Latest / CXOTalk / AI-Enabled Software Development: AI Coding at a Global Insurer, with Blitzy | CXOTalk #917
Transcript
- 0:00Probabilistic code is you just vibe coding cute ideas and put
- 0:04on the website and you can get away with this if you are start
- 0:08up at coffee shop but if you are regulated industry you need
- 0:12deterministic code. Chris Tokashevsky is Group Chief
- 0:17Technology Information Officer of Vitality, A multinational
- 0:22insurer with 14,000 employees. He's at Blitzie's headquarters
- 0:28rethinking Vitality software process using autonomous AI.
- 0:32Our legacy systems were not scaling fast enough, so we
- 0:37embarked on modernization plan to scale them up faster and
- 0:42future proof the business. We see ourselves as a
- 0:45challengers in the UK market. We've got about 12% market share
- 0:49and we compete against large institutions that have been in
- 0:54business for like hundreds of years.
- 0:56So as a challenger, we are growing and we need flexibility.
- 1:02So opportunities are all over the place and our systems have
- 1:07to be flexible to capitalize on those opportunities.
- 1:10So we've realized that that flexibility has to manifest
- 1:17itself in the systems because when there's new market
- 1:20opportunity, we need to build products and services that
- 1:24actually that actually support those those market needs.
- 1:27So we've been modernizing our systems for quite a while.
- 1:30We are building event driven architecture that supports our
- 1:37product requirements. So as our R&D colleagues come up
- 1:43with new products and services, we want to launch in the product
- 1:46that we want to launch in the market, rebuilding micro
- 1:50services, which actually event driven services.
- 1:54And that's quite important from a customer perspective because
- 2:00historically it was OK to make a phone call to an insurance
- 2:04company and make a request and somebody would do something.
- 2:09Nowadays everything happens in real time.
- 2:12So in order to be able to service our customers in real
- 2:15time, you need event driven architecture and event driven
- 2:20services. And also when it comes to the
- 2:23agentic AI and AI war that we entering into event driven
- 2:30services and natural partners of agentic AI.
- 2:35So when you combine event driven services with agentic AI, now
- 2:40you've got a very powerful combination because they
- 2:44actually like amplifiers that AI is amplifying the event driven
- 2:49capabilities and the event driven capabilities can leverage
- 2:54the AI capabilities and ultimately we can service our
- 2:57our members better in that way. You're using autonomous
- 3:03development working with Blitzie.
- 3:06With Blitzie, we're doing number of things.
- 3:08Autonomous development is part of it.
- 3:11Because we are in the regulated industry, the autonomous
- 3:15development has to be augmented with other processes.
- 3:20So we call it AI, native software development life cycle
- 3:24and then autonomous development is just part of it.
- 3:27But actually development life cycle spends right across.
- 3:32I'm looking at what are the limiting factors.
- 3:35So what is slowing us down because what happens is AI is
- 3:39accelerating specific processes and it exposes are the
- 3:44bottlenecks. So it's it's a continuous
- 3:48refocusing of the team. Now we've automated this part of
- 3:52the business or this part of the Sdoc process.
- 3:57What limiting factors have been exposed?
- 4:00So for example, we were able to generate the code very fast.
- 4:04Now who's going to review that code?
- 4:07So if we generated to, let's say 100,000 lines of code, who is
- 4:11going to review 100,000 lines of code and probably not a human,
- 4:15We're going to need another AI to review that AIAI developed
- 4:20code and then have a human review that part.
- 4:24So it's a continuous process of optimizing the throughput.
- 4:28So I'm quite keen on optimizing the throughput because local
- 4:33efficiency does not translate into overall throughput.
- 4:38So you have to optimize the entire software developer
- 4:42delivery life cycle to achieve optimal throughput.
- 4:45So you're working with Blitzy and you have essentially
- 4:50reimagined that entire software development life cycle.
- 4:56That's exactly right. We asking ourselves what are the
- 5:00limiting factors in our software development life cycle process?
- 5:05How do we leverage AI in that process to our advantage so that
- 5:11we get we generate better code, higher quality at the faster
- 5:17velocity and that translates into throughput?
