Transcription
Hello. Today we are diving into a program that, well, many consider the most boring on their computer, into the world of Excel financial spreadsheets. But what if I told you that right there, among all these cells and formulas, a real revolution is unfolding right now? We will be analyzing how the latest artificial intelligence model, Cloud 45 OPUS, is changing not just the speed of work, but the very essence of financial analysis. Our task today, relying on the presentation materials we have, is to understand how we are moving from manually entering numbers to neural reasoning. To grasp the scale, let's go back in time for a second. What immediately caught my attention in these materials was a very visual timeline. The eighties, USUS 1 2 3, everything manual. The two-thousands, macros, VBA, some automation, and the forecast for 2026 sounds like something completely different. Cloud 45 opus and neural reasoning. What does reasoning even mean in the context of Excel? It's just a grid of numbers. But therein lies the fundamental shift. Previously, for Excel, a model was, well, just a set of cells. Numbers, formulas, they were not connected in meaning. If something broke, the system didn't understand why. But now, AI can see the entire Excel workbook as a single semantic system. It understands, this is the sheet with source data, this is the calculation block, and this is the final report. It sees the logic. Essentially, this is the end of the black box era, when a model was so complex that even its creator couldn't always quickly find the source of a problem. >> Okay, it sees the logic, sounds intriguing. But how does it work under the hood? What enables this understanding? The key technology here, as indicated in the documents, is the cell dependency graph, or cell dependency graph. Imagine you can highlight a single number in the source data, say, the forecast for oil prices, and see how this number travels through your entire model. It kind of colors all the cells that depend on it, and ultimately changes the final profit. It's like taking an X-ray of your Excel model in real-time. >> Got it? So it's a tool for visualizing connections. But how much can we trust the conclusions the system draws based on this X-ray? I see a number here that, frankly, makes you think. 55.3% accuracy in the Finance Agent benchmark. That's, well, slightly better than flipping a coin. What's the breakthrough here? >> Excellent question, and it's absolutely valid. The context is important here. This benchmark tests not simple calculations. It tests complex, multi-layered reasoning where you need to find data, interpret it, and build a whole chain of conclusions. In such tasks, even a human won't show 100%. But you're right, 55% is not a figure to be blindly trusted. And the authors of the materials emphasize this. High accuracy does not mean a complete absence of hallucinations, so the golden rule remains in effect. Human in the loop. >> Human in the loop. Absolutely. A human must be aware and perform the final check. >> Okay, we've covered that. But let's get back to more down-to-earth problems. Anyone who has worked with Excel has seen that terrible #VALUE! error. It can break the entire model, and you spend hours searching for where you made a typo. How does the new system solve this old pain? >> It acts as a smart auditor. Instead of just showing an error, it explains its cause. In natural language, for example, you see not just #VALUE!, but a pop-up comment. Error in cell D15 caused by data type mismatch. And then cell assumptions. B5 contains text, while the formula expects a number. >> Uh-huh. >> This is no longer just a statement of fact, but a diagnosis with the source of the problem indicated. And speaking of scale, the materials provide a staggering figure from the NBB benchmark. A potential saving of 213,000 person-hours per year just on model auditing. >> So the system not only finds errors but also explains its conclusions. What about more complex things? If I ask, "Why did revenue drop?" And the system answers, "Due to seasonality adjustment." How can I be sure it didn't invent this answer? That's precisely why the function called cell level citations has been introduced. Trust, but verify at the cell level. For every assertion it makes, the model provides an interactive link. That is, for the answer "revenue decreased due to seasonality adjustment," you will receive not just text, but a clickable link that will take you to a specific cell, say, B45, and show a number, for example, -$10.2 million. That is, the number on which this conclusion is based. You see all the evidence. >> And how is this technically possible? >> This has become possible thanks to a giant context window of 200,000 tokens. This allows even annual reports in 10K format to be analyzed in their entirety. >> A system that not only gives an answer but also shows the evidence. Sounds convincing, but does it actually help speed up work? I see a very interesting example from the company IG in the materials. >> Yes, this is a very telling case. Their standard business review process, data