Andrea Read of Convex Insurance on balancing human expertise and AI in modern engineering

Well, welcome to another episode of Unstructured Unlocked. I'm your co-host, Tom Wild. Well, let me start by saying I want to welcome Parul Calgreen as our new co-host to Unstructured Unlocked. Parul, welcome, maybe... provide a little bit of background about yourself. We're so excited to have you as our new co-host here. Thank you very much, Tom. Really, really excited to be here podcasting with Tom and hoping that our Unstructure Unlocked will rival Pivot.

of Kara Swisher and Scott Galloway. So that's my ambition. I am a forward insurance executive and now advisor and really, really happy to be here. Great. I'm excited to have you as our co-host here. Pearl, we're really excited today. Another great guest. We've got Andrea Reid from Convex. Andrea is the head of technology engineering. And of course, technology engineering is at the center of...

all that is gen AI right now, the good, the bad, and what's to come. So very excited to have you, Andrea. Hi, thank you for having me. Excellent. So this is a great topic, and we were just talking before the show here that... This is something that changes seemingly week to week, not month to month or year to year anymore. Maybe start by describing your role at Convex, and then we can kind of work into this really interesting topic of to what degree is generative AI?

supercharging your role and your team's role in creating technology advantages, you know, in this very competitive insurance space. Okay. All right. So I look after all of the engineering teams, DevOps teams, integration teams at Convex and a lot of the products that we build here at Convex. So that means that... In my day-to-day, I deal with a lot of either AI products, AI builds, or just using AI to help us develop, which is a little bit about what we talked about.

the power of that. So there's many different aspects of Gen AI that I touch on in my role. So using Gen AI to help us with... Engineering and development at the moment is a big key topic and a lot of pressure that you have externally when you get Facebook making statements like 75% of our code is written by AI agents.

tried to use as much as we possibly can and we definitely do we have a co-pilot and we have chat enabled within our ides and we can use some of that ai generated code i would say Looking at the stats that I see, about 25% of our code is generated by AI, but that still needs to be reviewed and refined by expert developers. It definitely helps some of the more junior ones get started and some of the more senior ones do more easy tasks. So it's helped.

Over the last year, it's definitely increased productivity and I can see that and the output of the teams. But it's not really there yet in my world of that I'm swapping out developers for.

But I don't know if you kind of see the same thing in your world. Well, I think it is interesting that the pressure you mentioned is real. You know, as an earlier stage company, we get the same... view from from the board and from investors because they read the same press that everyone else does which you know i think and they're aware of this that you know there's clearly a vested interest from from the uh the

from the metas and from the AWSs and from the Anthropics to put out these breathtaking releases about the degree of which they've been able to produce code or reduce headcount. So there's... There's too much of a vested interest there to sort of believe it, you know, with blind faith. But no question it's having an impact. Pearl, you know, in your role, you're kind of consulting both to insurtechs and major insurers in your background there.

What have you seen in terms of how the enterprise is trying to understand this adoption curve and get benefits from it? You know, it's very interesting that Andrea has quoted the 25% statistics. Because when I'm advising to people in the boardroom, they have this perception that the biggest productivity boost will come in the engineering organization.

because they feel that the customer facing organizations still need the human touch and the processes cannot have human taken out of it so it's all all of the gen ai is basically boosting the productivity of an employee rather than face outwards. So the fact that your experience is that with the current state of development, It's still early days for developer productivity is super interesting. But the pressure, I hear it all the time because they're big numbers, big efficiency.

gains that people are expecting from this technology. Yeah, definitely. The pressure is real. But yeah, I don't know if you've seen any different, if you've seen any higher. statistic in your specific experience rather than the press but I live it across a lot of teams at the moment and obviously you've got an existing estate which is reasonably complicated and

Now, over the last six months, the AI has changed. So it has the context of the code that you're editing and it can produce new versions of that. Whereas six months ago, it obviously could not. So that has really helped, but it's still not kind of, yeah, that's 100% right all the time. Do you think that's the big challenge still to solve is while the context windows have grown, like you said, you can...

You can give it the code base, but is that still enough context or what's the missing context? I think, is it more sort of the overall architecture of the problem you're trying to solve? How has your team looked at that?

