Can artificial intelligence genuinely solve any task? More importantly, can it solve your particular task? Tens of millions of people and millions of companies are investing money, time, and substantial resources in AI. A small number obtain extraordinary results. Others get nothing, while some lose serious money, time, and opportunity. Why does that happen? And how can you determine in advance whether AI will actually solve your problem or merely consume your time and budget?
Today I want to explore that question through a very interesting McDonald's case. I will begin with another example because it is instructive, concrete, and easy to understand. Before automating a task or building a new solution, you need to ask whether the task should be automated at all. Is this worth studying deeply now, or is it a minor experiment that may become pointless because the entire problem will be solved six months from now?
A parallel experience arose for me two days ago. In one of my small and straightforward US businesses—a compact structure that manages certain assets—I decided to connect AI to the financial records. The company uses QuickBooks, the American financial-management system. QuickBooks offers many services; I used its bookkeeping and finance component. My accountant also accesses it periodically to prepare and file tax returns. The setup is not especially complicated. I have seen many systems of this kind.
Financial automation looks like a solved and simple problem from one angle and an extremely difficult one from another. In many businesses where I am an owner or participate in management, finance is still not properly automated. At best, there is a well-maintained Excel workbook. Tax accounting and regulatory filings are a somewhat different matter, but that is not today's subject. I went through the system's compliance process—the long set of questions required before it allows an integration and API access.
I connected the API and ran everything locally. The risk was limited. Even if some data had been partly erased, the consequences would not have been critical. This is not a company with a thousand transactions every day; it records roughly one hundred to one thousand transactions a month. There are no dedicated finance employees inside the business. I enter periodically, and a few other people perform specific tasks. I connected the AI system, and what it did was extraordinary.
It calculated everything and understood the structure. I then gave it the company's tax returns, account exports, credit-card statements, bank statements, and records from several payment systems. The system assembled all of the data and produced a very good HTML report in a simple, readable format. It was ordinary Codex—OpenAI, your merchandise has arrived, by the way. The entire setup was local and straightforward. There were a few additional security and access-control steps, but they were manageable.
My first reaction was: excellent. I had obtained an additional audit that I genuinely needed. It reduced my concern about numbers failing to reconcile. The people who normally review the company's records are primarily focused on filing accurate returns, not on examining every material detail inside the business. The next morning, while speaking with my partner, I realized something clearly: I no longer needed QuickBooks. The previous evening had shown me not only that this particular product might be unnecessary, but that I might not need any conventional financial system at all.
If I did need one, it would have to be a serious system of a very different kind. As I said, many of my businesses still use Excel because standard financial systems cannot meet the actual requirements. When they can, you often have to assign programmers, analysts, and support staff, and you still remain inside rigid constraints. I am no longer engaged in intensive daily operational control across every business, so I have stepped away from some of the most demanding requirements.
But I know that I personally need a completely different quality of information. Even Excel no longer satisfies me. I feed those spreadsheets into different models and AI systems anyway in order to obtain the information I need. What the new setup did today was remarkably good. But an important distinction remains. It is one thing to receive an additional layer of support almost as an experiment, even when that experiment saves time and reveals useful detail. It is another thing to hand the system full responsibility.
During the conversation with my partner, I understood that QuickBooks had become an intermediary. I query it through the API, but I could load the same bank statements directly. The AI can classify every transaction and allocate it across accounts immediately. Notice that this moves us away from normalized charts of accounts. Most viewers may not know what that means, but normalization is one of the largest problems in almost every software project: you must define fixed reference directories and data structures.
AI makes it possible to create dynamic reference systems and change them continuously. The structure in which data is stored can evolve as the system works. This is not a technical podcast, but that point is fundamental. Anyone who wants to advance significantly in the practical use of AI—in business and in daily life—should understand it. Some episodes explain ideas more simply or discuss broader philosophical questions. But I still want to share cases like this. After three or four hours of total work—and while the system continued running additional calculations today—I realized QuickBooks might no longer be necessary.
The process used very few tokens because the data volume was not remotely astronomical. I began imagining how many startups will disappear. At the same time, people may finally gain access to a level of automation that has never existed. Owners of financial-management companies will naturally say that their systems are excellent. I disagree. That is my personal view, and anyone interested in the broader subject can find my lectures outside ToTheMoon on financial accounting and company strategy.
