I love a good—fictional—murder mystery. Give me a quaint Scottish village where someone disappeared way before the newly minted detective with her dorky fanny pack was born (Karen Pirie, 2022), or a dark, gritty Copenhagen crime so gory I’ll never go to the Memorial Park again (Forbrydelsen, 2007). What I like less is a sloppy plotline that can only get resolved by some long-lost cousin-in-law with a massive grudge being introduced in the last seven minutes of the season finale. Deus ex machina—"God in the machine"—no thank you!
Or as Gillian Tett said in the Financial Times about something else entirely:
“Intelligent machines do not automatically deploy themselves either for good or bad. Human strategy is crucial.”
Gillian Tett in FT, May 8th, 2026.
It is, however, very human to want the intelligent machines to be magic wands that will do everything for us, at least until they threaten our livelihood.
As we are rolling with the AI tsunami, there are hints that point to this new technology being useful for collaboration rather than replacement and that we, the humans, still need to carefully stitch the storyline together before engaging the heavy machinery.
What we are learning is that AI does not, by itself, fix weak processes, fragmented workflows, or bad data; it exposes them. If we want to get the most out of the intelligent machines, we still need to bring our house in order beforehand. That can seem a little disappointing, but at least now we know where to start.
Another surfacing point is that being powerful is not the same as being altruistic or unbiased. Generative AI is built to give us the most likely answer to our question, and while the pool of potential answers is vast, it is still both branded and increasingly conditioned on the framing of the question.
A recent study from the University of Maryland, the National University of Singapore, and Ohio State looked into whether LLMs tasked with screening job applications prefer applications they have written themselves, as in the same LLM, to those written by humans or other LLMs. The conclusion is OVERWHELMINGLY SO, with 65-80% self-preference. And to pile it on, the self-preference is strongest in business-related fields like accounting, sales, and finance.
While the result makes sense because the LLM deems the most likely answer as the one it would have given itself, it also shows that, in the most literal sense, we get what we ask for. When we choose a particular flavour, we largely exclude all others, and we must take that into account, especially when we fold generative AI into agents that work autonomously.
A good team employs both generalists and specialists; juniors and seniors; those who follow instructions and those who vibe. What makes such a team even better is if the team members do what they do best without holding up anyone else. Less trivial is it that the best-doing does not scale equally across the team. As it turns out, experience, AKA institutional knowledge, scales differently than operational tasks.
And I am not (just) saying that because I am, hhmmm, seasoned. The next generations, of both human and digital persuasion, will do everything better and more efficiently because that is what progress and innovation look like. But integrating the lessons already learned and the knowledge amassed will make them get there so much faster.
So rather than blindly tasking AI with everything, a better way to deploy these intelligent machines is to provide them with context or metadata, use that context to build practical instructions or blueprints for what they should do, and then continually supercharge the collaboration with as much experience, knowledge, and human strategy as possible.
Less deus ex machina, more deus ex collaborationis.
Regitze Ladekarl, FRM, is FRG’s Director of Company Intelligence. She has 25-plus years of experience where finance meets technology.
This article is part of the FRG Risk Report, published weekly on the FRG blog. To read other entries of the Risk Report, visit frgrisk.com/category/risk-report/.
