Latest / DX Today | No-Hype Podcast & News About AI & DX / Goldman Pulls Claude Code From Its Traders and Bankers

Transcript

Goldman Pulls Claude Code From Its Traders and Bankers · DX Today | No-Hype Podcast & News About AI & DX

• 0:00
  SPEAKER_01: Welcome to the DX Today Podcast, your daily deep dive into the AI ecosystem.
• 0:05
  SPEAKER_01: I'm Brian, and joining me as always is Jessica.
• 0:08
  SPEAKER_00: Thanks, Brian.
• 0:09
  SPEAKER_00: Today we are talking about a story that every company handing powerful AI tools to its whole workforce should be paying close attention to.
• 0:17
  SPEAKER_00: Because one of the most famous banks on Wall Street just took some of those tools back.
• 0:21
  SPEAKER_01: Right, this is Goldman Sachs.
• 0:24
  SPEAKER_01: According to Bloomberg, the bank has told its traders and bankers that they can no longer use clawed code from Anthropic or GitHub Copilot from Microsoft for their work.
• 0:33
  SPEAKER_01: And those tools are now reserved for engineers.
• 0:36
  SPEAKER_00: And the timing matters here.
• 0:38
  SPEAKER_00: The internal guidance was updated in October, and Bloomberg reported it on November 2nd.
• 0:44
  SPEAKER_00: For months before that, bankers and traders had been allowed to use these coding assistants, so this is a real reversal, not a policy that was always on the books.
• 0:52
  SPEAKER_01: So walk me through who actually keeps access, because I think the details tell us a lot about what the bank decided it was really buying with these tools in the first place.
• 1:00
  SPEAKER_00: Goldman has roughly 12,000 engineers and they keep the tools.
• 1:04
  SPEAKER_00: There is also a group the bank calls strats, people who sit across trading, quantitative research, and software engineering, and the new rule does not apply to them either.
• 1:14
  SPEAKER_00: Everyone else, which is most of the front office, loses access.
• 1:18
  SPEAKER_01: That is a pretty clean line.
• 1:20
  SPEAKER_01: If your job title says you write software, or you live halfway between the trading desk and the code base, you keep your assistant.
• 1:28
  SPEAKER_01: If you are a banker who taught yourself to build little tools on the side, you do not.
• 1:32
  SPEAKER_00: Exactly.
• 1:33
  SPEAKER_00: And some of those people are not happy about it.
• 1:36
  SPEAKER_00: One person familiar with the change told Bloomberg that traders had been using the tools to monitor client trading flows, to model market moves and positions, and even to create written and graphic content for their work.
• 1:47
  SPEAKER_01: I have to say, as someone who has spent a lot of time keeping real systems running, the phrase a trader built a tool to monitor client flows makes me a little nervous.
• 1:55
  SPEAKER_01: And I suspect it made somebody in the risk department nervous too.
• 1:59
  SPEAKER_00: That is basically the official explanation.
• 2:02
  SPEAKER_00: A Goldman spokesperson said the bank makes certain AI tools available to determine the best use cases across the firm, and that after a period of experimentation, where it assessed factors including productivity and risk management, it puts them in the hands of those who are best suited to leverage them for the long term.
• 2:19
  SPEAKER_01: Which is a very polite corporate way of saying the experiment ran.
• 2:22
  SPEAKER_01: The bank looked at the results, and it decided that the people who were not trained engineers were not producing enough value to justify the risk they were creating.
• 2:30
  SPEAKER_00: I would push back a little on that reading, Brian, because the bank has not published any numbers.
• 2:35
  SPEAKER_00: We do not actually know whether the traders were unproductive.
• 2:38
  SPEAKER_00: We only know that productivity and risk were weighed together.
• 2:41
  SPEAKER_00: And the combined answer came out in favor of narrowing access.
• 2:45
  SPEAKER_01: Fair point.
• 2:46
  SPEAKER_01: And I will take the correction.
• 2:47
  SPEAKER_01: We do not have the scorecard.
• 2:49
  SPEAKER_01: But what makes this story so interesting to me is how sharply it contrasts with what the same bank was saying out loud only a few months ago.
• 2:57
  SPEAKER_00: Right.
• 2:58
  SPEAKER_00: This is the part that jumps out.
• 3:00
  SPEAKER_00: Back in July, the chief information officer at Goldman, Marco Argenti, was publicly praising vibe coding by non-technical staff.
