Artificial intelligence in investment banking: Machines read, humans decide – who really wins?

AI in investment banking: These are the tasks at which the algorithms (still) fail miserably

Creative deal structures, multi-million-dollar transactions, and visionary judgment – ​​that's how investment banks like to market their services. But the reality for junior analysts, away from the glossy presentations, often looks quite different: endless nights in virtual data rooms, manually sifting through hundreds of contract PDFs, and painstakingly transferring figures into complex Excel models. It is precisely in this discrepancy between expensively paid expertise and tedious information gathering that artificial intelligence (AI) is currently unfolding its greatest potential.

While Wall Street giants like JPMorgan and Goldman Sachs have long been investing tens of billions of dollars in AI systems to free their analysts from mere tedious tasks, reality paints a stark picture: Not every highly praised technology can withstand the relentless pace and pressure of a real M&A deal. Moreover, automation raises profound questions: If algorithms take over research in the future, how will the next generation of bankers learn their craft? And at what point will an experienced dealmaker even trust a machine's analysis? This article unflinchingly examines which AI applications actually work in the highly regulated financial sector, where automation must be deliberately halted, and why blind faith in technology without sound evidence quickly becomes a massive liability risk.

Evidence instead of rumors: When top bankers truly trust the results of artificial intelligence

Investment banks officially sell judgment, relationships, and creative deal structures. However, a look at the actual daily work of young analysts reveals a completely different picture: sifting through data rooms, painstakingly transferring figures from PDF documents into Excel models, searching for old presentations that the bank created three years ago for a similar client.

This discrepancy is precisely where the core of the problem lies. Investment banking sells judgment, yet spends the vast majority of junior staff hours simply gathering information. This gap between what clients actually pay for and what working hours are actually spent on is exactly where the use of artificial intelligence in investment banking becomes economically viable.

Against this backdrop, it's worth taking a closer look at how financial institutions can actually leverage AI for tangible business benefits. The short answer is this: the banks that win here aren't those with the boldest strategy presentations. They are the ones that have consistently deployed machines to read and search documents, while leaving the actual judgment with humans and operating the entire process within their own controlled system boundaries.

The invisible price of simply gathering information

The reason no one is addressing this problem is that it doesn't appear anywhere as a separate cost item. Salaries are already budgeted, analysts have already been hired and assigned. Reading and reviewing documents is precisely what the labor hours were originally purchased for – the effort therefore remains as invisible as the rent once the lease is signed.

However, when you consider the actual cost of this hidden effort, the problem quickly becomes apparent. It leads to spot checks instead of comprehensive reviews, because thoroughly reading every single contract in a data room was never financially feasible anyway. It results in outdated company profiles, because updating forty profiles per quarter regularly loses out in the competition with ongoing, urgent transactions. And it leads to promising deals not even being pursued, because the team is already completely occupied with current transactions. Each of these decisions is effectively made by a time constraint, not by a banker personally.

The volume aspect of this problem is becoming increasingly acute. Financial sponsor-driven activity is growing in tandem with the growth of the private capital market, with Preqin analysts predicting global assets in alternative asset classes will reach approximately $32 trillion by 2030. This translates to more transaction processes, more competitive bidding processes, and more due diligence per banker than any staffing plan can realistically accommodate. Banks that continue to accept this hidden expense will find themselves paying for it more heavily each year. Conversely, those who can expand their coverage without increasing their cost base are those who have stopped viewing the tedious sifting through documents as a kind of initiation rite for junior analysts.

Which AI applications actually stand up to practical testing?

Not all use cases are equally ready for productive deployment, and the order in which they are addressed is more important than the technical choice of software itself. Three filter criteria help with quick classification.

First and foremost is volume. You should choose a batch of documents that accumulates constantly and regularly, rather than just occasionally, because a workflow that's only run twice a year will never gather enough experience to build genuine trust. Second is repeatability: the same information fields, always queried in the same way, across documents that differ in wording but hardly in structure. Third and most important is verifiability. You want to obtain a result that someone can compare against the source within seconds, because it's precisely this verification that ultimately convinces skeptics.