Yeah, it definitely is some element of that. We also do a lot of integration. So we have a lot of different... microservices and components that we fit together that's our IT strategy and architecture so it's knowing about those different components and how they all fit together in that. overall picture, which is hard for AI to grasp and know about. You're exactly right. Solving any insurance problem is multilayered.

And even with the large context window, there is so much data which comes from external providers, especially in specialty where Convex operates. and reinsurance where external data sets play such an important role in risk management and also claims processing. We've seen some progress is kind of writing those business requirements. Writing the stories as well, that definitely has helped productivity in the engineering space. And I think that will grow more.

It's almost easier and easier to review and easier to get on the right line. So that's definitely helping our teams. I think, again, change management around that is a little bit. tricky because and it is in both areas right in in the engineering and writing stories business analysis side of things that not everyone is receptive to it I've tried to different ways of

encouraging my teams to do it. But yeah, that can be difficult. In a previous episode, I sort of made the statement that in some ways, Gen AI has the opportunity to sort of answer the question of Why did we spend all this time and money on building data warehouses and data lake houses and all those things? Like, is this the moment where all that becomes justified because

Of course, using AI requires access to, as Pearl just mentioned, quality data. How has access to these technologies and these technologies changed your data strategy, if it has, in terms of the landscape of... structured and unstructured data, which, you know, insurance is just, well, it runs on data, right? I mean, at the end of the day, insurance companies are in the business of making good decisions, but those decisions run on data. How has that changed your

your approach on sort of the data side of the equation? So I think the unstructured side of the data equation that Gen AI can bring is going to enhance our data strategy. So we can now incorporate the two. We've got the structured data, but if we can also vectorise the unstructured data, all of the documents that come in, search and ask questions across that. And that is the exact...

kind of use case that we are working on now. So enhancing our data strategy to be able to add more data that we didn't have access to before.

I think that that's our kind of key goal for the next couple of months, really. So I have a question on developer experience, which is central to your work. Do you think your colleagues in underwriting and claims processing understand how developer productivity really delivers their goals as well as they own these claims and underwriting processes. Do you think they get that engineering team is really there to help them or?

Do they still do you feel culturally they still throw tickets around their shoulder and say, hey, solve this and come back? No, I think we have quite a close link to the business. You know, we we have run agile. projects we release every week we work quite closely alongside them and we're I like to think but obviously unbiased that we're there to help them and that yeah it's not that like ticket

ticket-based request and we kind of solve problems together, which seems to work well for us. Talk about maybe the overarching kind of human in the loop. opportunity and challenge here and and maybe connected to agentic ai right which is a whole nother both opportunity instead of risks around how do you know these are how do you know these agents are performant how do you

deal with future sort of governance and compliance requests? That is a very tricky question. How do you know that they're compliant and giving the right answers? Until we do, we have to have the human in the loop. We are, and you'll know this is part of like the... the data extraction that we we work with it indico on you know we can get those confidence levels in the data which again means that we can have the right answers but

Obviously, we've got to build that confidence up and we have to do checking, especially in the kind of compliance space. But we are getting there. We are getting users more confident in the results as we get better results. So this is probably one question you get asked quite a bit. So you've built a career as a woman in tech leading engineering organization, which is not very typical.

What lessons from that experience do you feel is most relevant to our industry as it embraces gen AI and AI-driven change? So you talked about change management and... also your experience with your business colleagues so how does how does is your experience different as a woman in tech versus is it improving because it's been over a period of time I guess It changes because AI is becoming more relevant, more people want to talk about AI and technology was...

Before technology was just something that sort of happened in the background that nobody was interested in. So it's become a little bit more high profile and people want to talk about it, which is great. I suppose as... It has been hard as a woman in technology because I am very much a minority and that can sometimes mean that it is difficult to be listened to, I think. But that is changing and I think...

As you've suggested, having that closer partnership with the business stakeholders has definitely given me an advantage as someone who leads that engineering team. What advice would you give? So in retrospect now, what advice would you give a sort of young engineer entering the field?

that you think is different than when you entered the field, given these realities. I mean, I think there's, again, back to the hype machine, you know, we've heard in the last year, everything from it's the end of the software engineer to... No, not really. So where do you think the pendulum is there? And what advice would you give a university grad thinking about coming into both the engineering role and the engineering role within the insurance space?

