What Happens When Banks Can License AI Instead of Hiring You?
If you’re an investment banking analyst—or planning to become one—here’s what you need to know: OpenAI has assembled over 100 former bankers from JPMorgan, Goldman Sachs, Morgan Stanley, Evercore, and KKR. They’re paying them $150/hour. And they’re systematically teaching AI to build the exact financial models you build every day.
This isn’t a productivity tool. This isn’t something to “augment” your work. They’re training AI to receive instructions from a senior banker, construct a properly formatted Excel model for an LBO or restructuring, iterate based on feedback, and deliver a finished product ready for client presentation.
In other words: They’re building a system to do your job without you.
The project is called Mercury. And based on what Bloomberg revealed on October 21, 2025, it’s a lot further along—and moving a lot faster—than most people in the industry realize.
Here’s why this is different from every other “AI is coming for banking” story you’ve read: The specificity. OpenAI isn’t building experimental technology. They’re building something that can slot into existing bank workflows tomorrow. Models must follow exact industry formatting. Specific margin sizes. Italicized percentages. Proper Excel structure. The same reviewer feedback loops that analysts go through right now.
And here’s the timeline that should terrify you: Contractors already have early access to test the AI. That means it’s not in early research—it’s in late-stage product development. With 100+ expert trainers submitting models weekly, OpenAI is accumulating hundreds of training examples monthly. At this pace, they could have a production-ready system within 6-9 months of focused training.
Now add OpenAI’s financial pressure: $500 billion valuation, no profit yet, desperate need for enterprise revenue. Add the market demand: Goldman just cut 1,000 positions explicitly for AI replacement during their best year ever. JPMorgan restructuring to 4-to-1 junior ratios. Every major bank actively shopping for exactly this technology.
Put it together and you’re not looking at a 3-5 year horizon. You’re looking at 12-18 months before major banks start deploying this.
Why Mercury Is Different From Every Other Banking AI
Most AI tools in banking have been assistive—helping with parts of the workflow, speeding up specific tasks, augmenting what analysts do. Mercury is something else entirely.
OpenAI is replicating the complete analyst workflow:
- Receive instruction from senior banker (“build an LBO model, 8x EBITDA, 40% debt, model out 5 years”)
- Construct multi-tab Excel model with proper formatting
- Receive feedback and iterate (“sensitize the exit multiple from 6x to 10x”)
- Deliver final product ready for client deck
That’s not automation. That’s replacement.
And the training process reveals how serious this is. Contractors must follow precise industry standards—the same margin sizes, the same formatting conventions, the same iterative refinement process that analysts learn in their first weeks. OpenAI isn’t building something experimental that requires banks to change their processes. They’re building something that drops directly into how banks work today.
The application process itself is telling: AI chatbot interview, financial statements test, modeling assessment. OpenAI has already built the infrastructure to evaluate whether someone understands banking and can produce quality work. That same infrastructure evaluates whether their AI is producing quality output.
This is late-stage product development dressed up as a training program.
The Math That Makes This Inevitable
A first-year analyst: $200,000 all-in, 80-100 hour weeks, building maybe 30 models annually.
Mercury training data: $150/hour per contractor, one model weekly. Roughly $1,500-2,250 per model for training.
Mercury at scale: A bank pays OpenAI perhaps $75,000 per seat annually. Generates unlimited models. Marginal cost per additional model: essentially zero.
If one VP with Mercury access can supervise what previously took three analysts, the bank either cuts two-thirds of headcount or triples deal capacity per team. Either way, profitability per employee transforms overnight.
Goldman spending $18 billion on technology suddenly makes perfect sense—it pays for itself in quarters, not years.
The Compensation Question No One Wants to Answer
We’ve already covered how AI might compress analyst compensation in previous articles. But Mercury creates a specific new pressure point: What happens when every bank can license the exact same AI trained by the exact same expert bankers?
Previous banking technology gave competitive advantages to whoever built it first or implemented it best. Mercury will likely be available to all major banks simultaneously through OpenAI licensing. No technological moat. No implementation advantage. Just: do you adopt or not?
This creates a race to the bottom on analyst headcount and compensation because no bank gets a lasting edge from the technology. If Goldman deploys Mercury and cuts 50% of analysts, JPMorgan faces immediate pressure to match just to maintain cost parity. If JPMorgan then drops new analyst compensation 25%, Goldman has to decide whether to match or accept higher costs for the same work.