Applying these three filters results in almost every bank having the same short priority list. Data rooms score highest, which is why the intelligent analysis of due diligence documents is usually implemented first. Internal precedent documents follow in second place, as the bank's own history is vast, endlessly tedious to search, but trivially verifiable once located. Regulatory filings for ongoing market monitoring come in third. Anything whose result represents an opinion or assessment is automated last – more on that later.

When does an experienced banker trust the machine's results?

Essentially, it all comes down to one crucial characteristic: every claim must be supported by its source, and verifying that source must take no more than ten seconds. This might sound like a minor detail in product design, but it's actually the entire product promise.

An analysis claiming that the liability limit in the contract is 12 percent of the purchase price is initially just an assertion. An analysis that states the same thing and refers directly to page 84 of the third contract amendment, however, is proof. The first requires blind trust, the second practically invites verification – and that is precisely what a Vice President wants to do with any information concerning an ongoing transaction.

One can observe how trust develops quite naturally under these conditions. The vice president reviews the first twenty results, and all twenty pass scrutiny. The next time, only ten out of fifty results are checked. On the second deal processed, he only checks the results that surprise him and accepts the rest without further review. No one officially sanctioned this change in behavior—it simply arose because the reviews consistently delivered clean results, and no software vendor's sales pitch could ever have conjured up this level of trust.

Unfortunately, the opposite is just as reliable. A single, confidently worded, but incorrect answer without a verifiable source can set a project back further than a whole month of system downtime, because the story of this error spreads throughout the entire organization. The entire processing chain, from document to decision, stands or falls on the citable nature of the sources, and a system that provides answers without this evidence ultimately produces nothing but rumors.

Where must automation deliberately reach its limits?

Every reputable AI program needs a clearly defined list of what should not be automated – ideally established before the first pilot project. Machines provide evidence, humans provide positions and assessments. This boundary list is where this distinction is explicitly drawn, instead of being painfully discovered in the middle of an ongoing, critical process.

One piece of evidence can be verified: that every single contract in the data room contains a change-of-control clause, that a comparable company is substantiated with every single key performance indicator and supported by sources, or that the exact wording of an exclusivity clause, including its expiration date, is accurately reproduced. A position, on the other hand, is an assessment based on a context that no single document can capture: whether a particular change-of-control clause actually presents an obstacle in practice or is more of a negotiating point, what a company is really worth, or which of two bidders a client should ultimately award the contract to and how to deal with the losing bidder afterward.

Programs that ultimately fail almost always start by tackling positional problems, usually valuation models or direct transaction recommendations, based on the logic that the most valuable work deserves the earliest attention. The result, however, was always the same: Experienced bankers simply refuse to rely on conclusions they cannot critically examine themselves. And this refusal is justified for good reason. In these cases, the problem wasn't the technology, but the incorrect choice of the initial use case.

This boundary does shift over time, and it shifts just as trust develops: step by step, one verified result after another. Such a boundary list should therefore be reviewed and adjusted annually, not remain set in stone. What it actually protects against is the far more frequent failure caused by mere assertion – when the boundary is simply shifted in a presentation, in front of people who, in case of doubt, simply decide quietly not to abide by it.

Managed AI Platform - Image: Xpert.Digital

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

• Managed AI Platform

What happens when multiple deals arrive simultaneously?

A consistent throughput on quiet days is ultimately a meaningless metric. Deal teams never collapse during average weeks – they collapse when a third customer calls with an urgent, live situation, and at the same time, a seasoned executive is on a plane and unavailable.

It is precisely at this point that the sample size decreases, secondary due diligence questions remain unanswered, and the risk of overlooking a critical finding is highest because attention is most diluted. Every tool worth buying must prove itself under this exact pressure.

