Is AI really just about saving time or is there something bigger at stake for the London market?
“Human in the loop” is one of the most used phrases in AI, but what does it actually mean for insurers?
According to Tim Hardcastle of Instanda, there are three ways people and AI can work together: human-led, human in the loop and human-enabled. Each one trades control against scale.
He tells Insurance Insider’s Abbie Day that roles are shifting from doing the work to supervising it, raising questions about how junior staff will learn the ropes as entry-level tasks are automated.
And with regulators holding firms accountable for every decision, he explains why humans won’t be taken out of the loop any time soon.
This transcript was generated using AI and may contain errors. Please check against the audio before quoting.
Abbie Day 0:04
Hello everyone, and welcome to this week’s episode of In the Loop, where we bring you in-depth discussions about AI in the insurance industry. I’m Abbie Day, Senior Reporter for Insurance Insider, and today we will be diving deeper into the podcast namesake to learn a bit more about the phrase so often heard when discussing AI, human in the loop. Joining me is Instanda CEO Tim Hardcastle, and we will be discussing what the phrase human in the loop really means, the difference between human-led and human-enabled AI, and what approach is best for your business. As well as whether the insurance industry will ever use AI without a human in the loop. Without further ado, let’s get started. So this week we are deep diving into human interaction with AI with a very exciting guest. We have Tim Hardcastle, CEO and co-founder of Instanda. Hello and welcome, Tim.
Tim Hardcastle 0:58
Great to see you.
Abbie Day 0:59
So we are looking to understand more about what it means to have a human in the loop, kind of going off of our podcast namesake. So maybe we can start for any listeners who may not know Instanda or know who you are with a quick introduction to the company.
Tim Hardcastle 1:12
Okay, Instanda was created just over 10 years ago. We’re part of the insurtech wave, and it’s one of the world’s few, if not the only, no-code policy admin platforms. We serve MGAs and tier one, two and three carriers in the US, UK and Europe, and Australia. We currently have around about 100 clients using the platform, and obviously we’re very privileged to work with them because what we do is exceptionally important to their business success. And we’re at this great inflection point, which we’ll talk about, around how the future of technology will evolve in the world of AI.
Abbie Day 1:47
Amazing. So we’ve had conversations about AI in the past, and one of the conversations we’ve had that’s really stuck out to me is the meaning of the words human in the loop. We’ve spoken about human-enabled, human-led AI. Maybe we could start with the phrase human in the loop, and you could explain to me exactly what it really means. It’s a phrase we hear thrown around so much.
Tim Hardcastle 2:08
Well, like so many parts of the AI landscape that everyone is experiencing interacting with, I think things are evolving and moving quite quickly. From when we were talking, we talked about human in the loop. And I think I’d look at it through three different ways to express the human interaction with AI. So simplistically, the first one would probably be what I’d describe as human-led, where the person does the work and decides, and AI prepares and suggests. And I think many people would recognise that’s how the casual user might interact with AI today. Human in the loop, I would characterise that as where the AI does the work and the person checks or approves it at set points. And then the third way that you can look at the human and AI interaction is human-enabled, where AI works on its own within hard limits, so well-defined limits. So people set the rules, people watch the results and then deal with the exceptions. So there are three layers: the human-led, the human in the loop, and what I’d call human-enabled.
Abbie Day 3:12
Okay, so maybe we could go a bit more into how to decide which way you want your company to go. So we can start maybe with the underwriting side. Is it fair to kind of assign one type to one type of company?
Tim Hardcastle 3:26
Yeah, I think each level is trading control against scale. So human-led, where the human is making the decision, the AI is providing the draft, if you like, and a bit of a suggestion. It gives the best judgment. So from an underwriting perspective, it would be something where you’re dealing with complex, high-consequence risks, but it’s the slowest part of, let’s say, the decision-making process to scale for obvious reasons. In contrast, human-enabled is where, as I said, the AI is doing the work and the human is looking at the exceptions. So you get real scale with this, for example, pulling data out of documents. Obviously, you need to have good solid foundations around clean data, you’ve got your set parameters, and you need to record everything that AI is doing. But essentially, AI is doing the work for you, and, as I said, you’re looking at the exceptions. So if you contrast those two things, human-led, where the human is making the decision and AI is providing some of the drafting and suggestions, versus human-enabled, you can see the difference between the control and the scale. And then the one in the middle, human in the loop, you get throughput with some control because the AI may be triaging submissions, drafting a quote, for example, inside an underwriter’s authority, and then the underwriter approves. So the human in the loop in this context is the one in the middle from a scale and control perspective. And I think an organisation should look at the kind of activities, tasks and decisions it needs to make. And it would decide whether it’s going to be human-led, as in very heavily oriented around the underwriter making these complex decisions, or human-enabled, where the AI is doing a lot more of the processing and the humans are looking at just the exceptions.