But there is another boundary: am I ready to let the system file the company's tax return? The accountant may still be necessary because the business has many small local legal details. Yet no individual can know everything a strong model can discover when it performs a broad review. I am not proposing a fully automatic tax filing. That distinction is crucial. Whenever companies implement AI, they immediately want complete automation. Then they discover that what they actually need is to improve a few of their own processes.
I have worked in business automation for almost twenty-five years, and the pattern never changes: everyone initially wants a spaceship, then returns to fixing ordinary workflows. Today's question is not automation itself. It is what can be entrusted to which AI, and at what moment. Even when I talk about preparing a tax return, I do not assume the process should be fully
automatic. AI can assemble a large, detailed body of data and calculate it. I can then give the result to a person. A human reviewer may examine some issues narrowly and may know special practical techniques. In theory, that knowledge can add value. For the final 2026 return, perhaps I will provide the materials both to a tax professional and to the models. Or perhaps by the beginning of the filing period I will no longer need to send anything to a tax professional, because the available tools may be entirely different and I will know that the system can handle the work.
A year ago there were tasks I would never have given ChatGPT. Today there are tasks I assign to it without hesitation. You probably have your own examples; tell us what you would not delegate before but now do. One of my examples is turning transcribed videos into complete, publishable articles. I now have video material that must be converted into text while preserving about ninety-nine percent of the substance, along with my style and details. Preserving voice and detail used to be extremely difficult.
Now it works. Why do I need it? Some people prefer to read. Structured text is also easier to work with later and can support search discovery. The information must be organized properly: fillers and repetitions removed, speech corrected, while perhaps ninety-five percent of my stylistic voice remains. I have an entire set of requirements for that process. A year ago I would not delegate it; now I do. Tax filing may follow the same path. If the new system works reliably for a month, reconciles every balance, handles every detail, and sends me reports each day for a week, I will see that it is functioning as a living process.
This particular company does not even need operational financial reporting; I mainly monitor balances, review certain figures, and manage assets. It is a different kind of business. Once I see that everything is structured and stable, I will of course cancel QuickBooks. Perhaps QuickBooks costs fifty or sixty dollars a month, saving roughly six hundred dollars a year. But the six hundred dollars are not the point. Nor is the point whether I later pay an accountant five hundred, one thousand, or two thousand dollars to file a return.
The point is efficiency, quality of work, confidence in the data, and one additional consideration. For me, the critical issue is not saving one or two thousand dollars. It is ensuring that every record in the company is maintained carefully and legally: every document entered correctly, every transaction classified properly. I cannot guarantee that level of quality today. Achieving it manually requires not merely hiring a person, but investing an enormous amount of time. I want the company to be able to pass any review effortlessly five or seven years from now.
A retrospective audit should not create panic. Every record should be recoverable, traceable, and easy to inspect. Reviewing 2024 or 2025 would be relatively easy today because the company is only about two and a half years old. The emails still exist, most records are electronic, and the data can be retrieved quickly. But many businesses have periods in which records were not preserved. The information may sit in an inaccessible email account or in an employee's personal Excel file that later disappeared.
This is an excellent moment for integration. We can connect different systems rapidly through APIs or MCP, download Excel files, access websites, and automate an enormous range of work. But the decisive question remains: will we obtain the result we need? Could the system create a problem when the tax return is filed? If I do not use it now, could a gap appear six months from today that can no longer be repaired? Data may disappear, work may be recorded poorly, or AI itself may delete something.
Employees and system administrators have always been able to delete data. You can copy and duplicate everything, yet the possibility of loss never disappears completely. AI adds another version of the same risk. You give it access and restrictive rules, but you never fully know what it will do. Claude and Codex run for me almost twenty-four hours a day across a huge number of sessions. I see how often even the newest systems make mistakes. Opus 5 has just been released in Claude, and I can see both how quickly these systems are improving and how often they still perform work badly.
I recently recorded an episode about model memory and how much ChatGPT's memory has improved. The capability is extraordinary. You can start one session, ask a question, and immediately open another session. I did exactly that. I submitted a question to the Pro version but needed an answer urgently, so I opened another session and asked Extra High. Extra High responded after thirty seconds, while Pro took about ten minutes. The striking part was that Extra High had already noticed the other session, read it, and supplemented its response.