Therefore, this is precisely what you should test for. A sudden influx of fifty contracts under time pressure reveals more about the quality of a system than a whole month of comfortable use in normal operation. Three things are crucial to observe: whether accuracy decreases with increasing volume; whether the queue delivers usable partial results or simply forces the entire team to wait; and whether the system openly displays what it itself could not determine with certainty.

Under such pressure, a sensible system for marking uncertainties is even more important than a pure hit rate. A system that delivers 400 unambiguous results and 30 cases explicitly marked as uncertain provides the team with a clear worklist. Conversely, a system that delivers 430 seemingly unambiguous results, 30 of which are actually wrong, presents the team with a real liability risk. Under stressful conditions, no one simply has the time to retrospectively determine which of the two approaches they actually adopted.

How can sensitive documents be kept within the bank's own walls?

This question brings most AI programs to a halt in practice – and usually only late in the process. A team finds a tool that analyzes contracts well and initially tests it on publicly available company announcements where there is no real risk. The results are convincing, so someone suggests applying the system to a real, live data room. This immediately triggers a security audit.

The software provider offers contractual assurances: no training of the model with the company's own data, deletion after thirty days, and a customized corporate contract is also available. However, the compliance department asks the much simpler and more crucial question: where are the documents actually physically transferred to, and who else has access to them? The honest answer is usually: to a shared server outside the bank – and rightly so, that's where the conversation ends.

Price-sensitive insider information, confidential customer data, documents subject to confidentiality agreements, and the information barriers between departments, scrutinized by regulatory authorities, make this end inevitable. The real error lay not in the pilot project itself or in the justified caution of the compliance department, but in the choice of a deployment model that could never have overcome this hurdle from the outset.

The model that ultimately prevails keeps every evaluation and every conclusion consistently within the bank's own system boundaries. With its isolation of each individual transaction, it precisely reflects the prescribed information barriers, enforces access controls directly with every query, and maintains a complete log of every single interaction. Viewed in this way, the compliance department itself becomes an advocate for the project rather than an obstacle, because the control level answers the questions of the second line of defense even before they need to be asked. Making this fundamental decision before the first pilot project ultimately saves around six months.

What the big houses actually spend and invest

The abstract discussion surrounding retrieval management and trust building becomes concrete when one looks at the actual investment sums and rollouts of the largest institutions. JPMorgan now operates its in-house LLM Suite for more than 200,000 employees, funding over 400 documented AI use cases with a technology budget of approximately $18 billion for 2026. Goldman Sachs, after a pilot phase with around 10,000 employees, has rolled out its GS AI Assistant, which is based on models from OpenAI, Google, and Anthropic behind its own security layer, to approximately 46,000 employees company-wide and reports productivity gains of over 20 percent in software development. Morgan Stanley, on the other hand, has already achieved a usage rate of over 98 percent of all advisory teams with its AI assistant for financial advisors and has been testing digital assistants since the summer of 2026 that can interact with clients around the clock, with the decision-making authority over actual portfolio decisions explicitly remaining with humans.

A striking feature of all three cases is a common architectural pattern: availability and decision-making authority are deliberately separated. The machine is permitted to analyze, summarize, and suggest at any time, but the authority to actually act remains contractually and technically excluded. This separation aligns precisely with the previously described boundary between verifiable facts and genuine entrepreneurial judgment and empirically confirms that the institutions with the greatest rollout success adhere to exactly this principle.

From a macroeconomic perspective, this is no longer a niche phenomenon. Goldman Sachs estimates that total AI-related investments in the United States will approach nearly $600 billion in 2026, equivalent to roughly 2 percent of US GDP, while global AI investments are expected to exceed $1 trillion for the first time. At the same time, the same analysis warns that approximately two-thirds of these expenditures will be financed by reallocations within existing corporate budgets, indicating a real, albeit still moderate, crowding-out effect on other investments. For investment banks, this means that competitive pressure to invest in AI themselves is growing, not only for efficiency reasons, but also because their own client base, particularly technology companies and their bond issuances, is already bringing the topic into every capital markets advisory.