So my advice would be to know how to program any language. So nowadays in university, they don't just teach a language, they teach how to know.

what the future is, what is the different languages, understand Gen AI, understand how to control it, to take it forward. I disagree that there isn't a future in... in engineering i think there very much is but it will be different it will be different to the future that i had which is just developing there's so much more that you can do nowadays but there's still a lot and And people still have to look after the AI. There's never been a time in the history of technology where...

There hasn't been more people employed to manage the technology than there was previously doing the operational roles. So I see a future in that. that AI world and I think that university grads need to increase their knowledge and understanding of that but at the same time that the sort of basics of understanding operating systems understanding really how to work a you know work a computer what happens when it it goes wrong that's what they need to know like

AI can do a certain extent of that, but at the moment it can't do everything. So just really understanding those basics, but understanding AI as well would be the advice I have. And that is kind of pertinent to me.

daughter is doing computer science at university. So we talk about these things. Very relevant. Yes, very relevant. Andrea, tell me, I mean, I'll...

You must have heard Jensen Wong saying nobody needs to learn coding. And you have refuted that because you think that the innate knowledge of... architecture operating systems and languages is important even if you take a QA kind of a role do you think that QA is the quality audit is this kind of role that human will have in the future or do you still

believe it'll be more nuanced than that. So he may well be right, and I'm not refusing that, but I think a little bit of what the younger generation lack that we had in our generation is actually knowing. how the core of computing works because when when we first started you know we had to go into the operating system we had to control it if your computers go wrong now

Like our generation know how to fix it and how to restart it. Whereas the younger generation don't. And what if the AI, you know, what if AI goes wrong? What if I still think. Maybe not coding as a job, but there has to be some underlying knowledge of some of the basics for if and when it does go wrong.

As a vendor in this space, you know, a solution provider in this space, we think about this a lot now, which is there is a future where technical roles and technical leaders will be, the job description might be. you know manager of agents right that may be one of the jobs is you know you have a suite of agents that you know you're responsible for understanding their behavior their performance their accuracy and i think what

at least for now, we'll have to see how this changes in the future. You know, the human loop pieces is vital in that just like kind of a new hire you might make, you don't, you don't go sit them down and then check in with them six months later. Right. I mean, that, that would be. Exactly. Yeah. So it's a weekly process, right? It's weekly process. I think the metaphor holds here, right? Which is...

Right now, for agents, it's a daily process where you're checking its behavior, especially in the early days, to see if it's functioning. Because what AI still struggles with, and again, we don't know what the future holds, is... AI is really good at pattern matching within a sort of standard deviation from the mean, right? But when it's two standard deviations away, and I'm just using this metaphorically, not literally.

um it breaks right it's not good at high variability problems still and that's what humans are extraordinarily good at is is is being able to solve and generalize you know even with with high standard deviations to a problem that's presented to them. So, you know, as a solution provider, I think we think a lot about when we have a customer, what will that look like?

Because as Convex, you'll have multiple internal builds and solution providers, all of which eventually will have some kind of agentic functionality. So how do you have telemetry that is standardized enough that you can know internally? What's the state of all our agents today, right? Which ones are misbehaving? Which ones need attention? So that's kind of a, and then you have the whole compliance piece, right? As regulators will show up and say, hey, we're going to randomly sample, you know.

thousand transactions from last month. And you're going to have to be able to provide us with the sort of explainability around that. How is that affecting some of your architectural? It's very early days. I know I'm not suggesting anyone has this sold. QA is very much the future of agents and agentic AI. And I was talking to one of the QA managers this morning actually about this and she was saying, well, what do I do? I write an agent to test my agents.

You know, and it can go on and on in that loop. But at the end of the day, you've got to have somebody checking it and knowing that it's running correctly, right? This is very interesting because a lot of people that I speak to talk about hallucinations. Have you found a way to reduce the hallucinations of any deployment of Gen AI within your environment? Are there any strategies that you found?

particularly useful. This would be very interesting for our other insurance listeners who are grappling with similar problems. I haven't found any specific ways of reducing. the hallucination, really. I guess there's some strategies that you can employ and ways that you can configure the AI and obviously...

better training and more specific training on our data and vectorization and all of those things. But I can't think of anything specifically that... we can use to reduce hallucination or that we use i'm not aware of sorry it is kind of interesting right it's in many ways i've always felt that

hallucinations are actually a feature not a bug right the the strength of the generative the g in gen ai generative means that this technology was built to be good at generation it also is built to to be a pleaser right it is it is by default it seeks to give you an answer uh

It doesn't understand, and this is one of the major weaknesses, I think, of Gen AI that may never be solved, just architecturally. We may need a different architecture someday in the future. Because it doesn't have a understanding of the real world, it doesn't know that...