The only differentiation becomes: who’s willing to cut deepest, fastest. That’s not a recipe for stable analyst compensation.
The Question That Should Terrify Banks: Can You Judge What You Didn’t Build?
Investment banking has always run on apprenticeship. You build hundreds of models until the patterns become instinctive. You learn what drives value by modeling it repeatedly. You develop judgment by making mistakes at 3am and fixing them.
But Mercury upends this model in a specific way: If AI builds your first 500 models and you just review them, do you develop the same instincts as someone who built them from scratch?
Goldman’s CEO says AI can handle 95% of an S-1 and “the last 5% now matters because the rest is a commodity.” That assumes the person handling the 5% understands the other 95%. But how do you develop that understanding without building the 95% yourself?
This isn’t abstract. In 18 months, if you’re recruiting for private equity, the interviewer asks: “Walk me through how you’d model this LBO.” You’ve spent two years reviewing Mercury’s output, fixing edge cases, reformatting for client preferences. Can you build a complex LBO model on a blank Excel sheet and defend every assumption?
Maybe you can. Maybe AI supervision is actually better training because you see more models faster and focus on judgment over mechanics. But that’s a massive assumption banks are making with no evidence yet.
And here’s the uncomfortable part: We won’t know if this assumption is wrong for 5-7 years, when banks need to promote people who learned primarily through AI supervision. By then, if there’s a gap, it’s too late to fix.
What Mercury Changes About Bank Strategy
We’ve written before about how different banks might respond to AI. But Mercury specifically creates a new dynamic: elite boutiques can license the exact same technology as Goldman.
Previous banking technology required massive infrastructure investment. Boutiques couldn’t compete with JPMorgan’s $18 billion annual technology budget. But if Mercury is available through OpenAI subscription, an Evercore or Lazard analyst gets access to the same AI trained by the same expert bankers as a Goldman analyst.
This changes the strategic calculus. Boutiques can’t out-spend bulge brackets on technology development. But they can out-train on human judgment. If Mercury becomes commodity infrastructure available to everyone, the differentiator shifts back to: who develops better bankers through their training program?
That actually advantages boutiques—if they execute. Smaller training classes, more partner attention, traditional apprenticeship alongside AI tools. If bulge brackets are betting on AI supervision as sufficient training and boutiques bet on traditional foundations plus AI augmentation, we’ll find out in 3-5 years which approach produces better senior bankers.
But here’s the catch: Boutiques need to articulate this strategy explicitly and recruit on it. “We maintain traditional training while using AI to improve work-life balance” is a compelling pitch to candidates worried about skill development at bulge brackets. But only if boutiques actually deliver on it, not just use it as recruiting rhetoric.
Three Timelines for How Mercury Plays Out
Fast deployment (40% probability): OpenAI ships production-ready Mercury by Q3 2026. Goldman and JPMorgan adopt immediately. By end of 2027, analyst classes at bulge brackets shrink 50-60%. Compensation for new hires drops 20-30%. The transformation happens faster than almost anyone expected. Banking in 2028 looks fundamentally different—smaller analyst classes, lower compensation, different skill requirements. The career path still exists but the value proposition has changed dramatically.
Phased rollout (45% probability): Mercury deploys late 2026/early 2027. Banks pilot cautiously before scaling. Headcount and compensation adjust gradually over 3-4 years rather than overnight. Analysts work on more deals simultaneously rather than working fewer hours. The role evolves similar to how consulting transformed—still prestigious, still demanding, but different. By 2029, banking looks different but maintains its position.
Something breaks (15% probability): Technical limitations, regulatory concerns, or high-profile mistakes slow adoption. Mercury works for routine deals but struggles with complex situations. Banks implement cautiously. The transformation takes 5-7 years instead of 18 months. Or boutiques that maintained traditional training start winning complex mandates because their bankers have deeper expertise, creating market segmentation that slows AI adoption.
The most likely outcome is somewhere between the first two. The technology will work well enough for broad adoption. The question is whether it happens in 18 months or 36 months, and whether banks successfully evolve training models or create quality gaps they don’t discover until it’s too late.