• 3:07
  SPEAKER_00: The idea was that a banker could build a rough solution with an AI tool and then hand it over to one of the engineers to clean up and harden.
• 3:15
  SPEAKER_01: And that sounds wonderful in a keynote.
• 3:17
  SPEAKER_01: In practice, though, every one of those handoffs lands on an engineer's desk as a piece of software they did not design, did not choose the libraries for, and now have to own, test, and support forever.
• 3:29
  SPEAKER_00: That is the hidden cost people miss.
• 3:31
  SPEAKER_00: If you have 12,000 engineers and a front office of thousands more people who are all now producing prototypes, the bottleneck is not writing code anymore.
• 3:40
  SPEAKER_00: The bottleneck becomes reviewing it, securing it, and deciding which of it deserves to live.
• 3:45
  SPEAKER_01: That is exactly what happens on Monday morning.
• 3:48
  SPEAKER_01: Somebody shows up with a spreadsheet replacement they built over the weekend.
• 3:51
  SPEAKER_01: It touches real client data, and now the question is whether it goes through change control, whether anyone audits it, and who gets paged when it breaks at four in the morning.
• 3:59
  SPEAKER_00: And in a bank, that question has regulators attached to it.
• 4:03
  SPEAKER_00: Banks have very specific obligations around model risk management, data handling, and record keeping.
• 4:08
  SPEAKER_00: A coding agent that can read files, run commands, and connect to internal systems is a very different animal from a chatbot that answers questions.
• 4:16
  SPEAKER_01: That distinction is the heart of it for me.
• 4:19
  SPEAKER_01: A chat assistant gives you text, and a human decides what to do with it.
• 4:24
  SPEAKER_01: An agentic coding tool actually takes actions on a machine, and when the person supervising it cannot read the code, the human in the loop becomes more of a spectator than a supervisor.
• 4:34
  SPEAKER_00: There is also some history here that suggests this is part of a pattern rather than a one-time reaction.
• 4:40
  SPEAKER_00: Earlier this year, Goldman restricted staff in Hong Kong from using Claude, although that one was specific to that location.
• 4:48
  SPEAKER_00: So the bank has been drawing these boundaries for a while.
• 4:50
  SPEAKER_01: And let us not forget, Goldman is hardly an AI skeptic.
• 4:55
  SPEAKER_01: It built its own internal assistant that combines models from OpenAII, Anthropic, and Google.
• 5:00
  SPEAKER_01: So this is not a company retreating from the technology.
• 5:03
  SPEAKER_01: It is a company deciding which version of the technology belongs in which hands.
• 5:08
  SPEAKER_00: There is a delicious irony in the business ties too.
• 5:11
  SPEAKER_00: Goldman is the lead underwriter for the planned anthropic initial public offering.
• 5:15
  SPEAKER_00: So the bank is helping sell the company to investors while also telling most of its own staff they cannot use one of its flagship products at work.
• 5:22
  SPEAKER_01: I would not read too much into that, honestly.
• 5:25
  SPEAKER_01: A bank can believe a company is a great investment and still believe its own traders should not be shipping code.
• 5:31
  SPEAKER_01: Those are two separate judgments, and frankly, I would be more worried if the bank blurred them.
• 5:35
  SPEAKER_00: That is fair.
• 5:37
  SPEAKER_00: But it does tell you something about the market signal.
• 5:39
  SPEAKER_00: Anthropic has been pitching Claude Code as a tool for everyone, not just developers.
• 5:44
  SPEAKER_00: And a marquee customer saying, actually, only our engineers is a data point that other enterprise buyers are absolutely going to notice and quote in their own meetings.
• 5:53
  SPEAKER_01: Let me bring in the cost side because I think that is the piece that gets underplayed.
• 5:57
  SPEAKER_01: These agentic tools are not cheap to run.
• 6:00
  SPEAKER_01: They read huge amounts of context, they loop, they retry, and a curious trader exploring an idea can burn through a lot of expensive tokens very quickly.
• 6:09
  SPEAKER_00: And Wall Street has been thinking hard about that.
• 6:11
  SPEAKER_00: Bloomberg pointed out that JP Morgan Chase said in July it wanted employees to match models to tasks, and specifically not to use costly models for simple jobs like summarizing an analyst report.
• 6:23
  SPEAKER_00: That is a cost discipline message dressed up as a productivity message.
• 6:27
  SPEAKER_01: Which makes sense to me.
