By Alexander VolchekThis essay grew out of several recent conversations on ToTheMoon, the YouTube channel where we discuss AI and technology markets. It is not a recap of an episode. It is the one line I keep coming back to after using Codex, CRM data, agents, AI shopping, and a few pieces of extremely annoying professional software.\ Most conversations about AI still begin with the same panic: who gets fired first? Developers? Analysts? Designers? Sales reps? Lawyers? Junior people? Middle managers?\It is a fair fear. It is about jobs, money, mortgages, children, status. But I think we are looking one layer too late. AI does not have to come for the person first. It can come for the software that sits between the person and the result.That sounds abstract until you see it happen inside a very boring business process.Take CRM. You have customers, calls, deals, managers, statuses, funnels, dashboards, reports. You want to understand what is happening in sales. The old route is obvious: ask the head of sales, ask an analyst, wait, get a spreadsheet, ask another question, wait again, open six numbers, and realize that you now have twenty more questions.Then you open Codex and ask, in normal language: what did the salespeople do in the last 14 days?The system connects to the CRM. It looks at the data. It checks. It pulls the sample. It answers. Not “please build a report.” Not “go to this tab.” Not “ask someone with admin access.” Just a question, then an answer from the actual database.At that moment the CRM loses its throne.The deals do not disappear. The call history does not disappear. Permissions do not disappear. The database does not disappear. But the interface stops being the center of work. The CRM becomes a source of data. A pipe. A layer. Something the intelligence reaches into when it needs to.And if the CRM is only a source of data, then the rude question becomes unavoidable: why should I live inside the CRM at all?\The CRM MomentIn one business, I got so tired of human analytics that I asked for the entire CRM and deal system to be exported. Calls, employees, deals, tasks, raw data, all of it. I asked a senior analyst I know well to give me a summary. At the same time, I gave the data to ChatGPT.The human summary came the next day. It was not bad. It was clean, normal, familiar. A few numbers, a little framing, and immediately I wanted to ask twenty follow-up questions.ChatGPT came back much faster with a large, ugly, useful canvas of questions and slices: where the strange patterns were, which cuts to check next, which parts of the company had probably never been analyzed properly, what was missing, what looked suspicious.I do not like cheap fearmongering. I do not like channels that keep telling people: “you are finished, AI will fire everyone, learn prompts or die.” That is not my view. People will still work. Good people will be worth more.But when you see that kind of report, an unpleasant thought does appear: why am I waiting for this entire chain of people if the model gives me a better map of the problem?Then I asked for more data. It was not given to me quickly enough. So I got super-admin access to the CRM, because I am an owner. I created a token. In the evening, I gave Codex a task: connect to this system and let me query the data.I did not want to program. I did not want to open the code. I did not want to read API documentation, create GitHub drama, deploy servers, or pretend that I enjoy DevOps vocabulary in a business question.In the morning, Codex said it was ready.I could ask for any slice. Calls last week from these managers with this duration. Deals stuck at this stage. Employees who called less than two hours. Words appearing in notes. Patterns that should be checked.Very quickly, even CSV felt stupid. Why export a file if I can just ask the system: who is underperforming this week, where are the strange patterns, what should I verify next?This is not “AI made a report faster.” That is too weak. AI removed a whole layer of waiting, interface, and mediation. I connected, asked, received, and checked.That is the real shock.\Software Becomes PlumbingFor the last 15 or 20 years, we built a business around applications. Sales lived in CRM. Numbers lived in Excel. Documents lived in Word or Google Docs. Tasks lived in a task tracker. Websites live in a CMS. Design lived in Figma. Analytics lived in BI. Every new process needed a new app, a new subscription, a new admin, a new training document, and a new person who “knows how to use it.”\ We accepted this as normal. Need sales data? Open the CRM. Need numbers? Open the dashboard. Need a report? Ask the analyst.