The figure I noticed first in the 2024 Bank of England and FCA survey of AI in UK financial services was not the adoption rate. It was this one: "46% of respondent firms reported having only 'partial understanding' of the AI technologies they use." Only 34% claimed complete understanding. The survey puts most of the gap down to models bought from third parties.
Governance looks better on paper. "84% of firms reported having an accountable person for their AI framework." So in many firms someone is accountable for a system that the firm itself only partly understands. I keep thinking about that pairing.
The headline numbers are the familiar ones. The survey found that "75% of firms are already using AI, with a further 10% planning to use AI over the next three years." Firms using AI reported a median of nine use cases each.
Autonomy is another matter. "Only 2% of use cases have fully autonomous decision-making." A much larger share, 55%, has some automated decision-making, usually with a person stepping in on critical or unclear cases.
I turned the percentages into counts, because counts are easier to argue with. The survey covered 118 firms. At 75% adoption that is 88.5 firms using AI. At a median of nine use cases each, the total is about 796. Two percent of those is roughly 16 use cases, across the whole surveyed population, where a machine decides with no person in the loop.
Sixteen, out of roughly 796. For all the talk about autonomous finance, that is the supervisors' count, and I find it hard to read past.
The number is low partly because I used a median. Large UK and international banks reported medians of 39 and 49, which drags the average well above nine. At a mean of 1.5 times the median the total is 1,195 use cases and about 24 autonomous ones. At twice the median it is 1,593 and about 32. The survey also expects "the median number of use cases to more than double over the next three years (from 9 to 21)", which would put the count near 1,858 with around 37 autonomous.
If you are doing diligence on a target that claims AI across the enterprise, use these numbers. Ask for a list of use cases, classify each by autonomy and by materiality, and compare the count with nine today and twenty-one in three years. Ask early, because the list takes time to build. A claim that cannot survive it is marketing.
The price side is in Mastercard's 2024 annual report. "In December 2024, Mastercard acquired a 100% equity interest in RF Ultimate Parent, Inc. ("Recorded Future"), a global threat intelligence company, for cash consideration of $2.7 billion." Goodwill was $1.7 billion, which the company says "is primarily attributable to the synergies expected to arise after the acquisition date. None of the goodwill is expected to be deductible for local tax purposes."
I divided one by the other. 63.0% of the price is goodwill, and the identifiable assets, the part you could point at on the day of closing, come to $1.0 billion, or 37.0%, which for a threat intelligence business with a long list of customers is less than I expected. Mastercard's earlier purchase of Dynamic Yield in 2022 followed the same pattern: "cash consideration of $325 million", with "goodwill of $200 million", or 61.5%. Across both deals, $1,900 million of goodwill on $3,025 million paid is 62.8%.
That is a lot of hope in one number. A goodwill share above 60% tells me the price is mostly a forecast of synergies. If I were on the buying side, I would ask for the synergy model before the product roadmap. A 25% shortfall against those synergies would put $425 million of the Recorded Future goodwill at risk, 15.7% of what Mastercard paid.
The research I trust most on the downside is "Predictably Unequal? The Effects of Machine Learning on Credit Markets" by Andreas Fuster, Paul Goldsmith-Pinkham, Tarun Ramadorai and Ansgar Walther, in The Journal of Finance. On US mortgage data they find that "Black and Hispanic borrowers are disproportionately less likely to gain from the introduction of machine learning", and that in equilibrium "machine learning increases disparity in rates between and within groups, with these changes attributable primarily to greater flexibility."
The last clause is the one for diligence. The gap comes from the flexibility of the model, not from banned inputs. A lender can show a clean list of variables and still produce the outcome a supervisor will question. I now think a variable list alone answers nothing.
I suspect the 2% figure hides where the exposure actually sits. The survey reports that "55% of all AI use cases have some degree of automated decision-making with 24% of those being semi-autonomous", and that "62% of all AI use cases are rated low materiality by the firms that use them with 16% rated high materiality."
Materiality is self-assessed, a third of use cases come from third parties, and nearly half the firms only partly understand what they run. So the same firm that cannot fully explain a vendor model also decides how much that model matters. I cannot tell from the survey how often a person signs off on output nobody in the room can explain, and that is the number I would want before paying 63% goodwill for anything.