Legacy Data, AI, and the Chasm of Risk: Reimagining Digital Content Management

Thank you very much, Daniel. I'm looking forward to it.

I think very early on, I realized that technology moves exceptionally quickly. And in the end, everything is legacy. You know, it requires a certain amount of strategic insight that I may or may not have had to be at the front of technology. But, you know, just a couple of the one example that I often like to use is I went to college.

I studied electronics for four years. I was understanding and developing radios myself before I went to college. And, you know, and I learned how what a p n junction was and how to dope silicon or germanium materials of which now nobody knows what what I'm even talking about with that. And then I get out of college, and what I've learned in terms of how to design amplifiers, for example, is redundant.

And you can now buy them from a small integrated circuit out of China for a dollar and twenty, and you quickly think, alright. I'm out of date. And I was out of date at twenty two years of age.

My next job was to work for a company that built the world's first 64K static RAM out of wafer fabrication. It took us maybe three or four months to perfect it. And just when we were perfecting it, the next one, one hundred twenty eight or two fifty six comes along. And then when I got into tape, it was the same thing. The tape cartridge that we were developing and building at the time was a two hundred megabyte tape cartridge, which was by industry leading as a square tape cartridge.

As soon as you're manufacturing it, the next generation comes through. So those three experiences really made me realize we're gonna be out of date very quickly. And I think it probably set the seeds for focus on development of an industry that is or a company that can be identified as the keepers of lost knowledge, so to speak, legacy data experts. And really, it was the fact that technology moves so fast that it kind of pushed me into thinking this is a good space to be because, you know, the other thought I'd leave you with is that, you know, what is legacy data?

Legacy data is actually modern data. And what I mean by that is you have a much better chance of picking up a piece of media, whatever form that is, that's two hundred years ago and reading it and understanding it than you do of picking something up that was created fifteen years ago. And the reason is the technology changes are so rapid and media the type of media format that this data is landed on is changing so rapidly. Different hardware, different software, different formats.

People forget how to do things fifteen years ago, but it is much more likely that they will remember how to read stuff or or restore stuff that might be two hundred years old. So legacy data is really, it's kind of like from one year old to thirty years old, and it's not an industry that's gonna go away.

Whether or not something is or isn't reasonably accessible is still is still a hot topic. And burdensome, it's a term that legal practitioners have often used when they don't want to encumber their client with the cost of doing something. But the reality is it's never been as burdensome as it's portrayed.

But the AI era, I have to say when, you know, we all started to hear much more about the involvement of machine learning than ultimately AI and we all started to, you know, particularly in my company, myself included, we're a little bit nervous about it. What does this mean? Is everything gonna go away? And gradually as we've become a little bit comfortable with it, we're now starting to get excited by it for exactly the point that you outlined in that AI is enabling us to solve some problems on large legacy data, no storage amounts that were never able to be tackled economically before, and it's now giving us a different angle.

So it's actually it's actually quite exciting. I think the burdensome argument is always gonna be there for sure. But at the same time, with every problem that comes along, a technologist will come up with a solution. In the end, that'll win.

I've got a couple of thoughts on that. The first thing I'd say is a more generalized statement is that when I got into this particular legal technology landscape, which was in the early two thousands, having been in purely data storage after that, the one thing that struck me as an uphill battle was records management, RIM, you know, records information management, the groups that would the technology based librarian types that would try to classify and categorize data at the time the data is created.

And then it would be organized and it would be controlled. And this is the way that we've got a handle on our data and this is the way that we manage it. I've always remembered thinking at the time that this in many ways, it's a battle that's only ever gonna be lost because the number of sources for new data is accelerating. It's just it's been increasing for decades and it's accelerating. How do you manage that? And then we've seen it with the bring your own device or the, you know, the company owned stuff and the phone and the now there's twenty seven different types of messaging apps.

How do you control that? How can your records manage control that to a point where it's got this legal defensibility feel to it? And of course, the answer is I don't think you can.

I think we can make a best effort at it, but I don't think you can. But we've been trying to control all data sources in organizations. With AI, it's kind of interesting because if you think about it from a dataset, one is control at the start.

The other is understand at the end. And I think what might be happening here is we're moving from an era of trying to control it to an era trying to understand it. That's a whole different landscape and a whole different set of architectural requirements and opportunities that it can bring. Know, if you give me a pool of massive amounts of data and say, what have I got here? Why do we bother needing to control and categorize and classify if I can do this at the end? And, of course, speed and time and cost is major drivers from that. But the AI is initiating, in my view, a shift from control to understand.

