PANews has published a lengthy essay by Digital Life Kha'Zix on what AI adoption has looked like inside his company after two and a half years in business. He wrote that the company has survived several make-or-break moments over that period and is now preparing to move again, this time because the team has grown to the point that the current office no longer has enough room.
That outcome, he said, runs against the common expectation around AI. Friends were surprised that a company using so much AI would still be hiring, because in theory the headcount should be shrinking and the office should be getting smaller. His answer was blunt: that is not how it has played out. According to the essay, AI penetration at the company is now close to 100%, and agents are being used across finance, HR, legal, business development, talent management, operations, and other roles. The company has also built many workflows and tools around those agents. Even so, some parts of the job cannot be accelerated by AI, and those parts have become more central to what human employees do.
Why he argues against a heavy AI middle platform
The essay says many management teams react to AI in a similar way: they want to build a system first, appoint a person in charge, or create an internal AI platform team. On the surface, that approach looks organized and safe.
But the author said that model creates its own problems. In his view, it often turns AI into another internal scheduling center. A problem appears on the front line, then gets turned into a request, passed through several layers, and only then reaches the people building the tool. He used examples such as HR trying to improve AI screening for candidates or a business team wanting an AI tool for client analysis. By the time the request reaches the builders, information has already been lost, and the business context may have changed again before the tool is ready.
He also said this setup tends to split capabilities inside the company in an unhealthy way. Front-line teams get better at writing requests, while the platform team gets better at using AI. In the end, the people who can actually create things are still limited to a small group. He argues that an AI-era organization should look different: people across the company should be able to use AI directly, or use AI to build tools that solve the problems closest to them.
That leads to one of the essay's clearest recommendations. Companies with fewer than 100 native employees, he wrote, should be cautious about setting up a formal AI middle platform. He did not dismiss the value of centralized work altogether. Large systems, Feishu, finance software, unified permissions, and the security foundation still need professional oversight. But he said that layer should stay thin. Its job is to guard permissions, security, cost, data standards, and hard red lines, while helping with shared needs such as servers and tokens.
Everything close to day-to-day business, he wrote, should be solved by the people nearest to the problem. Whether they use an agent directly, build a tool through an agent, write a script, build a crawler, or use RPA does not matter much to him. What matters is that business staff solve it themselves. He said that in his own company, when something in a workflow needs improvement, there is often no fixed role waiting to build the solution for you. The people who can solve it are you and your agent. He named tools including Codex, Workbuddy, and Claude Code.
He acknowledged that the process is messy at first. Some people struggle to describe what they want, some are intimidated by code, and some spend a long time producing only a barely usable tool. In plenty of cases, the first attempt with an agent is slower than doing the work by hand. Still, he called that awkward stage essential. Once someone turns a frustrating part of their own work into something that actually runs, even if it is rough, their relationship with the job changes. They stop seeing the workflow as something to endure and start seeing it as something they can alter.
That, he wrote, may be the most valuable thing AI gives ordinary employees. It is not only efficiency, but a form of agency they had not felt in a long time. For that reason, he said he now believes more strongly in a structure he described as a "thin middle platform, thick front line."
Data matters more than agents, in his view
The next section of the essay shifts from tools to data. Simply handing agents to everyone does not automatically produce an AI-native organization, he wrote. When companies ask him about AI transformation, he keeps running into the same problem: they do not have the data foundation.
He listed familiar examples. Meetings were not fully transcribed. Client communication was scattered across different chat histories. Contracts and quotations had no single standard version. Projects ended without proper reviews. Some of the most important know-how existed only in the heads of a few veteran employees. Under those conditions, he argued, agents cannot do much. His conclusion was direct: agents are not the key issue; data is.
He wrote that his company, Xushi Media, runs an MCN business, has signed hundreds of creators, and has worked with nearly several hundred brands. The company tries to preserve past content data, business data, and collaboration data as internal assets. Information that does not fit neatly into standard fields is turned into unstructured tags and stored in Feishu multi-dimensional tables. Full meeting transcripts, documents, knowledge materials, and SOPs are also placed into a knowledge base.
He added that one of the talent management team's daily jobs now is to continue tagging many of the creators the firm has worked with. The work is tiring. Many records do not show obvious value right away, and no tagging system will be perfect on the first pass. Some details stored today may not be useful for half a year. Even so, he said they should still be kept. A company's real assets, in his view, are not limited to cash in the bank, equipment in the office, or the employee roster. They also include the data and context accumulated over years of real operating activity.
Models can be bought. Tokens can be bought. Codex is not unique to one firm. But the context a company leaves behind over countless specific days, he wrote, is almost impossible to buy from outside. That is why he pushes an internal strategy of storing everything that should be stored.
He was careful to add that "store everything" does not mean dumping everything into a knowledge base or a table and calling it done. Real organizations are full of conflicts. Old rules sit next to new ones. Two departments may use different definitions for the same metric. Obsolete contract templates may still rank near the top of search results. An agent will not absorb that mess and fix it for the company. It will pick something that looks like an answer and then carry the mistake forward at AI speed.