Regulatory requirements that define the framework

Parallel to the investment boom, the regulatory landscape for operating these systems is also shifting. The European AI Act originally stipulated that high-risk applications in the financial sector, such as creditworthiness assessments of individuals, must be fully compliance-tested, documented, and subject to mandatory human oversight by August 2, 2026. However, the so-called Digital Omnibus on AI postponed this deadline by sixteen months to December 2, 2027, in the summer of 2026, because the necessary technical testing standards simply weren't available in time. For core investment banking activities—M&A advisory, capital market issuances, and due diligence—this high-risk classification is largely irrelevant, as these primarily focus on lending, insurance pricing, and biometric applications. Nevertheless, a fundamental transparency obligation has been in place since August 2026: Customer communication that is supported or created by AI systems must be identifiable as such, which is also relevant for automatically generated analyses and customer correspondence from banks.

In the United States, the Securities and Exchange Commission (SEC) pursues a technology-neutral, but significantly stricter, audit approach. The audit priorities for 2026 explicitly identify AI systems in investment advice, compliance processes, and customer communication as a key area of ​​focus. Institutions must demonstrate the optimization goal pursued by each deployed AI system and whether this could create conflicts of interest between the bank and the customer. Contract reviews with external AI providers are also receiving greater attention, particularly regarding whether customer data is used to train third-party models. This corresponds precisely to the compliance obstacle described in the article concerning externally hosted data room solutions. Both regulatory regimes are thus moving in the same direction: not the prohibition of AI, but rather the mandatory documentation of control, data origin, and ultimate human responsibility.

The uncomfortable downside: staff reductions and skepticism

The narrative of AI as a pure efficiency gain ignores an uncomfortable reality that is being discussed with increasing openness. According to McKinsey partner Debasish Patnaik, banks are already cutting their entry-level analyst programs by up to two-thirds, while at the same time, roughly 62 percent of the required AI specialists are recruited from precisely these same junior ranks. This raises a structural question that the original article deliberately omits: If the traditional training of young bankers relied on the very tedious reading work that is now being automated, where will the next generations of experienced decision-makers come from if they are deprived of the training ground of information gathering? This question remains largely unanswered in the current debate and is likely to prove to be one of the real risks of automation in the coming years, extending far beyond mere data privacy or accuracy concerns.

Furthermore, there is a methodological caution, which Goldman Sachs itself acknowledges: The macroeconomic contribution of the AI ​​investment wave to the gross domestic product and the associated displacement effects on other investments are, overall, smaller than is often assumed in public debate. For individual banks, this means: Those who rely solely on current trends and make sweeping AI announcements without adhering to the discipline described in the article regarding automation limits, source verification, and data sovereignty ultimately risk precisely the loss of trust that a single false, unsubstantiated result can trigger.

Why real returns in investment banking don't come from AI show effects

The economic benefits of artificial intelligence in investment banking don't arise from spectacular announcements or the promise of fully autonomous transaction processing. They arise where banks have the courage to honestly distinguish between verifiable facts and genuine business judgment, and consistently translate this distinction into systems, processes, and access rules. The actual figures from JPMorgan, Goldman Sachs, and Morgan Stanley demonstrate that those institutions achieving the greatest productive rollout are precisely those that strictly separate availability and decision-making authority and understand regulatory documentation requirements as a tool rather than an obstacle. The institutions that take this path will not be more successful because their technology appears more impressive, but because they free their bankers' actual value creation from the time trap of mere information gathering and redirect it back to what customers are truly willing to pay for: sound, transparent, and ultimately human judgment.

Konrad Wolfenstein

I would be happy to serve as your personal advisor.

You can contact me at wolfenstein∂xpert.digital or

Just call me on +49 7348 4088 965 .

LinkedIn