collection, reconciliation, manual checks, took five business days. >> A week, you mean? >> A whole work week. A long, laborious process. With the introduction of Cloud, where the workflow changed to synthesized data, human validation, strategy development. The same cycle now takes only one day. >> Five times faster. >> Exactly, a five-fold acceleration. And critically, the data accuracy, according to their measurements, increased to over 90%. Simply because the machine doesn't get tired and doesn't make mistakes due to inattention. >> But isn't there a downside? If it does all the grunt work, aren't we risking a generation of lazy analysts who lose their deep understanding of data and just click buttons? >> This is the main risk noted in the materials. Automation should not lead to superficial analysis, and this situation creates a very interesting dilemma for Microsoft itself. On the one hand, they have their own powerful flagship product, GPT4 O, and Copilot. On the other hand, they have invested over $30 billion in Anthropic, the creators of Cloud. >> And how are they solving this? >> The solution they offer is the Model Choice ecosystem. Within the secure environment of Microsoft 365, the user will have a choice. >> And how should a user make this choice if I have both options, when do I need one tool and when the other? >> The sources provide a very apt analogy. Microsoft Copilot is, let's say, a universal soldier. It's good for quick commands, formatting tables, creating Outlook emails, or summarizing for Teams. And Cloud from Anthropic is a financial scalpel. It excels in deep reasoning, in analysis of long, complex documents, and, critically for finance, demonstrates significantly fewer hallucinations. >> Is there confirmation of this? Yes, there's a link to the Human Eval benchmark, where Cloud leads specifically in tasks requiring planning. However, there's also a downside that's important to note. Currently, Cloud does not support old macros and VBA. >> So, for a company with old models, this could be a problem? >> Yes, for a company with a large number of legacy models, this could be an obstacle. Okay, this all sounds great, but my imaginary security director is already sounding the alarm. We can't just send our confidential financial data to some third-party AI model, can we? How is this issue being addressed? >> Absolutely a valid concern. And the materials pay special attention to it. The integration is happening as part of the Frontier Program. This will make Cloud a fully enterprise-ready solution by February 2026. From a security perspective, the process is structured as follows. Data from your Excel is first encrypted within what Microsoft calls Ten and Boundly. >> So it doesn't leave my company's perimeter, so to speak. >> Exactly. And only then, in encrypted form, is it processed by the model. And here's the key point. The model is stateless. >> Stateless. What does that mean? >> It means it doesn't remember anything. It's like a calculator; it receives a task, solves it, and immediately forgets everything. Your data is in no way used for its further training, and the entire process is certified according to the SOC2 security standard. >> But there are probably some nuances. >> Yes, there are a couple of nuances. For now, there's an exception for Data Boundary. And your query history is not saved between work sessions. >> So, the system is secure. Is there a simple rule for when I need this financial scalpel? >> Yes, for this, the documents have a very clear strategic matrix. Imagine two axes. The horizontal one is the data type, from fully structured on the left to unstructured, like contract texts, on the right. The vertical axis is the task complexity, from low at the bottom to high at the top. So, the Cloud zone is the upper right quadrant: high complexity tasks with unstructured data. >> And for everything else? >> For everything else, simple tasks with structured data, native Excel or Copilot are perfectly sufficient. >> But isn't the power of such systems often not just in the model itself, but in how it interacts with the external world? Can Cloud, for example, pull data from external sources? >> Absolutely. And this transforms Excel from a static table into, well, a living organism. The technology called Model Context Protocol, or MCP, allows for secure connection to external APIs without a single line of code. >> For example? >> The materials list key partners like IHS Markit for macroeconomic data and currency exchange rates. Moody's with their OrbC database of 600 million companies. >> Impressive. >> Yes, for credit ratings. S&P Capital IQ for fundamental indicators. Moreover, you can set up connectors to internal ERP systems or, for example, to SharePoint, to analyze thousands of legal contracts. >> All of this sounds incredibly powerful, but let's talk about money. This must be very expensive. What's the economic effect here? >> The Total Cost of Ownership (TCO) section is very telling. A hypothetical but typical analytical task is taken. With the traditional approach, it takes, say, 40 hours of analyst work. In monetary terms, this is estimated