When you think about the off-site storage model, the highly regulated or many industries, all industries so to speak, but particularly the highly regulated industries have been working in for the last twenty, thirty years, there's off-site storage companies, massive companies, massive amounts of data pool, and they sit there and they don't move.

And they don't move because a, we don't know what's in them. If we don't know what's in them, how can we confidently get rid of them or how can we confidently access them? Or where is the data of interest anyway? If we don't know that, we can't touch it. And then the other thing is because there's a risk. You know, the off-site storage model is the stone that you don't wanna turn over because you don't know what's gonna be underneath it.

So from that perspective, it sits and it grows. And AI is an opportunity to tackle that because really when I think what we've been asked to do for the last twenty years is all risk based.

Can you delete this? Can you get rid of this? Can you identify it, and can you get rid of it? And when you think about it, the whole of the legal industry is largely in terms of the cost structure.

It's all about ruling out, not about ruling in. Everybody makes their money on the rule out process rather than the rule in process. We could go directly to something immediately, then we wouldn't have all of the massive extents of filtering, review, collection, etcetera, etcetera. We just go straight to the pertinent point, and that would be it.

Now I'll give you a kind of an example, a technology example that that that we've seen that involves the off-site storage model that we've obviously been part of, and it's a voice logging. It was a nice voice logging environment. And in that voice logging environment, the way that those voice logging systems work, the broker dealer type stuff where millions and millions of voice calls are recorded for, you know, transactions that have taken place, buying frozen orange juice type of stuff.

And the tracking of the voice call to creator of the voice call was gone. It was gone because of an encryption issue between the source system. So we have tens of millions of voice calls and no ability to understand who made the call and to basically recreate a database. But at the same time, the regulated industry requires that they must be maintained, they must be kept, and, you know, because they potentially are gonna be in breach of some compliance as a result.

What AI has enabled us to do mixed with legacy technology, so to speak, is actually to go through the content and recreate the metadata from the content. So it's a case of once I recognize your voice, and then when I hear your voice the second time, I'll know it's you. And so you can go through that content, and you can recreate with defensibility to say, yes. This was the originator and other pieces of metadata that you can pick up in those transcripts.

You do the speech to text transcription. You pick up all sorts of other potential metadata and you can recreate the catalog, so to speak, of the voice logging system.

That could never have been tackled without AI. Not in a commercially viable way. Not without somebody sitting there and going through five hundred million hours of voice recordings.

That's probably a great example of how AI has has absolutely helped transform this particular records management challenge.

It's the same idea. What we've become is we merely become the curation layer, which is still required.

You know, the certain amount of intelligence about what's trying to be achieved because the content, you don't know what the objective is of going through the content. So our technology that we've had for years really just ends up being the curation layer, the piece that actually means that it can be somewhat pinpointed because there are significant disadvantages, I would say, the most in terms of implementation of AI with with legacy data. Probably the most single biggest limit issue is cost and speed. You'd think that, you know, something that's fully processed and indexed is a fast solution, but the reality of it is you need incredibly powerful processes because ninety nine percent of legacy data is junk or valueless.

Maybe not junk, but valueless. And and so if you're gonna go through that, then you've gotta go through it all to know that it's valueless. And AI requires a huge amount of power and a huge amount of processing power. And, typically, that has to take place in temporary storage.

It has to take place in RAM on whatever the computer system is. So you have to tackle these things saying, what is this file? Can you tell me about this file? Maybe a predetermined set of questions that can be applied to the AI model to learn more about what it is, but you're not gonna talk to that in its legacy format.

You're gonna have to talk to that actually on on RAM. And that means that it's expensive and not incredibly quick, but it works. But it does work. So and what we always know is that if something works, you can make it faster and cheaper.

It's a matter of how you go about it.

I hope so. So what we're dabbling right now in, you know, in full disclosure, we're dabbling around with our own technologies and applying different large language models to them to see what can be achieved.

It's exciting us, but there's always that element of somebody's got to give you money to do something. Otherwise, what are you doing it for?

So in terms of that, I definitely wouldn't say we're at that point.

