Note

“AI Saved Us RM40 Million” Is Not a Number a Board Should Accept

A bank tells its board that AI saved RM40 million this year. It sounds precise. It isn't. Blended into that one figure could be genuine savings, normal cost discipline that would have happened anyway, a good year for the loan book, and a bit of rounding that makes the number look tidier than it is. A board can't audit a single blended figure — it can only accept it on faith.

JPMorgan doesn't report AI value that way, and it's worth looking at exactly how it does. The bank tracks over 600 separate AI use cases, each with its own named metric tied to a specific business outcome — not one company-wide savings number. Its contract-review AI, for instance, is measured in hours of manual legal work eliminated (roughly 360,000 a year) and the resulting drop in legal costs (30%) and compliance errors (80%). Its trading AI is measured by win rate, which moved from 52% to 63%, and by a specific figure for reduced slippage cost ($25 million) — deliberately narrower than "AI made us more money," because a narrow, named metric is one a board can actually check against the trading desk's own numbers. Across all of it, JPMorgan puts the aggregate value at $1–1.5 billion — but the number that matters isn't the total. It's that every dollar of it can be traced back to one specific, defensible use case.

The three places the money actually has to show up

For a bank, an honest accounting of what AI changed splits into three separate lines, not one:

The margin line. A loan decision that takes days instead of hours doesn't show up as a loss anywhere on the books — it shows up as a customer who quietly took their business to a faster competitor. That's margin that was never visibly lost, only never won, which is exactly why it's the easiest of the three to overlook and often the largest.

The cost-of-operations line. This is the most directly traceable: rework, exceptions, and manual reconciliation that happen specifically because a decision arrived late or under-informed. Real, countable, and — unlike the margin line — usually already being tracked by someone before AI enters the picture at all.

The cost-to-income line — and this is the one banks tend to get wrong. The honest claim isn't "we cut headcount." It's avoided growth: a bank that would otherwise have had to hire more underwriters and analysts to keep up with rising loan volume doesn't have to, and the people it already employs spend more of their time on judgment than on coordination. The right way to show a board this line is a cost curve growing more slowly than loan volume — not a smaller number than last year.

A credible board presentation names all three, sized separately — not one blended efficiency percentage nobody can actually trace back to a cause.

What this looks like against Malaysia's own numbers

Malaysian banks are heading into 2026 with loan growth projected at 5.5%, profit growth expected to nearly quintuple to 4.9% from just 0.9% in 2025, and return on equity holding around 9.5% sector-wide (some individual banks are running considerably higher — Q3 2025 came in at 13.3%). Capital ratios sit comfortably above 17%.

That's the backdrop any AI savings claim has to be measured against. If a bank's loan book is already growing 5.5% and profit is already recovering sharply for reasons that have nothing to do with AI — a better credit cycle, stronger fee income, market conditions — then an AI-attributed "RM40 million saved" needs to answer a specific question before a board accepts it: how much of that is separable from a genuinely good year the bank was already having?

The part almost nobody puts a number on

There's a second effect worth knowing about, because it rarely makes it into a first board presentation. A bank that closes its information gaps faster isn't just cheaper on average — it's more predictable, because the same buffers that used to absorb cost also absorbed variance: swings in cycle time, in margin, in how much capital the bank needs quarter to quarter. A business whose earnings swing less from one quarter to the next is worth more to anyone pricing a claim on those earnings, independent of the average return itself. A board that evaluates an AI programme purely on expected savings is pricing only half of what it's actually worth.

Before the next number lands on the board table

The next time someone reports that AI "saved" a specific amount, the sharper question isn't whether the number is real. It's which line it actually sits on — margin, operating cost, or cost-to-income — and whether it could survive being tested the way JPMorgan tests its own: not at the platform level, but at the level of one specific, named use case someone in the room could check.

Sources: JPMorgan AI value measurement and use-case attribution, Olakai; Malaysia banking sector 2026 outlook, AmInvestment Bank via Bernama; Malaysia banking sector performance 2025, Malay Mail.

← All notes