My second prompt had been incomplete and contained only part of the constraints, yet the
system connected the two contexts. The quality of cross-checking, memory, detail, and interaction among sessions is becoming very interesting. That brings us to a real case at McDonald's. It is especially relevant because McDonald's began working on this problem long before the current generative-AI era. Around 2019, before its later IBM partnership, the company acquired a California startup called Apprente. The goal was to process complex drive-through orders involving different accents, languages, multiple items, and menu variations.
McDonald's had tested systems before the acquisition as well. The task was to let a customer place an order by voice without a human operator participating. Public testing began in ten Chicago restaurants in 2021. At the time, a McDonald's executive said the system entered orders correctly about eighty-five percent of the time. That was before today's AI wave. Even then, employees still had to intervene and continue the conversation in roughly one out of five orders. That is an excellent illustration of the boundary between what a system can and cannot do.
In October 2021, IBM agreed to acquire McD Tech Labs, the technology unit created from Apprente— The California company McDonald's had acquired. IBM was expected to continue developing automated order-taking and improve the system's ability to handle different languages, accents, dialects, and menu combinations. We recently recorded an episode about the quality—good and bad—of ChatGPT's new voice. Who has actually continued using it? I tried it, found it amusing, and then stopped.
Tell us whether you kept using it or switched to another voice product. By 2022, analysts estimated that accuracy across twenty-four restaurants in Illinois remained around eighty percent—lower than the earlier eighty-five percent. McDonald's, however, wanted ninety-five percent before a broad rollout. Why did the company want the system? It may sound like a plan to eliminate operators, but the deeper issue was efficiency and service quality. Anyone who watches a McDonald's or Starbucks operate in a small space sees five, seven, or nine people performing highly coordinated work.
Elsewhere, a coffee shop may have a line of forty people and one employee unable to serve them. The operating system matters. A year later, the technology was still far from McDonald's required threshold. By 2024, the test had expanded to more than one hundred US restaurants. But in June 2024—after ChatGPT, Claude, Gemini, and the current generation of systems already existed—McDonald's told franchisees that automated order-taking would be shut down at every test location within weeks.
The separate voice-ordering partnership with IBM was ending. Why did the system fail? This is the fundamental question. Over the last three years, in roughly ninety-five cases out of one hundred when people told me they were beginning an AI implementation, my recommendation was: do not implement it yet. Do not automate this task now. Use ChatGPT manually inside the workflow. If you need to analyze a sales department, export the data once a week, upload it to ChatGPT, and conduct the analysis.
If it must be done daily, assign an employee to do it daily with ChatGPT. But do not automate the entire process, build a dedicated system, spend money on programmers, or buy servers before the technology and the task are ready. The number of mistakes companies have made is extraordinary. In the United States, an enormous number of businesses have already suffered from premature implementations and continue to suffer. People still do not understand the issue because everyone around them claims to know exactly
what to build and insists that everything is simple. What does eighty-five percent accuracy mean in real work? First, today's generative AI can make mistakes. That must be accepted calmly. There are automated systems that do not make probabilistic errors, but those are strongly programmed boxes rather than the kind of AI we are discussing. When you use current AI, error is part of the operating model. I often tell directors, colleagues, and partners not to panic when a system produces an incorrect report or misjudges something.
In some tasks, the error is harmless. If AI analyzes one hundred thousand sales-management events and misclassifies five thousand, that may be acceptable. But if a system prepares the one tax return you file each year, the same error rate makes it unusable. If it controls the ordinary money in your operating bank account, you cannot accept a five-percent error. A trading strategy may deliberately tolerate five percent if the risk is understood and constrained. But in ordinary finance, five percent can be catastrophic.
The same reasoning applies to advertising campaigns and conversations with customers. You must define exactly what the missing fifteen percent at McDonald's means for your own process. The acceptable threshold differs by task. In one application a system may be wrong ninety percent of the time and still produce useful creative options. In another, a three-percent error rate is already intolerable. It is also important to understand that McDonald's voice system was only partly AI.
It contained a large amount of tightly programmed code. We can see the same tradeoff in smart speakers such as Amazon Alexa or Yandex Alice. Earlier versions answered standard questions quickly and clearly. Now a speaker may spend almost a minute reasoning about the weather, still make a mistake, and give the wrong answer. AI brings broader knowledge and greater capability, but also uncertainty. At McDonald's, fifteen percent error meant fifteen problematic orders out of every hundred.