• 6:29
  SPEAKER_01: You would not send a senior structural engineer to change a light bulb, and you probably should not point the most powerful coding agent on the market at a task that a cheap model or a simple template could handle in seconds.
• 6:40
  SPEAKER_00: Now here is the contrast I really wanted to get to.
• 6:43
  SPEAKER_00: Because another financial company is going in a completely different direction.
• 6:48
  SPEAKER_00: Revolut, the European fintech, reportedly moved about 2,000 engineers off metered commercial models and onto open models that it hosts itself.
• 6:56
  SPEAKER_01: Okay.
• 6:57
  SPEAKER_01: Now that is interesting because it attacks the same problem from the opposite end.
• 7:01
  SPEAKER_01: Goldman is narrowing who gets the tools.
• 7:04
  SPEAKER_01: Revolute is apparently keeping the tools broad for its engineers, but changing what is under the hood so the meter stops spinning.
• 7:11
  SPEAKER_00: Right.
• 7:13
  SPEAKER_00: According to a conference talk summarized by Startup Hub, Revolute runs a model called GLM 4.5 on 56 NVIDIA B200 chips, and it claims its AI coding costs fell by about 90% to roughly 30 cents per million tokens.
• 7:30
  SPEAKER_01: I want to be careful with those numbers, because a talk at a conference is not an audited financial statement.
• 7:35
  SPEAKER_01: And the 30 cent figure reportedly leaves out cloud spending, staff time, and observability, which in my experience is exactly where the real money disappears.
• 7:45
  SPEAKER_00: Completely agree, and the article itself flagged that.
• 7:48
  SPEAKER_00: But the reported productivity numbers are striking.
• 7:51
  SPEAKER_00: The claim is about a 60% improvement in coding time and a 70% improvement in lead time.
• 7:57
  SPEAKER_00: Across a sample of roughly 2,000 engineers, once tokens stopped being rationed.
• 8:02
  SPEAKER_01: And notice the guardrails they describe.
• 8:04
  SPEAKER_01: Every agent run is supposedly human in the loop, inside a sandbox with an all-alo list, so a command like adding files to version control is permitted, while broader destructive commands are blocked.
• 8:16
  SPEAKER_01: That is real engineering discipline, not a hope and a prayer.
• 8:20
  SPEAKER_00: Which brings us right back to Goldman.
• 8:22
  SPEAKER_00: Revolute is handing these tools to engineers who can read what the agent is doing.
• 8:27
  SPEAKER_00: Goldman just decided the same thing.
• 8:29
  SPEAKER_00: Neither company is saying that a banker with no software training should be running autonomous agents against internal systems.
• 8:35
  SPEAKER_01: So maybe the headline is not really Goldman pulls AI tools.
• 8:40
  SPEAKER_01: Maybe the more accurate headline is that the era of everyone becomes a programmer is colliding with the reality that somebody still has to own the software after the demo ends.
• 8:49
  SPEAKER_00: I like that framing.
• 8:50
  SPEAKER_00: But I want to defend the traders for a second.
• 8:53
  SPEAKER_00: Some of the work they were doing, like modeling positions or tracking client flows, is genuinely valuable.
• 9:00
  SPEAKER_00: And if you take the tool away without giving them a safer alternative, you just push them back to spreadsheets and email.
• 9:06
  SPEAKER_01: That is a really good point.
• 9:08
  SPEAKER_01: And it is the classic shadow technology problem.
• 9:11
  SPEAKER_01: When you ban a useful tool without offering a sanctioned path, people do not stop needing the capability.
• 9:17
  SPEAKER_01: They quietly go find a worse version of it on a personal laptop, and now you have even less visibility.
• 9:23
  SPEAKER_00: So the smart move, and I suspect Goldman knows this, is to give the front office a governed lane.
• 9:28
  SPEAKER_00: Maybe that means the internal assistant.
• 9:31
  SPEAKER_00: Maybe it means request-based builds where an engineer partners with a banker, but it needs to be faster than filing a ticket and waiting a month.
• 9:38
  SPEAKER_01: And that is the practical takeaway for anyone running a technology team.
• 9:42
  SPEAKER_01: Decide which of your people are allowed to let AI take actions, not just give answers.
• 9:47
  SPEAKER_01: Those are two very different risk categories, and most companies are still treating them as one big bucket labeled AI tools.
• 9:55
  SPEAKER_00: There is a workforce angle, too, that I do not want to skip.