\Need Telegram connected to an internal system? Find someone who understands APIs, tokens, servers, permissions, and the small hell that comes with all of it.But once the model becomes the working environment, the application no longer has to be the place where the human sits with his hands on the keyboard. The application becomes a data layer. And a data layer is not a product in the old sense. It is storage. It is a pipe. It is a place the agent enters, reads from, writes to if allowed, and leaves.The word is ugly, but accurate: a lot of software becomes plumbing.Not because the software is useless. Good software will still matter. Security matters. Permissions matter. Data quality matters. History matters. Reliability matters. Legal responsibility matters. But the user does not want to worship the interface. The user wants the result.What happened? What should I do? Where is the risk? What failed? Who should I check? What is the next action?If a single AI workspace can answer that by moving across tools, then the old screen full of buttons becomes secondary.\This Is an Attack on SaaS UX, Not Just a FeatureSomeone will say: “Fine, you built an integration. Integrations existed before.”Correct. Technically, a lot of this was possible before. You could connect CRM to BI. You could hire a freelancer. You could write a script. You could ask a developer to create a dashboard. You could buy an add-on. You could spend three weeks explaining to someone what you wanted and then spend another week checking whether they understood you.Markets do not change when something is possible. Markets change when it becomes cheap enough, fast enough, and mentally available to people who do not want to become programmers.That is the difference.I do not want to know where each file is. I do not want to maintain a little server. I do not want to read your API documentation. I want the result. If Codex, Claude Code, or another agentic system takes over the dirty part, then the layer of “find a person, explain the task, wait, test, fix, repeat” starts to shrink.This is the attack on old SaaS.Old software often sold the route to the result: here is the screen, here is the tab, here is the filter, here is the export button, here is the tutorial. AI says: I do not need the route. Give me the data and the goal.If your SaaS moat is data, trust, workflow, permissions, compliance, and deep business logic, you still have a business.If your moat is “we have a screen with buttons,” that moat is not deep.\Kubios, Sensors, Excel, and My Patience for InterfacesI had another small example, and it made the same point.I bought a professional heart sensor and used Kubios, a serious piece of software for heart rhythm analysis. It opened MATLAB. It wanted Java. It had plugins, settings, fields, and all the usual professional software energy. I looked at it and thought: I am not in the age where another window with 86 settings makes me happy.The important part was not Kubios itself. I needed raw data. I needed the software to pull the sensor data into a readable form: Excel, a chart, a report. After that, I could put the data into ChatGPT and ask for a human explanation.So again, the expensive professional tool became an intermediate layer. It collected and exported the data. The thinking moved elsewhere.That changes how I look at devices. If a wearable, a sensor, a camera, or a business system locks me away from raw data, I now see it as a future problem. I do not just want a nice app. I want access. I want to know what the device measured, what it missed, where the signal disappeared, and what assumptions the software made.Excel is similar. Excel is probably still the number one business interface in the world. I send people Excel files. People ask for Excel files. But I increasingly do not think inside Excel. I ask AI to create the table, split sheets, normalize data, add views, and package the result for humans who still want the format.Excel becomes a delivery format. Not the place where the work necessarily happens.\Small Projects Stop Being Small ProjectsThis is also where the freelance market starts to change.Not the whole market. I am not saying all freelancers disappear. Good engineers, good architects, strong product people, serious designers, people who understand real systems — they will have work.But the small annoying jobs change.A little admin panel. A Telegram flow. A sales assessment bot. A connector. A simple internal dashboard. A script that pulls data and creates a report. A tiny tool that used to cost $100, $200, $500, plus a lot of explaining.These jobs are moving into the range of a subscription and a determined operator.I recently built more infrastructure than I expected: database, server, sources, admin panels, API connections, reporting. Normally I would have had to involve developers, maybe a DevOps person, maybe someone to polish the interface. The quality of the deployment and admin panel was better than my human expectation for that first version.That does not mean I suddenly love doing it all myself. I actually resist it. I do not want to become the person who personally builds every little internal tool.But when people in a company boycott new tools, do not understand them, or keep waiting for the old process, the temptation is obvious: fine, I will make the agent do it.That is a hard sentence, but it is real.