at around $2,000. With the Cloud approach, the same task takes 4 hours of work. That's $200 plus the cost of model computations and tokens, about $100. Total $300. >> $300 versus $2,000. >> Yes. >> Wow. This changes not only the analytics budget but the very nature of the specialist's work. So, their value is no longer in meticulously checking data for hours and mastering functions. >> Exactly, a tectonic shift is happening from the role of Data Gatherer to the role of Data Architect. The materials present a new pyramid of competencies. At its base is now simple data handling, above it is the ability to correctly formulate AI queries, i.e., prompting, and at the very top is critical validation of the results produced by the machine and developing strategy based on them. >> And is this already working somewhere? >> The experience of Coinbase and Citi, which is mentioned, shows savings of 8-10 hours per week per employee. That's a whole workday. >> That's impressive. But where is all this heading next? What future do these materials paint? >> And here we move to things that sound like science fiction but are already in development. First, Autonomous Analytical Agents. >> So, it's like hiring a junior analyst who lives inside my computer. >> Perfect analogy. You no longer give it step-by-step instructions. You set it a goal. For example, "Analyze competitor reports for the last quarter and identify key risks for our business." The agent itself plans the steps. Find reports, extract data, analyze, compile a summary. It uses Excel tools, APIs, has memory and the ability to reflect. >> Sounds incredible. What else? The second concept is vibe coding. It sounds a bit strange, but the essence is that you create complex models not by writing formulas, but by describing the desired behavior or analysis style. Instead of writing hundreds of lines of VBA code, you can write a prompt: "Create a liquidity stress test model in the style of Goldman Sachs." >> In the style of Goldman Sachs. Exactly for this, ready-made expert skill packages, Agent Skills, are being developed, for example, competitor analysis, DCF model building, or preparing a coverage initiation report. >> So, I can literally ask Excel to think like an investment bank analyst, and it will do it. Can you give an example of how this would look in real work, step by step? >> Of course, the materials have a very clear example of such an agent chain. Step one. Excel, with the help of Cloud, detects a serious anomaly in sales data. Step two. Cloud, understanding the context, automatically prepares a draft email for the head of sales with a brief analysis of this anomaly. Step three, and it's mandatory. Unbreakable - human confirmation. The user reviews the text, makes edits, and clicks approve. Only after that, step four, Outlook automatically sends this email. All this orchestration is configured through Microsoft Copilot Studio. >> An impressive picture. And how should companies prepare for this? Surely, you can't just implement it in a day. >> Of course, in conclusion, the materials offer a very pragmatic roadmap for the CFO for 12 months. It's broken down by quarters. First quarter: Access. Ensure access, purchase licenses, conduct basic training. Second quarter: Integration. Set up data connectors, calculate TCO for pilot projects. Third quarter: Agency. Create a sandbox, a secure environment for testing the first autonomous agents. And fourth quarter: Scale. Implement vibe coding standards and scale the most successful cases across the entire company. >> And what's the goal? >> The ultimate goal, according to the same NBM benchmark, is a 20% increase in productivity. And, of course, the thought about the main risk runs like a red thread. It's not the technology, but change management. Working with people. >> Well, let's summarize. The picture emerging is grand. Excel, essentially, ceases to be just a calculator on steroids. It transforms into an intelligent platform for reasoning. Capabilities range from automatic auditing to creating simulations in natural language. This truly changes everything. Yes, and if you look at it from a bird's-eye view, the main conclusion is: there is a fundamental shift in what we consider valuable analytical work. All the routine, data collection, cleaning, verification, which took up to 80% of an analyst's time, is disappearing. And entirely different skills are coming to the forefront: the ability to ask the system the right, deep questions, the ability to critically evaluate results, and, most importantly, the art of building the right strategy based on this synthesized data. >> And here's a final thought to ponder, which these materials offer and which keeps me up at night. We talked about analysts becoming architects. But if an agent can not only find an insight but also propose specific actions and then execute them after a single click-confirmation from a human, where does the new boundary now lie between the role of an analyst who researches and the role of a manager who makes decisions and acts? Doesn't the analysis process itself become an integral part of strategy execution? M.