If it doesn't know the answer, it shouldn't try to give you an answer, which, you know, unless you're a very narcissistic person, you know that if someone asks you a question and you genuinely have no idea, you don't just generate an answer and say, how does that sound? Right.

and do it enthusiastically and tell you that it's a great question you know which is sort of the the gen ai experience you know even though you know you try i put in system prompts to say like don't be sycophantic right don't i don't care if you think my question's good and now i've found that

at least with some of the LLMs, it will somewhat sarcastically say to me, you know, okay, this is just a fact-based response. No, you know, no sycopancy. So it sort of seems to be almost passive aggressive about it. But yeah, I think... Insurance is maybe you could say FinServe, pharma. These are industries where deterministic outcomes are critical, right? So...

In a weird way, while insurance is in the business of uncertainty and trying to assess risk and make risk quantifiable, which we all know is like, you can't actually do it. You can approximate it. We're now using a technology that is stochastic by nature, and we're trying to apply a stochastic tool to a deterministic outcome, which makes the whole thing really fascinating to figure out.

Can we do that at scale? The number one challenge or complaint that I hear from CIOs is, look, it's a great technology, but the thing that we still can't quite figure out is how to... find either vendor solutions or internal builds, how to make them scale successfully, scale across lines of business, scale across use cases and insurance that would be underwriting claims, policy servicing.

And scale, you know, elastically to volumes, because obviously in insurance, you've got the renewal season where you might have a 10x spike in volumes. Is that...

How have you thought about the scale problem with these technologies, you know, in general, either with your vendor partners or with your internal builds? Is that resonate for you as well? Not really. I don't think we're kind of... ahead enough in the journey we are not you know so so we we are using ai for data extraction i use it in the engineering that we we talked about we are starting to look at agents for doing some

certain tasks but i wouldn't say we have nailed this done it have it in production where where now my problem is is scale that is not the place that we are in I think we will get there probably if we have this chat again in a couple of months. Today, that's a second order problem. Yeah, yeah. Yeah, yeah. So for us, it's a second. Right now, it's making sure that we...

Or trying to find the right use cases for actually using Gen AI and implementing agents to be able to do tasks effectively and have the right management of that. position that we are in right now, not the very nice problem to have is how do we scale that? Align to the scale question, which is a question which will come up. How much do you think is the culture of the organization and senior sponsorship and realistic view of what the technology can deliver?

important to get to that scale and what is your wish list i wish the ceo of the organization would understand this we at

Convex have very good sponsorship in the technology space from our CEO. He understands and knows technology. The strategy of Convex is to build and make decisions based on... data and technology so we have the sponsorship and the ability to be able to innovate both within my teams and the business teams and workers partnership so we definitely have that senior level sponsorship and capability to be able to do the things that we need and try

Like from my perspective, trying to kind of manage that and the day-to-day can be a little bit overwhelming sometimes and try to make the progress that you need. But it is certainly not through sponsorship. We definitely have a really good... level of sponsorship and enablement from everyone at Convex. I always like a good wrap-up question, and those are always good prediction-based, right? So if we were to do this again in a year...

What do you think we'll be talking about a year from now on this topic? Or maybe ask differently, you know, what do you hope we're talking about a year from now on this topic? So I think the future topics, so we talked, we touched on scale. I think we can come back to that.

Obviously, security is a big aspect of Gen AI. That's something that I'm conscious of and we're touching on it at the moment. But I think as it grows, that's going to... become more of a of a topic of conversation yeah for sure well great andrew it's been great talking to you up rule we've been talking to andrew reed a head of technology engineering at convex insurance

This has been another episode of Unstructured Unlocked. I'm your co-host, Tom Wild. And I'm Paragol Green. Andrea, thank you very much. Thank you, guys. Thank you for having me on a great discussion. Perfect. Thanks. Thank you.