\Fire People or Amplify Them? There Is No Beautiful AnswerThe politically nice version is: AI should amplify people, not replace them.I agree with that. I really do.But the sentence becomes empty if we do not ask: which people?What is an analyst who gives me six numbers and no picture? What is a developer who does not use AI to check architecture, code quality, and risk? What is a manager who operates a CRM screen and cannot explain the business process underneath it? What is a designer who only knows how to move pixels inside the old tool, but cannot reason about the system, customer, cost, and constraint?\ The problem is not “developers.” The problem is weak developers. The problem is not “analysts.” The problem is analysts who behave like slow UI. The problem is not “managers.” The problem is managers who do not understand the process and only pass information from one screen to another.\Strong people become stronger. The person who can see the system, formulate the task, control the risk, judge the output, and push the next question becomes dangerous in a good way. The title matters less. Product manager, founder, analyst, engineer, designer — fine. The real question is: can this person think at the level of the task, not just at the level of the interface?That is where the money will move.\Cheap AI Gives a Pretty Answer. Expensive AI Does Real WorkThere is another thing most people still do not understand: not all AI usage is the same.People ask a complex question, get an answer in two seconds, and believe they used modern AI. Maybe they did. Maybe they just received a polished surface.For converting Fahrenheit to Celsius, two seconds is fine. For analyzing sales calls, a medical image, a contract risk, a real estate deal, a company acquisition, or an architecture decision, a two-second answer can be a toy.A serious answer may take 30 minutes. Or an hour. Or three hours. Sometimes you run it overnight. Sometimes you ask it to check itself, build control totals, look for missing data, find contradictions, and then explain where it is uncertain.This is why many AI apps feel mediocre. It is not always because AI is bad. It is because the product costs $19 or $29 a month and cannot spend $50 of compute on every request. So it uses a cheaper model, a cheaper API, a faster mode, and a thinner analysis.The user then says: “AI is not there yet.”No. That product is not paying for the expensive version of the work.A new divide is coming. Not just between people who use AI and people who do not. Between people who are addicted to instant answers and people who know when to run a slow, expensive, verified process.If your default is always instant, you may be using a new tool with an old search engine habit.\Agentic Commerce: The Agent Will Buy for You, But Whose Side Is It On?There is a second line that looks small and consumer, but I think it is huge: AI shopping.I was choosing a gymnastic ring for my daughter and the hardware to mount it. I am not a specialist in that hardware. I asked ChatGPT what to buy and how to hang it safely.It started recommending expensive parts. Sometimes the hardware costs more than the ring. I asked: " Why do I need an $80 hook if there is a $5 one? It said the cheap one might not be load-tested. I said: It says it is certified. Then it moved to a $15 option. So why did you first suggest the $80 one?This happened in several categories that week: bolts, hooks, and household things where I did not know the market well. The model often moved toward the premium option.Maybe it was right. Maybe safety matters. Maybe you should not save money on a child’s equipment. But that is exactly the problem.The model may not be manipulating you in a dark, conscious way. It may simply optimize for safety, brand, reliability, and fear. But if it knows your context — children, home, health, risk — the recommendation becomes personal. “Do not save money on your child” is not just product advice. It is an emotional lever.Before AI, the middleman was a salesperson, an ad, a marketplace ranking, a blogger, a consultant, a review score. Now the middleman is a system you talk to privately, a system that knows your preferences, your fears, your budget, your family situation, and your previous questions.That will become a fight for trust.\ Who controls the recommendation? The user? The model provider? The marketplace? The advertiser? The payment system? The manufacturer?\If AI becomes the buying interface, the “shopping assistant” is not a cute feature. It is a new gatekeeper of demand.