- 5:21Clearly this requires an end to end platform rather than
- 5:25specific point solutions. That's exactly right.
- 5:29You do need to have an integrated platform that
- 5:33automates and integrates the entire software development life
- 5:37cycle to achieve optimal throughput.
- 5:41Because if you've got fragmented process, that fragmented process
- 5:46is going to manifest itself in bottlenecks, which I call
- 5:50limiting factors. If you've got integrated process
- 5:53with integrated tooling, you can achieve optimal velocity,
- 5:57quality and throughput. Where does Blitzie fit into
- 6:01this? We started with Blitzie on
- 6:04reverse engineering of legacy code and we've learned very
- 6:09quickly that Blitzy is an excellent reverse engineering
- 6:12tool. But then we quickly learned that
- 6:16Blitzy can also generate a high quality deterministic code.
- 6:23And it is massively important the insurance business because
- 6:27many of AI companies can generate probabilistic code and
- 6:32probabilistic code you just vibe coding cute ideas and put on the
- 6:37website. And you can get away with this
- 6:40if you are start up at coffee shop or some other company.
- 6:44But if you are regulated industry, you need deterministic
- 6:48code and that's where Blitzy comes into it.
- 6:51That one we were able to reverse engineer our legacy systems and
- 6:57we are able to create and generate deterministic code and
- 7:02that translates into efficiency and throughput of the entire
- 7:08process. And of course, here again,
- 7:10you're operating within a regulatory framework and you're
- 7:15a very large organization, multinational organization.
- 7:19So that consistency is fundamental.
- 7:23Absolutely. So the quality of the code and
- 7:25the consistency is massively important to us because of the
- 7:29regulation and also because we've got to do the right thing
- 7:32for our customers. You know, when, when, when, when
- 7:35customers call in the insurance industry, they call with a
- 7:39specific problem. That problem has to be resolved
- 7:42correctly, not by a probabilistic scenario.
- 7:47You know, it's a very deterministic scenario.
- 7:49I've got a problem. This is a solution, and that's
- 7:52what Bitsy comes into it. There's lots of AI companies out
- 7:56there, but very few companies can actually generate
- 8:00deterministic code which is fit and proper for the insurance
- 8:04business. Why didn't you use traditional
- 8:07approaches to modernization such as hiring more engineers or
- 8:13adopting standard AI coding the systems?
- 8:16Why did you take this platform approach?
- 8:20We have been modernizing our systems progressively over the
- 8:24last couple of years. Every time we would build new
- 8:27product, we would then take a chunk of a heritage system and
- 8:32we would modernize it. So it was kind of a just in time
- 8:36modernization for specific new product.
- 8:39However, we believe that the whole world is accelerating and
- 8:43we have to accelerate our modernization process.
- 8:47So we believe that we can achieve acceleration of
- 8:52modernization with Blitzy. Chris, how does the platform
- 8:56support Vitality's internal validation rules, coding
- 9:00standards, governance and compliance requirements?
- 9:05We achieve it through context engineering and the context
- 9:09engineering is effectively ingestion of code base, millions
- 9:15of lines of code of of our systems, our engineering
- 9:19standards, our development standards, this number of
- 9:24documents and set of rules that get ingested into the context
- 9:30that help us to build higher quality code.
- 9:34So this context is massive and includes both technical and
- 9:40business requirements. That's right.
- 9:43So you've got the code, the code base, you've got rules,
- 9:49standards, and all of that put together generates higher
- 9:55quality code base. When you say higher quality, can
- 9:58you elaborate on that? We've gone to what we call a
- 10:02test driven development, and what it means is that that you
- 10:06define test cases before you actually write the code.
- 10:11And we found that when we generate code with Blitzie, with
- 10:14Test Driven Development, AI is naturally naturally aligned with
- 10:21that methodology and generates much higher quality of code.
- 10:26So that when we generate the code, not only we pass the test
- 10:30cases much easier but also the code quality is higher as well.
