I concluded our last article by suggesting that AI ‘is unlikely to be any simpler to employ than human capital in complex organisations’. There are a number of factors that influence this, but the most telling of these for me is alignment. Alignment in the distribution chain Almost all MGAs are familiar with the concept from the time they have spent negotiating their risk capital requirements. Every firm in the insurance distribution chain needs to be aligned, so that everybody in the chain wins together, or, when there are losses, everyone acts in concert to minimise losses in a way that is fair to the customer. Put more simply, everyone must have ‘skin in the game’ and, when possible, this must be of near-equal significance to them as it is to other parties in the distribution chain. In the world of insurance this is reflected by deductibles, commission returns / clawbacks, profit commissions, fronting fees, fronting risk retentions and so on, all along the insurance, reinsurance, and retrocession chain. These exist because alignment protects everyone’s profit. Also, both UK and EU law and regulation make alignment in the distribution chain a key priority. Competing priorities with AI providers If you attempt to apply these rules to an AI-intermediated business model, it quickly becomes clear that similar equivalence of alignment typically does not exist. In some cases, it can be argued that there is little alignment at all. This is because data is AI’s currency and good governance should prevent it from being harvested without good reason. This is perhaps the largest elephant in the room. AI firms have a demonstrated and persistent history of scraping data from the internet and elsewhere to train their large language and other AI models. What is more data is an MGAs key asset. If an MGA decides to crystallise its investment, what is for sale is distribution and data. I highly doubt a prudent MGA business that would put its key IP into the hands of a larger, third-party organisation with the funding and scale to duplicate its business model. However, with human factors involved, every time a business deploys AI to improve efficiency, there is a risk that key business IP and processes are learned by that AI whilst it is involved in a related task, such as, for example, formatting a spreadsheet or rewriting legacy code. Put simply, if data and intellectual property are not harvested by AI firms and resold for profit, then the current AI business model fails. The Economist estimates that AI firms have invested around USD1.4 trillion whilst only recovering around half of that. Therefore, it can be argued that there is a perverse incentive for firms providing both cloud storage and AI services to put relatively light guardrails around data in their care. Whilst it is perhaps unfair to say this is deliberate, each time an event occurs, their AI tools learn, which, ultimately is a business goal for such firms. Examples are already emerging, including AI accessing and potentially further disseminating the contents of confidential emails Add one final unpredictable variable into this mix, that of human nature, then it is probably safe to assume that, if you use AI without careful controls, it will get to look at pretty much everything essential to your business over time. So how do I make this work? There is, of course, the option of avoiding AI use altogether, although I suspect that any business that adopts such a strategy might be forcibly introduced to its use by human behaviour in any case, even where policies stating that AI must not be used are in place and regularly circulated. A more sustainable approach to AI must: • Increase Alignment •. Be structured around strong Data Management •. Manage Human Nature effectively •. Give primacy to protecting Intellectual Property •. Build Resilience not just by testing processes, but ensuring that structures are in place to maintain control and operational continuity …then engage carefully with AI, and only where an identified need exists and a suitable risk assessment has been undertaken. AI cannot be adopted safely overnight. MGAs must take time to evaluate their data and IP assets carefully and to decide what can and cannot be shared. This often requires that your existing data lake is carefully categorised and sorted, then renamed to identify types of content. Finally, the results of this process should be reviewed by at least a second pair of eyes, if not independently. Whilst I do not subscribe to an entirely bond-villain approach to AI providers, a recent effort by US government to pressurise AI provider Anthropic to reduce guardrails around its technology for military purposes suggest that reliance on third-party governance may be unwise, An AI firm is highly unlikely to, by default, offers terms that completely protect your interests and those of your customers. Therefore, you will need to refine your terms of engagement with them. A final warning AI is cheap and always on your doorstep because it wants your data, your business model, and it can be argued, your IP. As we will explore in the next article, the AI industry also wants your customers’ data, and you have a duty to manage the sharing of both commercial and protected data with proportionate care. MGAs should navigate any change to AI giving full weight to the fact that alignment in the distribution chain, the protection and location of data (both owned and in your control), and resilience in individual firms are enshrined in UK and EU law and regulation. Image: Generated by AI