But it does make perfect sense. I mean, we have if you think about how much is wrapped up in terms of cost and risk in storage vaults around the world, doesn't need to be there. It absolutely doesn't need to be there. So I believe that it's probably worth it. And as we've also already mentioned, you know, ninety nine percent of everything that we do is risk and remediation.

And I think if you ask the person in the street what's exciting about AI with legacy data, they'd say it's learning from the data that we have that we don't know we have and what can be achieved from it. You know? So I think I'm definitely acutely aware of that opportunity.

And I think whether or not it'll become in management's eyes a real, you know, have some real legs is whether or not we can demonstrate wins for the clients. And that's what we need to do is we need to demonstrate a win or two.

Okay. Sure. So Invenir, it's a Latin word. It means to find.

We actually had it for a fairly long period of time now.

And the stimulus for it was twenty years ago when I was at a previous company and Sean Strickler, cofounder of our company, he was there. It was actually his concept about twenty years ago.

It was solely tape at the time. The idea being is that when you want to go through a large amount, you know, that large box of unidentified tapes, want to know what's on there. How does a litigator or an investigator understand what's on there? And the traditional way to do it is restore it, land the data, and look at the data. That's expensive. So how do you make a pool of ten thousand tapes? How do you learn about what's on it economically?

And if you're gonna go the restore route where you land all the data, that's not economic. It's it's ridiculously expensive. So the idea being is that what if we could just scrape metadata from the files on the tape without landing the content. There's two pieces to Invenir, the platform that you interrogate.

But really, the smart thing is a program we have called Tracks. And what that does is it enables the scanning and extraction of metadata as it rolls through a tape. But not just a tape. It can be a disk.

It can be anything else. But at the time, it was a tape. And so the type of stuff that we pull off are things like backup sessions, policies that were you know, anything that we've learned over the years that's an interesting piece of information that might end up serving to locate particular data for a litigation or something like that.

Yeah. So think about it like this.

You create data, a Word document, for example, just for simplicity. It's on your computer. There's a backup server that is networked to your particular computer at your particular desk. And at the end of each day, it would suck that information from each of the network computers through the master media service, so to speak, and push it to tape.

And, therefore, at the end of the day, that's where you have the hourly backup, the daily backup, the weekly backup, the full backup, the monthly backup, etcetera, etcetera. And it just keeps pushing. But it is a sequential piece of data. And what that means is it's just it's just it's sequential.

It's just getting pushed to tape. So one of the problems or the biggest problem, I would say, with backups on tape, which gave it a bad rap, is that you cannot modify anything on there. If you're gonna create a new backup, you gotta start from scratch. Whereas, of course, if it's live on a disk, then you can go in and modify a file or delete a file and add another file, etcetera, etcetera.

Once it's on tape, it's there. So when it comes to discovery, it's difficult. So data on tape is a sequential replication of a block of data that you might have had on your own work computer that ultimately gets backed up. And there's servers that might be print servers, exchange servers, share file servers, etcetera.

They're all called backup sessions. So I might have a backup session of a file share, a backup session of an Exchange environment, a Lotus Notes environment, etcetera. So the data is separated by by sessions.

Okay. So I got an answer for you. We never should have got rid of the mainframe or we never should have moved away from the mainframe. So back in the days of the mainframe, and there's still many, many uses of mainframes out there, but it was, you know, you'd go to a bank and you'd put ten dollars in your bank account and it would immediately create and write to tape and you'd get a little DSN, a little dataset name that would be, okay, Erica's added ten dollars to her bank account and here's the record. And then every time your bank account changed, it would get rewritten, rewritten, rewritten, rewritten. So it's always live and there's no junk.

No junk. Nothing that you don't want. It was always a value. But then along came Windows and Unix for that matter, Unix and Windows, And they said, you know, I've got a much more efficient way of computing. And it's I'm gonna do it all locally. I'm gonna store it on a disk, and then I'm gonna dump it, otherwise called a backup, and then I'm gonna dump it to somewhere else.

And as soon as you move to that model, you lose control. And we wouldn't have a discovery market if we hadn't had that transition from mainframe slash a s four hundred where every single piece of data was fresh, real, with no junk. Once we moved to using Unix on Windows, so to speak, and a network network share, NAS or a SAM type of environment, and then just dump it to backup, that's why we're all in discovery now. And so that's the reason we shouldn't have gone away from the main frame.