• 9:58
  SPEAKER_00: Earlier this year, Goldman executive John Waldron described the bank as a human assembly line that was ripe for automation.
• 10:05
  SPEAKER_00: So this restriction does not mean the bank is slowing down on automation.
• 10:10
  SPEAKER_00: It means it wants automation built by specialists.
• 10:12
  SPEAKER_01: Which actually raises the value of those 12,000 engineers.
• 10:16
  SPEAKER_01: If the bank believes AI multiplies what a skilled engineer can do, then the engineer becomes the force multiplier, and the trader becomes the customer who describes the problem rather than the person who writes the fix.
• 10:28
  SPEAKER_00: And that's a second-order effect worth watching across the industry.
• 10:32
  SPEAKER_00: For two years, the narrative has been that AI makes technical skills less important.
• 10:37
  SPEAKER_00: Goldman is quietly betting the opposite: that technical skill is what makes these tools safe enough and valuable enough to deploy at scale.
• 10:45
  SPEAKER_01: I find that refreshing, to be honest.
• 10:47
  SPEAKER_01: A lot of the hype assumed that the hard part of software was typing.
• 10:52
  SPEAKER_01: The hard part was always understanding the system, the failure modes, and the people who depend on it.
• 10:57
  SPEAKER_01: And an AI agent does not remove the need for that understanding.
• 11:00
  SPEAKER_00: Let me play devil's advocate against both of us, though.
• 11:03
  SPEAKER_00: The tools are improving very fast.
• 11:05
  SPEAKER_00: Six months from now, the agents may be good enough at explaining and testing their own work that a careful banker really could supervise them safely.
• 11:13
  SPEAKER_00: Goldman might be locking in a rule for last year's technology.
• 11:16
  SPEAKER_01: That is possible, and I think the spokesperson left the door open on purpose with that language about determining the best use cases.
• 11:23
  SPEAKER_01: Policies like this tend to get revisited.
• 11:26
  SPEAKER_01: The question is whether the next review is driven by better tools, better controls, or simply competitive pressure from rival banks.
• 11:33
  SPEAKER_00: And competitive pressure is real.
• 11:36
  SPEAKER_00: If a rival bank figures out how to let its traders build safely and sees a measurable edge, Goldman will notice very quickly.
• 11:43
  SPEAKER_00: Wall Street does not leave money on the table for long, and that includes productivity gains from software its staff could be using.
• 11:50
  SPEAKER_01: So if I am a chief information officer listening to this on my commute, what is the one thing I should do differently this week because of what Goldman just did, other than panic and revoke everyone's access before lunch?
• 12:01
  SPEAKER_00: I would say run your own honest review.
• 12:03
  SPEAKER_00: Find out who is actually using a genic tools, what they are touching, and what it costs.
• 12:08
  SPEAKER_00: Then separate the people who need answers from the people who need actions and give each group a governed path that is genuinely easier than going around it.
• 12:16
  SPEAKER_01: And I would add one thing.
• 12:17
  SPEAKER_01: Measure the review burden on your engineers.
• 12:21
  SPEAKER_01: If every citizen-built prototype ends up in somebody's queue, you are not saving engineering time.
• 12:27
  SPEAKER_01: You are just moving it.
• 12:28
  SPEAKER_01: And that cost needs to show up on the same spreadsheet as the productivity gains.
• 12:32
  SPEAKER_00: That is the lesson for me as well.
• 12:34
  SPEAKER_00: The story is not that AI coding tools failed at Goldman.
• 12:38
  SPEAKER_00: The story is that one of the most sophisticated buyers in the world looked at the full bill, including risk and support, and decided that access has to follow accountability.
• 12:47
  SPEAKER_01: Access follows accountability.
• 12:50
  SPEAKER_01: I think that is a pretty good rule for the whole agent era, and probably one we will be repeating on this show a lot more as these tools move from writing suggestions to taking actions inside real businesses.
• 13:01
  SPEAKER_00: Agreed.
• 13:02
  SPEAKER_00: And I will be watching whether other banks follow Goldman or whether someone like Revolute proves that broad access with tight engineering guardrails is the better bet.
• 13:11
  SPEAKER_00: Either way, the debate has moved from whether to use these tools to who should be allowed to drive them.
• 13:17
  SPEAKER_01: That's all for today's episode of the DX Today podcast.
• 13:20
  SPEAKER_01: Thanks for listening, and we'll see you next time.