Tanmay's Mycroft Update: Their Numbers, My Arrows – Separating AI Narratives from Facts
Every number in a paragraph about HSBC's AI-driven savings was real, but the sentence connecting them into a story about AI making the bank leaner was invented, not reported.
Here's a paragraph about HSBC written before checking a single source: HSBC engineers got dramatically faster using AI. The bank switched off more than a thousand old systems, found $1.2 billion in savings, and is moving another $1.5 billion into growth. AI is making this bank leaner. Every number in that paragraph is real, and every number came from HSBC. The last sentence, the one that ties it all together and makes it a story, was made up. Not deliberately, just written because that's what those numbers look like sitting next to each other.
The two-column method
The fix Tanmay Kulkarni demonstrates is a simple exercise: two columns. On the left, everything HSBC actually said. On the right, everything added on top of that. Then, going through the right-hand column line by line, trying to move each item across to the left. Whatever won't move, whatever can't be traced back to something HSBC actually stated, gets marked as invented.
The left column holds HSBC's own figures, drawn from its own results and earnings call transcripts: the speed-ups, the applications retired, the savings realized, all sourced and on the record. All of it is self-reported and unaudited, with no published methodology behind it, which isn't a complaint, just a fact worth knowing. HSBC's own language is notably careful throughout: coding assistance that engineers use, described as assistive, never described as autonomous.
Five assumptions, checked one by one
The right column held five lines, all five things Tanmay believed, none of which actually came from HSBC.
The first: AI made engineers faster, so the bank needed fewer people, so it saved money. That connection isn't in the filings anywhere. HSBC never joins coding assistance to headcount savings in the transcript, the investor presentation, or the release. They don't deny the connection either. They simply never state it, so that line doesn't move to the left column.
The second: the reported speed numbers, like being faster by some multiple, imply a method someone could go inspect. Faster than what, measured how, using which tool against which baseline? None of it is disclosed anywhere in the public record. The most technical detail available is HSBC's chief executive calling the tools "vibe coding assistance." The outcome is real; the method behind it is absent.
The third assumption is the one Tanmay stopped and read twice: that $1.5 billion means one thing. In a single transcript, on one call, HSBC uses $1.5 billion twice, for two different things. The first use: "we are taking $1.5 billion of annualized simplification savings straight to the bottom line," money being kept. The second, later in the same call: "positive progress with the reallocation of circa $1.5 billion" from non-strategic or lower-returning businesses, money being moved. Same figure, opposite jobs, sitting minutes apart in HSBC's own words, with nothing hidden. It isn't a one-off either: a separate figure, $1.8 billion, meant severance and restructuring costs in HSBC's February 2025 reporting, then meant reallocation, including $300 million in Hang Seng cost synergies, in February 2026. Same number, different year, a completely different meaning both times.
The fourth assumption: that somebody out there must have gotten this wrong. Tanmay went looking for a publication that had misreported the $1.8 billion figure as new AI investment money, and found three that covered it: Kingy AI correctly spelled out that it was a reallocation, not new money; Met Intro gave the same breakdown; Infotech Lead put the word "reallocates" directly in its headline. There was no culprit to find, which turned out to be the most instructive result of the whole exercise: the answer had been sitting in HSBC's own transcript the entire time.
The fifth assumption: that HSBC's coding tools are not agentic AI. This one gets relabeled rather than struck out entirely, because HSBC does use the word "agentic," just not for the coding tools. Four months after the productivity figures, HSBC announced a financial crime detection system built with Google Cloud, and there it explicitly used the term "agentic," its own word. So one system gets called agentic and the other doesn't, which is a real and meaningful distinction, but HSBC has never stated outright that its coding assistants are not agentic AI. Nobody at HSBC put those two announcements side by side. That line stays on the right, grayed out and labeled as a personal reading, revisable if HSBC describes the coding tools in agentic terms in a future quarter.
Testing the boundary in code
The exercise extends into a small working simulation of HSBC's described code-patching pipeline: a vulnerability comes in, an AI assistant drafts a fix, a human reviews it, and the fix gets applied and tested. At the review stage, a real decision had to be made about approval criteria, auto-approving anything below a certain severity, requiring sign-off above some confidence threshold. That decision couldn't actually be made, because HSBC has never described what its code review gate looks like. It has stated a general principle, that AI deployments get reviewed, monitored, and audited, and that's the entire public record on the subject. So the simulated gate ships with no criteria at all. It refuses to run until someone hands it one, raising an error and stopping rather than inventing a threshold HSBC never specified.