- 10:36Can you describe the results you've seen so far across
- 10:39autonomous work, completed engineering velocity,
- 10:43development cost timelines, and also adherence to your
- 10:48compliance requirements? We're looking at different areas
- 10:52of software development life cycle and what we found is that,
- 10:56for example, reverse engineering of legacy code, we are able to
- 11:01reverse engineer our code from the five times faster than we
- 11:06were able to do it manually by people.
- 11:11So AI helps us to accelerate our reverse engineering of legacy
- 11:15systems by five times. We haven't fully quantified end
- 11:19to end the entire cycle, but we've got pockets of efficiency
- 11:25at this point in time and now we are in the process of actually
- 11:29optimize the full delivery life cycle.
- 11:32So we're going from local efficiencies of steps in the
- 11:36process to the overall effectiveness of the end to end
- 11:40process. And that's where right now we
- 11:42see major improvements in different steps of the of the
- 11:46cycle. But now I want to optimize the
- 11:50whole throughput because local efficiency of each step in the
- 11:54process doesn't necessarily translate to the overall
- 11:58effectiveness. For example, if somebody can
- 12:01generate code quickly, maybe they can have extra coffee break
- 12:04because nowadays they've got a bit of extra time until the next
- 12:08process kicks in. Now I want to optimize the whole
- 12:11thing. So the throughput get
- 12:13accelerated not only specific steps but the entire process
- 12:17gets accelerated because we are shared value insurance business,
- 12:22we providing value to our customers.
- 12:24And the question is does the new product deliver value as
- 12:29expected by the customers And then the question is how fast
- 12:33can I deliver that better proposition to the customers.
- 12:37So then accelerating the software delivery life cycle
- 12:42gives me the faster response time to the market.
- 12:47So there's always this linkage then between what you're doing
- 12:51with your software development and the architecture of your
- 12:55approach and the outcomes that your customers will then
- 13:00experience. Absolutely.
- 13:02So the architecture has got direct impact on the product
- 13:06that we're delivering and the software development life cycle
- 13:10accelerates the process of from the inception through the
- 13:15engineering to the product delivery.
- 13:17Now, how does all of this change the relationship between product
- 13:23and engineering teams? At Vitality we operate in matrix
- 13:28organization and R&D, product and technology or engineering
- 13:35are actually one team. So we actually got a blended
- 13:39team with R&D people, with product people and software
- 13:42engineering people working together to accelerate the
- 13:46process. So we are able to generate quick
- 13:49prototypes for R&D so that we can conceptualize our ideas
- 13:53faster. Then through the with with use
- 13:58of of Blitzy and AI, we can then accelerate the delivery of the
- 14:02product and then of course launch in the market.
- 14:06It's almost like you've transposed DevOps into product
- 14:10and engineering. That's right.
- 14:12So the DevOps is an integral part of the software delivery
- 14:16life cycle and we connecting DevOps with product.
- 14:21So Product and DevOps are then integrated.
- 14:24Now, when requirements can move more directly into working
- 14:29software, what happens to the steps in between, such as
- 14:32specifications, handoffs and iterate iterations?
- 14:37We're using AI to help us crystallize and document our
- 14:41requirements. Those are then ingested into the
- 14:45context that we spoke about few minutes ago.
- 14:48And of course that ultimately translates into test cases and
- 14:52code generated and into testing. So product requirements,
- 14:57engineering, software engineering, testing and release
- 15:03is an integrated process. What changed for your
- 15:06engineering teams in terms of their roles, their skills, the
- 15:09workflows and how they spend their time?
- 15:12We are still at the early stage of the transformation, but I
- 15:17can, I can already see early signs of what kind of a what
- 15:21what we call a levelling up. So the skills get uplifted.
- 15:26So we've got examples where requirements engineers dive into
- 15:31actually code and actually they want to have a look at what the
- 15:34code looks like. And you know, so requirements
- 15:37engineers build requirements, they boot test cases and they
- 15:41actually have a look at the code.
- 15:43And even though they are not software engineers with AI,
- 15:46they've got the ability to do that.
- 15:48So it is an integrated process and you can see the uplift, the
- 15:52skills being uplifted and then software engineers look upstream
- 15:57at the requirements closer. So, so the teams, the teams are
- 16:02actually getting well integrated.