Abbie Day 5:08
Okay, so maybe we could talk about the teams themselves. Does the team you’re working with, the team you’re managing, change the way you would approach this as well?
Tim Hardcastle 5:17
It’s a really interesting point because I think when we start to reconsider and reframe, if you like, the workplace environment, the structure almost of what the future will hold and what some companies are experiencing even now, what we can see is jobs, in the way I’ve just described those interactions between the AI and humans, jobs are shifting from doing the work to supervision and, as I said earlier, being the final decision maker. So underwriters, for example, can spend more time on exceptions, portfolio decisions and relationships, and the work that they previously might have been caught up in, which was stuck in their inbox and perhaps in spreadsheets and trying to chase topics around, that activity can be and should be passed to more of the AI capabilities so that the managers are spending more of their time on the high-quality activities. And I think that’s a good thing for MGAs and carriers. I think it’s a good feeling for the industry because, you know, often underwriters are the most highly experienced and quite well remunerated. I think the quality of their input should be on the right things rather than, let’s say, the slightly more rudimentary or admin-related activities. So that’s one thing. So the shift from doing the work to supervising, we’re seeing, and I think we can expect more of. But I think the other interesting thing around jobs and the workplace, and we’re experiencing this ourselves as well, is new accountabilities are appearing. So we think of agents as like new employees. They need boundaries, job descriptions, a manager, a limit on what they can and can’t do. So someone has to own and manage the agents that are being deployed across an organisation. So this is a new challenge for HR, for how team members interact with AI. And I think again, that’s an emerging area. The one that perhaps worries me most for the industry is the apprenticeship problem, which is where junior people, you know, leaving university or A levels and coming into the industry, need to learn by doing the more straightforward tasks. And obviously, AI can now do that. The question I think for many MGAs and carriers, as they’re building their businesses and taking the benefits of this more efficient way of working, is how do they find better routes for more junior members of the team to come through and learn how to make judgments and get the experience that the classic pyramid structure of an organisation would traditionally have given them, allowing them to start at a junior level and then move their way up. And that’s obviously not as available now as it was literally six or nine months ago. So I think there are quite profound impacts on organisation and roles within carriers and MGAs.
Abbie Day 7:58
I mean, the training point that you made is something that I’m talking to so many people about. How do we completely train our staff to work with this new technology? One thing I think about a lot is it feels a lot easier to check our own work than it is to check an AI’s work in some cases. What kind of training needs to be offered to new staff, to managers that are already in place? Does it go beyond just the skills that we’re using with AI into kind of new management styles and things like that?
Tim Hardcastle 8:28
There are multiple layers to that point you’re making. Whilst I’ve talked about the work shifting from trying to do all tasks to this higher-quality oversight and making the final judgment, I think in some roles, the skill development or the management of the activity by AI is quite straightforward because the underwriters that are typically in those positions, making those decisions on complex risks, are very experienced. They know what right looks like. They’ll use various different sources of information and other judgment to form a view. And I think for those kinds of groups and communities, I don’t think there will necessarily be particular retraining. Historically they’d have been managing the junior underwriters, and now increasingly they’ll be using or managing AI fulfilling some of the tasks of the junior underwriters. So I think they would just review and see that in the same way that they would have done with individual junior underwriters. So I think for that group, it’s straightforward. When you go to the other end of things, as we talked about earlier, with human-enabled, where I was talking about AI working at scale, the humans are allowing the AI to do the work and they’re actually judging the exceptions. That’s where what we’re seeing some organisations do is put in multiple layers of governance. And we’re doing the same thing with the way that we provide some of the AI support across our platform, where you’ll have a separation of governance from the AI that’s actually doing the work, and then another layer of governance, which again is a separate AI capability, which is doing some of the reviewing. So you end up with what some people are calling a council, where you have different AI models reviewing each other’s work, and you’re triangulating how the producer of the workflow, if you like, is doing that to a level of accuracy, because the humans are checking the exceptions, but they’re also using AI to do the reviewing of, let’s call it, the originator’s work. And the skills that you need for that start to become a set of skills that people need to know about how to run, manage, orchestrate and put controls around AI. That’s an area where training and development is literally having to catch up in real time with the AI capabilities that are getting better and better. And so what we’re seeing across our client base, or companies we’re talking to, is that there are activities going on where people are being helped to put themselves in a position to fully understand the limits and the constraints on AI activities in this workflow, and how they use AI to do the governance and the oversight on the AI itself. And then finally the human layer where they’re looking at exceptions. The training and development that’s going on is happening in a very dynamic way because people are scrambling to put in formal learning mechanisms. For example, you can see at the upper echelons of training that there are now master’s level courses being put in place with some of the leading universities in the US and the UK for AI strategy, AI governance and AI management. And some of the leaders that we talk to have already been on some of these courses, which are being put in in an accelerated way, say three months. And the final point I’d make comes back to my earlier issue around how you then nurture and develop talent as it comes through when you’re enriching the organisation by using more AI. That’s the one I don’t have an immediate answer for. I think companies are tackling that in different ways, but I think that is an industry-wide challenge and something that I’m, you know, all ears about if people are coming up with new initiatives to achieve that.