What's actually left standing
After the exercise, what remains defensible is smaller than the opening paragraph: HSBC engineers got measurably faster at specific tasks by HSBC's own measurement, HSBC switched off 1,165 old applications, HSBC realized $1.2 billion in simplification savings, and HSBC is moving $1.5 billion of costs into growth, a different $1.5 billion from the one going to the bottom line. Four things, all sourced, all dated, none of them explaining any of the others.
Key takeaways
- Every individual number in an AI-related news narrative can be accurate while the sentence connecting them into a story is invented.
- The two-column method works by separating what a company actually said (left) from what you added or assumed (right), then testing whether each right-hand line can move to the left.
- The same dollar figure, such as HSBC's $1.5 billion or $1.8 billion, can refer to entirely different things across different statements or reporting years.
- Absence of denial is not confirmation: HSBC never explicitly denied that AI-driven speed fed into headcount decisions, but it also never stated that connection.
- Building a simulated review pipeline exposed that inventing approval criteria HSBC never published would mean pretending the company works in a way it has never actually described.
Try it yourself
Take any AI-related announcement, a press release, an earnings call, a vendor page, and build the same two columns: on the left, only what was actually said; on the right, everything you added while reading it. If your right column barely empties when you try to move each line across, that's a sign of how much of the reading was yours rather than the source's. This exercise is part of the ongoing Mycroft Financial AI series from Humanitarians AI Fellows.
Chapters
- 0:00The HSBC Narrative: Real Numbers, Invented Story
- 0:40The Two-Column Method: Facts vs. Assumptions
- 1:15Checking the Record: Headcount Savings & "Vibe Coding"
- 2:05The $1.5 Billion Trap: Same Number, Different Meanings
- 2:50Building the Code Gate: Simulating HSBC's Review Pipeline
- 3:30"Your Turn": Try the Two-Column Challenge Yourself
Full transcript(auto-generated, with timestamps)
The HSBC Narrative: Real Numbers, Invented Story
[0:00]Hi, this is Tanmay Kolkarni from Humanitarians AI. Here's a paragraph I wrote about HSBC before I checked anything. HSBC engineers got dramatically faster using AI. The bank switched off more than a thousand old systems, found 1.2 billion dollars in savings, and is moving another 1 and 1/2 billion into growth. AI is making this bank leaner. Every number in that paragraph is real. Every number came from HSBC. And the last sentence, the one that ties it together, the one that makes it a story, I made that up. Not deliberately, I just wrote it because that's what those numbers look like when you put them next to each other. So, I did something
The Two-Column Method: Facts vs. Assumptions
[0:40]Simple, and it's the only method in this video. Two columns. On the left, everything HSBC actually said. On the right, everything I added. Then I go through the right-hand column line by line and try to move each one across to the left. Whatever won't move, I made up. Left column first. Everything on it is HSBC own figure from HSBC own results and transcripts, the speed ups, the applications retired, the savings realized. They're on screen with their sources. I'm not going to recite them at you. Two of them I do need you to notice though because they come back. 1 and 1/2
Checking the Record: Headcount Savings & "Vibe Coding"
[1:17]Billion in simplification saves going straight to the bottom line. And 1 and 1/2 billion in reallocation costs going into growth. All of these are self-reported and unaudited with no published methodology behind them. That isn't a complaint, it's just what they are and you should know it. And notice the care in HSBC language. Coding assistance that engineers use, assistive. They never once call them autonomous. Now, the right column. And this is the uncomfortable part because I have to write down what I actually thought. AI made them faster so the bank needed fewer people. The numbers imply a method I could go and inspect. 1 and 1/2 billion means one thing. Somebody out there has definitely got this wrong, and their coding tools are not agentic AI. Five lines. I believed all five of them.