\Agents Should Not Be FoldersEvery big company is talking about agents now. Microsoft wants management layers. Google wants its own agent environments. OpenAI has workspace agents. Anthropic is pushing Claude Code and desktop workflows. Everyone has a version of the same pitch: agents will do work across tools.For enterprise, I understand the admin layer. A large company needs permissions, roles, limits, logs, controls, approvals, budgets, and someone who can see which agent is burning tokens for no reason.But I do not believe the final form is a folder of agents.Agent for sales. Agent for email. Agent for code. Agent for reports. Agent for shopping. Agent for finance.That is just old software thinking with new names.The more interesting future is one entry point. You state the task. The system decides which agent is needed, which data to pull, where to wait, where to spend more compute, where to ask for approval, where to stop, and where a human must confirm the action.The user should not think: which agent do I launch?The user should think: what do I need, what is the risk, what data is allowed, and what must be checked before anything irreversible happens?Right now everything is messy. Projects, GPTs, Codex, Claude Desktop, connectors, cloud tasks, local files, memory, model names, modes, limits. Search inside chats is still bad. Interfaces are inconsistent. Sometimes you look at a company worth hundreds of billions and think: you still cannot make a normal search through my own history?But the direction is clear. The chat stops being a chat. It becomes the work environment.\A Seven-Day TestHere is what I would do if I were reading this and wanted a result, not just another AI opinion.Do not “implement AI in the company.” That phrase is too big and usually useless.Pick one process that annoys you.1. Choose a process where you regularly wait for a person, report, spreadsheet, export, email, or small technical task.2. Rewrite it as a real question: what do I want to understand, decide, or verify?3. Find where the data lives: CRM, Excel, Google Sheets, call logs, PDFs, admin panel, API, email, raw files.4. Start read-only. Let the model read, export, explain, summarize, and compare. Do not let it send, delete, refund, change, or approve anything yet.5. Compare AI output with a human output. Where did it see more? Where did it hallucinate? Where was the human better?6. Add verification: ask it to find missing data, contradictions, weak assumptions, and control totals.7. Only then automate. Start with “agent suggests, human confirms,” not “agent does everything.”If this gives you no value, maybe you chose the wrong process. Maybe the data is missing. Maybe the task is badly framed. Maybe you used a fast cheap mode where you needed deep analysis.But if it works, you will see the real point: the problem was often not that humans were slow. The problem was that the path to the data was built through too many unnecessary layers.\Where Not to Let the Agent Off LeashThere is also a line I would not cross too early.Do not give an agent your email and ask it to decide what is important. That is madness unless the boundaries are extremely clear.Do not let it delete files because it “thinks” they are not needed.Do not let it send messages to clients without review.Do not let it move money without a second factor and a human confirmation.Do not let it make medical, legal, or financial decisions as the final authority.The right early pattern is boring: read-only, summarize, propose, verify, ask for approval.Boring is good. Boring prevents stupid disasters.I want agents. I use them. But I also want them on a leash until the process, data, and failure mode are understood.\The End of Button WorshipI think the next few years will not be only about job titles. They will be about losing the habit of thinking in applications.CRM, Excel, admin panels, dashboards, cheap integrations, small freelance tasks, internal reports — they will not all die tomorrow. But many of them will become background. Raw material. Infrastructure underneath an AI workspace.The main question for a person is no longer just: will AI replace me?The better question is: do I understand my work at the level of the task, or only at the level of the interface?Can I explain what I need? Can I give the right data? Can I limit the risk? Can I check the result? Can I notice when the model is confidently wrong?If the answer is no, then the problem is not AI. The problem is that you depend on old software more than you depend on your own understanding of the process.If the answer is yes, the interesting part begins.Then ChatGPT, Codex, Claude, Gemini, or whatever wins the next round is not a toy and not a search engine. It becomes the environment where you assemble tools around yourself.Old programs stop being walls.They become material.The buttons will not disappear tomorrow. But their power is already disappearing. And that is a much bigger story than whether one particular job title survives another year on LinkedIn.