Abbie Day 12:03
And you mentioned earlier this kind of pyramid shape that we have in place for companies at the minute. It sounds like from what you’ve been saying, that could change as AI comes in and we’re managing people alongside AI. What are you expecting that shape to look like, say, five years in the future?
Tim Hardcastle 12:21
I think what we can foresee is that the classic pyramid is going to narrow at the bottom. And so there is an open question: where are tomorrow’s senior underwriters going to come from? There’s a challenge for the industry around the fact that, as I said earlier, senior underwriters have accumulated years of experience, and when they’re managing AI that’s supporting them on their decision making, they’re in a very comfortable space because the AI is removing some of the unnecessary work they were doing. And the question that remains open is how do you bring people through to get up to that level of seniority? I think inevitably what we’ll see is that there will still be processes. For example, I was just talking to one of the main consulting firms the other day. They’re still maintaining their graduate intake programme, putting their graduates through structured training. But as I said earlier, that training is now orientated and skewed very heavily towards understanding how to manage, orchestrate and govern AI. And I see companies doing the same thing, but I think we have to recognise from an economic standpoint that not as many graduates are going to be taken into the industry to fulfil tomorrow’s underwriting roles as would have been a year ago. I am concerned about that, but I think it’s an inevitability of what we’re seeing in terms of the displacement in the workplace, with AI doing the jobs that previously would have been done by the juniors in the underwriting team.
Abbie Day 13:46
The last question does ask you to look ahead slightly. I’d like to know whether you think the insurance industry or the companies within it lend themselves to use cases where humans eventually aren’t in the loop.
Tim Hardcastle 13:59
If you take a step back and look at the context and the environment in which MGAs and carriers, that we know particularly well, and even brokers operate, I can’t see a future where we’re going to have a world in which all transactions, all relationships, if you take the broker component, and all underwriting decisions across all lines of business are fully automated. I think that is fanciful. What we have to recognise is the context within which brokers, carriers and MGAs operate. They all work within regulated environments. They have responsibilities, not just to their customers, but obviously within that regulated context. And just because AI is available and is potentially a force for improved efficiency and streamlining operations, it doesn’t remove the accountability for ultimately making decisions. And so in any context in which AI is being used, there are two or three questions that people need to be able to ask, which are: what happens if it goes wrong? Can we validate, determine and be explicit about how decisions were made? Because whether you’re a carrier or an MGA, ultimately, if decisions are made by AI and go badly wrong, you’re accountable to the regulator for those decisions. So I can’t see any scenario in the near term, and I’m talking three to five years, where you’re going to move to a fully automated environment. There will be improvements, as I said, in underwriting operations and other parts of the insurance value chain where, you know, ingestion and data analysis can be done by AI. But I can’t see a future where humans, whether that’s human-led, human in the loop or human-enabled, will be removed completely.
Abbie Day 15:42
So it’s not about to take all our jobs then?
Tim Hardcastle 15:44
No, I think we can be confident that there will be plenty of employment. And I guess the other perspective, to end on a slightly more positive note, because we were talking about potentially some of the pyramid collapsing for some of the junior intake. You know, let’s take a step back and think about the world of insurance in terms of what the opportunity is for the insurance community and insurance players. And I would say this: there is a huge unmet need for insurance. And that crosses the span of P&C and life and health. So there’s nowhere near enough people who have life insurance or health insurance, and there’s nowhere near enough businesses that are fully covered from an insurance standpoint on all of their assets and the services that they deliver. And so there’s a huge unmet need across all lines of insurance, I would argue. And the opportunity is for all of the MGAs and carriers, while they can lean out and be more streamlined and become more efficient because of AI, to also look to grow their businesses by serving this unmet need. And that should, everything else being equal, bring in far more people to support that growth. Because notwithstanding there’s going to be a reduction in some of the activities that juniors would have undertaken, the businesses can be growing to fulfil these unmet needs. Timing is always an issue about what happens first. But in large part, I think the industry could certainly be confident that it should be employing the same, if not more, people over that same time horizon of three to five years, because again, they can get more efficient and they can serve these unmet needs more effectively. And therefore, there’s a big growth agenda, I think, for many companies to look forward to.
Abbie Day 17:29
Amazing. I think that’s a good note to finish on. So, Tim, thank you so much for chatting with me today. It’s been really, really interesting. Thank you.
Tim Hardcastle 17:36
Pleasure.