: You are right, I mean the volumes there speak for themselves right? There was a report that came out by Wrapped Media recently and it talked a little bit about the marketers dream around the future of content and we hear a lot about headless and that’s based around content and trying to blend that with product, but then if you don’t have those attributes at a product level, it makes it very hard to tag certain articles with certain product attributes but then you don’t have the content of the data that sits behind the product and those volumes that you can turn a hundred thousand hours into 48 hours is incredible. Can you tell me a little bit about the accuracy of the model? Does it take training and how as a retailer in that fashion space, how would they go around looking at Okkular from a reliability perspective or content production perspective?
Abhi: So what we have typically is for certain categories it’s out of the box, it’s already been trained it’s already been say if you take the apparel category as an example, that category we’ll retrain it for dresses, tops and anything within that apparel space so the accuracy is anywhere between 80% to 85% which is an acceptable level for most retailers. Where the subjectivity comes in is some might have a different occasion or style, some of that is a bit more nuanced to the actual retailer that’s where a bit more training is done to make it customised for that particular retailer. The way our product has been built is they can start off fairly quickly using the actual product and then with every correction that they make it gives out attributes and it will say it’s blue, it’s sleeveless, it’s v-neck etc but they can audit it and say ‘we don’t call it a v-neck we think this is a square neck from the way we have designed it.’ Then it starts learning from those corrections that a retailer makes or the audit that the retailer does. They don’t have to audit every single item. Most of what we have seen in practice is that they would do it for the first month, get to an acceptable level and work towards what we call a confidence score so then when we look and can say if the confidence score is 80% then the system automatically accepts it and you push it back into the PIM however if it is not then you show us and we can reviewing things which we call level one attributes where we know if it is a white or a black dress there is no subjectivity there so for those ones you know it’s far higher than the ones where there is a bit more subjectivity that can change depending on the retailer.
IT has been quite seamless to integrate and work mainly because we built it to ensure that if it took up too much time for them to get that output they’ll say we’d rather just do it manually and even though we know sometimes that thinking might not work, so far the retailers we have been working with have been able to do it quickly now and they’ve got it as part of their process so the behaviour has also changed and that’s important.
Ben: Definitely behaviour is always the hardest thing to change.
We’ve talked a little bit about fashion and apparel and you talked about some other industries, what industries do you see as being key for this kind of technology?
Abhi: So the way we have built our product is that it can learn different images it can learn different product categories etc but we are also moving in two ways. One is different features within the same categories that we are operating in for example the platform sees the image, it generates the tags and the attributes, but we are also generating all tags, we’re also generating synonyms. An example of this would be if the search word we have generated is polka dot, we are automatically generating the backend to identify ‘dot’, ‘spots’ and that way while the front end might have only ‘polka dot’ but if a user comes and searches for not that exact phrase then the product will still show up. So we look at it from the perspective of what can we do to increase product discovery on the site and increasing the metadata becomes crucial so that is where we are doing other things within the same category but talking about other categories apart from the ones that you mentioned, we look at it and we say wherever visual can play a role, wherever visual is important and it’s fast moving, it becomes an area of opportunity for us.
Ben: Yeah, and I guess that brings me along to my last question really. You know a lot of retailers are sitting there on day one saying ‘we’ve got this problem at hand’ and so how do they get started on this journey, what would be a quick win for them?
Abhi: I think I’m biased in this scenario but I would say that with most retailers that I’ve worked with I’ve seen that getting that foundation correct before they start thinking of the shiny toy is important and that key foundation is that you need to get your product attributes, your filters, you search items, those need to be put in place before it becomes too far down the road and too hard. At that point it becomes acceptable because this is the way it’s always been done. That complacency should not fit in so realistically, we don’t see it’s a technology issue because the technology is there, it’s a cultural issue. Sometimes the intent is that if we’re using technology, we want to see the outcome tomorrow morning we use it tonight and we want to see an ROI on this tomorrow and realistically that doesn’t quite fit in. There is a bit of a learning curve in anything. AI and automation or anything that we use should be treated as enablers or tools and of course, if an individual could curate every single product description, every single attribute and deliver it to their customer they would do that but the challenge becomes doing that at scale, and I think that’s what retailers should keep in mind, that having the ability to think of that ROI might be different on this and we should not expect an immediate return within a very short period. I think that will be good but the easy one would be you know if they’re using the Comestri PIM and being able to ensure that the baseline foundation, the basics are there before they further any of the opportunity that being able to be on marketplaces provides.
Ben: Yeah valuable points there I mean you know we’re all about owning more of the search space whether you’re in marketplaces or on your own ecommerce site and that’s got to start with the product data that you have in hand and the more beneficial it can be in the more product search space you own.
Always a pleasure to talk to you Abhi, hopefully, we will catch up again soon in person.
Abhi: Thanks Ben and the team at Comestri for putting this together, I think it’s been great to be able to share some insights and be able to help out.