The $1.5 Billion Trap: Same Number, Different Meanings
[2:05]Not one of them came from HSBC. First line, AI made them faster, so the bank needed fewer people, so it saved money. Go and look for that in the filings. It isn't there. I couldn't find HSBC connecting the coding assistance to the headcount savings anywhere. Not in the transcript, not in the investor presentation, not in the release. They don't deny it, either. They simply never join those two things up, so that line doesn't move. That line was mine. Second line, the numbers imply a method. 60% faster, five times faster, surely somebody measured something, and I can go and read how they did it. Faster than what? Measured how? Which tool, which model, against which baseline? None of it is disclosed anywhere. The most technical thing on the entire public
Building the Code Gate: Simulating HSBC's Review Pipeline
[2:50]Record is HSBC chief executive calling them, and I do genuinely love this, vibe coding assistance. The outcome is real. The method is absent, and I had quietly assumed one from the other. Third line, and this is where I stopped and read it twice. 1 and 1/2 billion means one thing. Except in that one transcript, one document, one morning in February, HSBC uses 1 and 1/2 billion dollars twice. Here's the first. We are taking 1.5 billion dollars of annualized simplification saved straight to the bottom line. That's money they're keeping. And here's the second, further down the same call. We are also making positive progress with the reallocation of circa 1.5
"Your Turn": Try the Two-Column Challenge Yourself
[3:31]Billion dollars from non-strategic or lower returning businesses. That's money they're moving. Same figure, opposite jobs, and nobody hid anything. Both sentences are sitting right there, in HSBC own words, minutes apart. But if you'd skimmed that call and walked away holding one number, there's a decent chance you'd be holding the wrong one. And it isn't a one-off. $1.8 billion. In February 2025, that figure meant severance and restructuring costs. In February 2026, the same figure means reallocation, 1 and 1/2 billion of it, plus 300 million in Hang Seng cost synergies recycled back into growth. Same number, different year, a completely different meaning. Fourth line, somebody out there must have got this wrong. This is the one I'm least proud of because my instinct wasn't to go and check my own reading. It was to go and find somebody else's mistake. Some outlet that had grabbed 1.8 billion and called it an AI investment. I found three publications that covered it. All three of them got it right. Kingy AI
Spelled out that it was a reallocation, not new money. Met Intro gave the same breakdown. Infotech Lead put the word reallocates in the headline. There was no culprit. So, I stopped looking for one and went and read HSBC transcript myself, which is where the good example had been sitting the whole time. That line doesn't move either. And it's the one that taught me the most. The danger here isn't that people are careless. It's that a number can honestly mean two things and you still have to go and check. Fifth line, their coding tools are not agentic AI. And this one I'm not striking out. I'm relabeling it because HSBC does use the word agentic. In June, 4 months after those productivity figures, they announced a financial crime system with Google Cloud and there they used it, agentic. Their word, not mine. So, one system gets called agentic and the other one doesn't. That's real and I think it means something. But, HSBC has never said, "Our coding
Assistants are not agentic AI." Nobody at HSBC put those two announcements next to each other. I did. So, this line stays on the right, grayed out with a label on it. My reading, revisable. If HSBC describes those coding tools in agentic terms next quarter, I'm not going to defend this. I'd just be wrong and I'd change it. There's one more place this happened and it's the part I actually enjoyed. I built a small working version of that patching pipeline. Four stages. A vulnerability comes in, the assistant drafts a fix, a human reviews it, then it gets applied and tested. And at the review stage, I had to decide what the criteria were. Auto approve anything below a certain severity, require a sign-off above some confidence score. I couldn't do it because HSBC has never described a code review gate. They've stated a general principle, AI deployments get reviewed, monitored, audited and that is the entire public record. So the gate in my version ships with no criteria
At all. None. It refuses to run until somebody hands it one. It raises an error and stops. That empty gate is this whole video written in code. If I'd invented a threshold there, I'd have implied HSBC works in a way nobody has said they work. So where does that leave you? Because both of the easy answers are wrong. You can't assume the connection, nobody stated it. But you can't assume there isn't one either. HSBC hasn't denied it and it would be a perfectly reasonable thing for faster engineering to feed into head count decisions. It just isn't on the record. So what are you actually allowed to walk away with? This, the left column on its own. HSBC engineers got measurably faster at two specific tasks by HSBC own measurement. HSBC switched off 1,165 old applications. HSBC realized 1.2 billion dollars in simplification savings and HSBC is moving 1.5 billion of costs into growth, which is a different 1.5 billion from the one going to the bottom line.
Four things, all sourced, all dated, not one of them explaining any of the others. That's smaller than the paragraph I opened with. It's also the version I could defend in a room full of people who'd read the filings. And that trade, a smaller thing you can actually hold instead of a bigger thing you can't, that's the whole point of the two columns. Your turn and it's one instruction. Take any AI announcement, a press release, an earnings call, a vendor page, make two columns. On the left, only what they actually said. On the right, everything you added. Then try to move each right-hand line across. You've done it right if your right column barely empties. That isn't failure, that's you finding out how much of the reading was yours. You've done it wrong if everything slides across on the first try. That means you're filling the left column in your own handwriting. Their numbers, my arrows. It's worth knowing which is which. This is Tanmay Kolkarni from Humanitarian's AI.
