00:00You've raised $100 million in the last couple of weeks.
00:02Your customers include the likes of Microsoft, Amazon Web Services, Cohere.
00:06And just broadly for our audience, what you do is data labeling for LLMs,
00:11but you are also working on the ground with enterprise in terms of actually adopting
00:16and putting into place these AI systems.
00:18Talk us about that proposition, how much demand you're seeing for that kind of solution.
00:22Yeah, so we're an AI software platform.
00:24I think simplicity, you think about it, we can bring data, structured and unstructured data together,
00:28bring it, build custom software on top of that, and then fine-tune models
00:32and actually statistically validate them.
00:34And I think the interesting kind of history and go-forward view of the business is
00:38historically we've served mostly the large language model players.
00:41That's going to continue to be a very large part of our business,
00:43but more and more we see the same demand set in the enterprise,
00:46which is large companies, let's say a bank, if they want to implement Gen AI,
00:50they need to make sure their data is clean and organized,
00:52they need to make sure the workflow design around it is good,
00:54and they need to statistically validate the model at the end.
00:57And so actually I see the journey the model builders have been on over the last couple of years
01:00being the same one the enterprise will in the years ahead.
01:03At a time where there are these kind of bubble concerns being voiced by some around the AI space,
01:08given some of the funding that we're seeing and some of the huge investments
01:11involving the likes of OpenAI and NVIDIA and AMD,
01:14when you look at the enterprise piece,
01:17and some of the research has suggested maybe companies are not able to actually adopt
01:22and embed this technology as effectively as maybe some had assumed,
01:25what are you seeing on the ground?
01:27Yeah, I think the interesting thing is that we are in the first or second inning of the enterprise,
01:31in my mind.
01:33There's a report that MIT published a couple of weeks ago
01:35that 5% of models make it to production at the moment, roughly.
01:395%.
01:395%.
01:40There's a report that Gartner came out with, I think a couple of weeks ago as well,
01:45that they think 40% of agentic AI projects will be scrapped by 2027
01:50because they don't have a clear pathway to production.
01:53So you take that context and you actually realize we're at an interesting inflection point
01:57where the models have gone from high school to PhD level in the past couple of years.
02:00The model improvement has been extraordinary.
02:03But the enterprise is really struggling to adopt that.
02:05And so the interesting thing in my mind is, with all this excitement,
02:08we're still in the first or second inning of actual realized value.
02:11What are the biggest impediments to that?
02:13Yeah, I think it's three things.
02:14The first is data.
02:16So 70% of the software in the world is over 20 years old.
02:20Most of that's heavily fragmented.
02:22It's very hard to do Gen AI on data structures that are fragmented and hard to pull together.
02:28I think the second one is there's not a clear definition of what a good outcome looks like
02:34in a Gen AI context in the enterprise.
02:36And I can walk through an example of this.
02:37But if you were a bank trying to generate a credit memo on 10,000 of your credits,
02:41and I asked you, does that memo generate a good 25-page outcome
02:46that looks like what one of your analysts would have generated?
02:49There's not really a statistically validated accepted way to do that yet.
02:52That's a lot of what we've spent the last decade doing.
02:54But the kind of evaluation frameworks and rubrics you do to actually assess Gen AI.
02:59And then I think the third one is actually just organizational and workflow design.
03:03The biggest takeaway for me in the MIT report is of the 5% that are successful,
03:08the main determinant of that is the initiatives are led by the line.
03:13They're led by the managers, not the IT department.
03:15So you have managerial ownership of them and a deep understanding of the workflows.
03:21And I think that's actually been the hardest shift in this is
03:23if you're a CEO today, you're looking around and you're saying,
03:26who do I bet my major Gen AI initiative I'm going to drive on?
03:29I have to find somebody who understands the technology, the process,
03:32the people in the organization.
03:33And that's been harder than people expected.
03:35I think people were hoping this would be a SaaS era
03:37where you could just push a button and it would all work.
03:39And it's not going to be that.
03:41It's going to require real change management.
03:42That's been harder than people expected.
03:43What is the priority for you and Invisible into 2026?
03:47Then you've raised this capital, $100 million.
03:49You're valued at a little over $2 billion right now.
03:51The priorities for next year?
03:53Yeah.
03:54So we have six major enterprise sectors.
03:55I think our model training for the LLM is going to continue to be a huge part of what we do.
03:59I think we view that as the next 10 years.
04:01It's going to continue to be enormous.
04:03The next 10 years?
04:03Really?
04:03I am a believer that actually reinforcing learning and human feedback, human data will be a huge part of model training for the next decade.
04:11And then on the enterprise side, we have six sectors, asset management, insurance, public sector, healthcare, sports, and food and beverage.
04:20And so a lot of our focus will be on those six sectors and continuing to refine our offerings there.
04:25We're also entering some pretty interesting new spaces.
04:28So, for example, contact centers.
04:29We now have our first couple customers.
04:30I think that's a really interesting one.
04:32Yeah.
04:32Because contact centers are one that if you thought about the ability to interact with a voice agent that will know all of your data, it's an area that a lot of consumers are pretty unhappy with.
04:43Okay.
04:43But it's also a pretty hard thing to implement well.
04:45And you've seen some pretty public examples of folks trying it.
04:48Like Klarna tried it and rolled it back.
04:51And so we think the unique combination of bringing data together, modified tuning, will be effective for us.
04:56Okay.
04:56We will watch for that.
04:57Matt, we've run out of time.
04:57Really appreciate your time now in the studio bright and early this morning.
04:59Matt Fitzpatrick, CEO of Invisible Technologies.
Be the first to comment