What Current Analysts Should Do About Mercury Specifically
Test yourself without AI assistance right now. Open a blank Excel sheet. Build a three-statement LBO model from scratch. No templates. No macros. No AI prompts. Time yourself. If you struggle or take significantly longer than you did as a first-year, you have a problem. Your PE interview in 12 months won’t have AI assistance.
Track your next analyst class compensation carefully. If the 2026 class gets offered 15-20% less than you made, that’s not about them—it’s about where the market is heading. Your future raises and bonuses will reflect that new baseline, not your current comp. Plan accordingly.
Ask your group directly about AI strategy. Not “are you using AI?” Everyone is. Ask: “What percentage of our models will AI generate first draft within 12 months? How is that changing what analysts learn?” Banks that can’t articulate a clear answer are winging it, and you’re the experiment.
Consider timing on lateral moves. If you’re thinking about moving to a boutique for better training, doing it before Mercury fully deploys gives you more traditional skill development. Waiting until after means you’ll interview having primarily done AI supervision—which is a weaker profile for boutiques hiring on training culture.
What Prospective Analysts Should Know About Mercury
The recruiting timeline matters. If you’re recruiting for 2026 positions, Mercury probably won’t be fully deployed at most banks when you start. You might get 6-12 months of relatively traditional training before things change. If you’re recruiting for 2027 positions, Mercury will likely be live at bulge brackets by the time you arrive. That’s a fundamentally different analyst experience.
Ask about AI deployment explicitly in interviews. “When do you expect to deploy AI tools like OpenAI’s banking models? What percentage of analyst work do you estimate AI will handle by the time I’m a second-year? How are you adapting training for that?” Banks that dodge these questions are either dishonest or unprepared. Neither is good.
Verify exit opportunities with recent alums, not banks. Don’t ask Goldman whether their analysts still place well at PE funds. Ask second-years who just recruited how many of their friends got offers, and whether interviewers asked about AI use. The data you want is whether PE firms are already changing how they evaluate AI-trained analysts.
Reconsider compensation assumptions entirely. Do not make a decision to go into banking based on $200K+ analyst compensation. That number might not exist by 2027. If you need that compensation for loans, lifestyle, or family obligations, you need a plan B. If you’re choosing banking over tech or consulting primarily for money, get current data on what 2027 offers actually look like—don’t assume 2024 numbers hold.
Why This Time Is Actually Different
Every few years there’s a story about how technology will disrupt investment banking. Most of them fizzle. Banks adopt new tools, workflows adapt, but the fundamental analyst role persists.
Here’s why Mercury is different:
It’s being built by actual bankers. Not engineers who think they understand finance. Not generic AI researchers. 100+ people who worked at JPMorgan, Goldman, Morgan Stanley, Evercore, KKR. People who know exactly what analysts do because they did it themselves. They’re teaching AI the precise workflows, formatting standards, and iterative processes that define the job.
It’s already in late-stage testing. Contractors have early access. This isn’t vaporware or research. This is product development with user testing.
The market is ready to buy. Goldman cutting 1,000 people during their best year ever. JPMorgan restructuring ratios. Every major bank explicitly shopping for this technology. OpenAI doesn’t need to create demand—they just need to deliver product to buyers who are waiting.
The economics are overwhelming. $200K analysts versus $75K software subscriptions. Banks that don’t adopt face 40-60% cost disadvantage against competitors who do. There’s no scenario where banks collectively decide not to use this technology.
The timeline is compressed. Not 5-7 years. Not 3-5 years. 12-18 months for initial deployments. Maybe 24-36 months for widespread adoption.
If you’re an analyst who started in 2024 expecting a traditional two-year program followed by standard exit opportunities and compensation progression, Mercury could upend all of that before you recruit for your next job.
If you’re recruiting for 2026 or 2027, you’re potentially signing up for a fundamentally different experience than the people who recruited just 2-3 years before you—without fully understanding what that difference means for your skills, compensation, and career options.
The question isn’t whether Mercury will impact investment banking. The question is whether you’re planning for the world where it succeeds—which based on the timeline, market demand, and economics looks increasingly likely—or whether you’re assuming the traditional analyst path will persist because it always has.
Because this time, the traditional path might not persist. And the window to plan for that reality is a lot shorter than most people think.