The restaurant also had to keep an employee ready to intervene in perhaps twenty or twenty-five percent of conversations. Even when the system had not yet made an obvious mistake, a person might need to check a complex order. In other cases, the employee would not realize that intervention was needed. At first glance the restaurant saves labor and serves customers faster. In practice, much of the saving disappears because the same employee still has to stand at
the window. The difficult conditions identified in the voice trials were predictable: vehicle and street noise; different accents and speaking styles; several people talking from one car at the same time; voices from a neighboring car entering the microphone; and constant changes to an order. Anyone who has ordered at a drive-through has heard the pattern: with onions, without onions, add sauce, remove the previous item, put it back. The structure becomes complicated quickly.
My own morning order in Los Altos is extremely simple: “Good morning, one Americano, please.” I do not eat sweets and never ask for additions. I am not ordering the kind of Starbucks drink that contains chocolate, an orange flavor, a purple-colored supplement, a tiny amount of ice, a special milk, and one-third of something else. Other customers place long family orders. Products may have similar names, and the system can genuinely confuse them. Social media collected many videos from the McDonald's trial in which the system repeatedly added nuggets or burgers, captured another car's order, or produced absurd combinations such as ketchup and butter with ice cream.
I recently visited Hawaii with my daughters. At breakfast, a waiter came to our table and we changed almost everything on the menu: remove this, add that, replace another component. We ordered almost fifteen dishes, each with a modification. The waiter wrote almost nothing down and walked away. I told my daughters, “Do you think he remembered all of that? We changed everything that could be changed.” He delivered the order perfectly. Good waiters often do not write anything down unless the request is exceptionally complex.
The project was not closed merely because customers complained.
It was closed because this task requires something close to ninety-five percent quality. Some errors are unavoidable, just as human employees make mistakes and customers press the wrong button in a kiosk or application. The next question is how the system resolves difficult cases. There is a coffee chain called Blue Bottle, which operates both in and outside the United States. I once drove to a location in San Jose, in the South Bay at the southern end of Silicon Valley, and placed an online order.
I arrived, but the order was nowhere to be found. I asked an employee where it was. They asked to see the application, and I discovered that I had ordered from Palo Alto by mistake. I apologized and placed the order again. The San Jose store did not say, “No problem—we will make it for you at no charge.” A month earlier, my family and I stopped at a Starbucks while driving on the highway. I had accidentally ordered from another Starbucks about three hundred meters away. The employees immediately said, “No problem.
Tell us the items and we will make everything here.” They prepared the full order and gave it to us without charging again. The quality of a system includes how it resolves cases like these, not merely whether it
recorded the initial order correctly. When you plan an AI project, first ask whether the underlying technology will advance so quickly that the entire implementation must soon be deleted and replaced. That, in my view, was part of McDonald's problem. The company started when the world was not ready. Even today, the technology is not fully ready to take every drive-through order at the required quality. McDonald's and Starbucks have enough money to conduct additional research.
A small chain does not. I would not advise a smaller business to spend heavily training a voice robot now. Voice systems will improve, and at some point the capability will be so good and inexpensive that the implementation will require far less investment. Eventually, an order at McDonald's or Starbucks may not involve the restaurant's voice hardware at all. You may place it through your phone and a general environment called ChatGPT, Claude, or something else. Whatever the platform is eventually called, it will not merely aggregate restaurants and orders.
It will be your own AI environment. If I am driving and want coffee, I should be able to say: “Find the Starbucks that requires the smallest detour. How far do I have to leave the route? Does it have a restroom?” Starbucks normally does, but some gas stations do not, which matters when that is what I need. Then I should say, “Place this order for me,” arrive, and collect it. That interaction is unlikely to happen inside a Starbucks-owned environment. If every company requires its own closed assistant, the result will be absurd.
This is why connectors are being created now. MCP connectors are a transitional mechanism that allows AI in one environment to understand and interact with a system in another. These are relatively simple problems and will be solved quickly. The strategic question for a company is therefore whether it should invest in building the entire layer itself or put the same money into something else that can produce a much greater return. Tell us what you think. Share the video with friends, subscribe, leave a like and a comment, and join us for the next episodes.