- 16:04I see a better handover and more integrated process or better
- 16:11integrated process which translates into greater velocity
- 16:15and effectiveness of the whole throughput.
- 16:18What was it like to shift from reviewing small incremental code
- 16:23changes to reviewing pull requests of 50 to 100,000 lines?
- 16:28This is exactly where we are right now and this is This is
- 16:32why I'm at right now at Blitzy head office here, because we
- 16:36actually analyse the situation, how do we optimize it?
- 16:40Because we find ourselves that now we can generate requirements
- 16:45rapidly, we can generate code rapidly, but we cannot consume
- 16:50that code fast enough. So the code review is now a next
- 16:55limiting factor. So now that's an exchange for
- 16:59the for the agentic AI. How do I accelerate code review
- 17:04to get to testing and of course release and, and launch of the
- 17:08product? So as we go through the process
- 17:11and we accelerate one part of the part of the SDOC process, we
- 17:16find other limiting factors and those shift, they shift
- 17:19downstream. Now they shift upstream.
- 17:21And you know, it's like all the whack, whack and mole game, but
- 17:25you whack this one, this one pops up, you whack this one, you
- 17:28and you Canada. You've got to optimise to
- 17:30achieve greater throughput through the whole process.
- 17:33This kind of development is very new, so you are analyzing each
- 17:41of the components as you go and optimizing along the way.
- 17:44That's how it sounds. That's exactly right.
- 17:46And as we optimize the process, we are measuring the velocity
- 17:52and the efficiency and the effectiveness of the whole
- 17:54process. That's right.
- 17:56It sounds like the measurements are also integral to what you're
- 17:59doing. Yes, because we want to
- 18:01evidence, you know we are consuming tokens.
- 18:05So the question is what do we get for those tokens, right.
- 18:08So we might accelerated specific activities, but if that activity
- 18:13doesn't translate to the overall effectiveness, it's a limited
- 18:17utility. So it's all about measuring the
- 18:20process throughout. And I have to admit we we
- 18:23haven't, we haven't completely instrumented the whole process.
- 18:27We the process of actually analysing and instrumenting the
- 18:30whole process. But that's exactly the objective
- 18:32to optimise the entire process. I have to imagine in a regulated
- 18:37industry such as insurance, your risk audit and compliance folks
- 18:42are going to be looking at this very, very carefully.
- 18:45Absolutely right. And you know, it's this is exact
- 18:47science and you know, us as technology professionals, we are
- 18:53equally motivated to get dried because it is a cutting edge
- 18:57brand new technology and we want to make sure that the code
- 19:01generated is as good or better to what we are, what we are able
- 19:07to do historically. This is a significant set of
- 19:11changes that goes right to the core of your business.
- 19:16Given that, what role did executive sponsorship and
- 19:20leadership play in driving adoption of this new approach to
- 19:26technology? It's massively important, I have
- 19:29to say that we've got a a full sponsorship and full support at
- 19:34all level of all the way from the board, through executive
- 19:38committee, through management. Everybody's equally motivated to
- 19:43get it right. We believe that AI is as
- 19:48important as the Internet. You know this is a game changer
- 19:52and I'm personally determined to be the fastest horse in the
- 19:55race. This is a strategic move for
- 20:00vitality. It is for us and we want to be
- 20:03the the leaders in the industry and that's I believe that
- 20:07whoever wins this race is going to outrun the competition and I
- 20:12intend Vitality to be the host, the the fastest horse in the
- 20:15race. Chris, any advice for CIOSCTOS
- 20:20who are listening to this and wondering how can I get started
- 20:24with this kind of AI driven autonomous development?
- 20:29I think first one has to decide why we doing this and why is it
- 20:34important to our organization and vitality because we are fast
- 20:40moving and agile organization. AI is massively important to us.
- 20:45You know it's for some companies might not be.
- 20:48So I think maybe it's not for everyone, but if anybody's got
- 20:52an ambition of being a agile, fast moving organization, AI is
- 20:59the is the obvious answer. OK.
- 21:01Chris Tokozewski, thank you so much for spending time with us.
- 21:06Thank you, Michael. Absolute pleasure.