But it is when you look at it, it's incredibly economic to just back up stuff at the end of the day. The cost per store per megabyte, just dump it all. Dump it all. Dump it all.

It's great. But it's when you need it's the egress side. When you wanna get it out, when you wanna understand it, that's where it gets incredibly complex. And the other side of it is the other party to blame, so to speak, is really the backup software companies in my view.

You know, the IT administrators that had increasingly increasing amounts of data to back up an archive on a daily basis. And they'd have a budget and they'd say, okay, I've got, you know, twenty terabytes of data to back up each day. And the software guys would say, I got a great way to do this. You don't know, we can spread it wide.

I can put it across ten drives. I can put a bit of the file here, a bit of the file there, a bit of the file there. And you know, so be it. The one hour backup window that you might have had gets used in incredibly efficient ways to move massive amounts of data off to a backup environment quickly.

But there's a reason it's called backup software and it became known as backup software and not backup and restore software. It's because all of the technology or the majority of the technology went into backing it up, not restoring it. And the whole reason why we're in business is because we've created technology that hopefully comes up with elegant ways to do the restoration side. But it is a little when you look at how it developed through the nineties and two thousands, it is a little dumb.

I'd probably go down all of the wrong rabbit holes that everybody else went down. I don't think I'm smart enough to know a better way. Classification and categorization and control of data at the source makes all the sense in the world. I don't think any of us knew or are likely to know what's coming around the following day.

It's almost like you back up data ten years ago and a new data privacy regulation comes out nine years ago. And all of a sudden, with all the best intentions in the world, I never knew there'd be a subject access record request when I created this data. And now there is, and now it's on legacy data. Tell me what I've done wrong.

I did it right at the time. I did it with all the best knowledge I had at the time. But privacy regulations, other compliance factors, they've destroyed IT administrators' best intentions.

I don't know if there's any way to forecast how you would do it without knowing how the end result is gonna be.

Answer is yes. And it absolutely comes back to, you know, the example I gave with the voice logging system. We're getting excited by it. You know, Invenir, as I was explaining it's or started to explain, it started on tape.

But how most clients now ask us to use it is actually on smartphones and laptops and PCs. So by extracting metadata from laptops, we have this automatic JSON creation file that then can go into Invenir and Trax is our engine that creates these JSON files, and therefore, we have this we're adding structure to largely, up to that point, unstructured population sets. What it can also be used for significantly is the retirement of legacy environments because a legacy environment, you know, it's a piece of hardware, it's people, it's utilities, it's tape libraries, etcetera. So if it's a

complete infrastructure, a SAN and NAS. What we advocate or what we like to suggest that clients should do, the cost of maintaining those environments becomes a problem unless they're creating new data. But if they're not creating new data because they've gone to Office three five or they've gone to some kind of more elegant modern solution, but regulation compliance and data privacy and legal holds have said, you know what? You can't get rid of that.

So they gotta keep the whole environment around.

And where the Invenir solution can work for that is it can say, okay. Just retire everything. Stop paying the support fees for all of this kind of stuff. Extract the metadata, and we'll mine it as and when we need it.

You don't touch the data at rest. I know. I admit it even though our business is based on it. Every time you touch tape, it's expensive.

So the idea being is that you provide solutions to people, to clients that don't require picking up a table, picking up a storage, and actually putting it in and going through it. That's kind of the approach. Now with AI, what we're trying to do with Invenir tracks is we're trying to interrogate files to get the story in a very, very economic way. So Invenir is just metadata.

In the future, what we see it as being useful is to be you know, just create that kinda highlighted text, so to speak, as as how it might work. So tell me about that file.

Tell me about that file. So imagine if, you know, right now you're just looking at the metadata, but behind that metadata, that Invenir, which we host on AWS, behind that, if there is a tracks engine with a tape library full of images of the data that is a question that sits there, does nothing at all until you want answers from it. Click on it. Tell me about that file.

At that point, it would invoke the AI portion in Trax. We'll call it Trax AI at that portion. And then you'll get a story towards that file, a story about that database, a story about that dataset that will be targeted as and when you need it and not before. And if you wanna go deeper and wider, great.

That's basically where I think there can be value. Nothing I can see happening without a great deal of intelligence or curation layer, so to speak, will be good. But the technology answers for AI into legacy data are are boundless. There's there's so many ways that it could be applied.