Because of that, he said stored data needs a source, a timestamp, and a person responsible. Outdated material has to be retired. Conflicting standards need someone to rule on them. Decisions that affect money, contracts, and people need regular checks and data cleaning. He described this kind of work as the dirty, exhausting labor behind AI adoption, but also as a necessary condition for a real leap in organizational AI capability.
As routine work fades, jobs start to look more human
Once data starts to accumulate and agents move into each role, the essay says something else happens: employees do not become more machine-like. They become more human.
He used business development as one example. In the past, a business employee might spend 50% of the time sorting data, building spreadsheets, checking materials, and revising contracts, and the other 50% meeting clients. With agents taking over much of the first part, that workload may shrink to 20%. The remaining 80% can go into seeing clients, understanding what they are really worried about, and making proposals more complete and more careful.
He said talent managers are seeing something similar. They used to spend large amounts of time collecting data, sorting records, and confirming details repeatedly. As agents take on more of that work, they can spend more time speaking with creators, listening closely to their recent situation, and maintaining relationships that are hard to measure. He said the same pattern applies across HR, legal, content, and operations.
His broader point is that AI is best at consuming work that can be described, repeated, and verified. Once those layers are peeled away, what remains is the hardest part to accelerate: sincere communication between people. For a company with a fairly traditional and almost entirely ToB business model, he wrote, that is not an abstract idea. It shows up in simple questions. After a problem appears, will the client still pick up your call? Will a creator really entrust the next few years of their career to you?
Codex can help prepare materials, remember details, and review past conversations, he wrote. It cannot experience time on your behalf, and it cannot replace face-to-face communication. That is why he says trust and brand are more valuable than diamonds in the AI era. They are painfully slow to build. You gain one point when you do what you promised. You gain another when something goes wrong and you do not run away. You gain another when the other side is having a hard time and you actually show up instead of sending a polite message. AI gets faster, he wrote, and that makes these things more expensive, not less.
This is the logic behind his line that the more agents the company uses, the more traditional it becomes. Business staff start to look more like the old model of people who go out, sit with clients, and talk. Talent managers start to look more like people who truly know creators and stay with them over time. Managers, meanwhile, have less room to hide behind reports. They have to face conflicts, make decisions, and carry the outcome.
What happens to the time that AI saves
The essay then turns to a question he called harsh but necessary: where does the time saved by agents actually go?
He wrote that many companies love to measure AI adoption in labor hours saved. A process that used to take four hours now takes 20 minutes. A role that used to serve 20 people now serves 50. A draft that used to take two days now appears in half an hour. Those numbers matter, he said. But if the freed-up time is swallowed by more meetings, more reporting, more approvals, and more forms nobody reads, then the company has only become busier.
He put the problem in personal terms. If an employee used to handle five things a day and, after getting an agent, is now expected to do 20 things a day, all of them immediately, that person will not feel liberated by technology. They will feel that the whip has only gotten faster. He thinks this is one of the easiest blind spots for managers, because from the company's side, higher efficiency naturally feels positive and opens the door to more ambition.
He admitted he sees that tendency in himself as well. Once the company's capacity expands, it becomes tempting to launch one more project. If talent managers can serve more creators, the company wants to sign more creators. If the business team can absorb more information, it wants to meet more clients. If content production speeds up, it wants to cover more topics. The capacity freed by AI quickly gets filled by new ambition. He said that is probably one reason the headcount keeps rising and the company needs another office move. AI does not necessarily make a company smaller. It may enlarge the ambitions of the company and its founder first.
At the same time, he used that point as a warning to himself. Ambition is not wrong, he wrote. A company should move forward. But if every gain in efficiency turns only into bigger numbers, denser schedules, and more work, then AI transformation becomes little more than a harder-to-refuse version of overtime for ordinary employees.
Because of that, he said he no longer wants to ask only how many hours agents saved. He wants to ask who got those hours back. His answer was specific: give business teams more time for clients, give talent managers more time for creators, give HR more time for employees who truly need to be heard, give legal teams more time for difficult judgment calls instead of a 27th formatting revision, give content teams more time for experience, curiosity, and stories worth writing, and give ordinary people back some time for themselves. That extra time could go to learning, doing better work, or simply getting home early enough to eat a proper meal.
Managers often reduce employees to output, he wrote, but people are not batteries waiting to be drained by agents. In his view, the best organizational use of AI is one that helps people feel more genuine interest in their work and more happiness doing it. That is also what allows a company to stay curious about the world and build trust outside the organization.
Managers have fewer places to hide in the AI era
The fifth section of the essay focuses on management itself. He wrote that in the past, managers could prove their value through visible action: holding meetings, chasing progress, collecting daily reports, approving work, and breaking one task into 10 steps before checking whether every person followed each one.
As agents take over large amounts of execution work, that style of management starts to look awkward. Employees with Codex can research material, prepare proposals, write scripts, and get a workflow moving even when that same process once needed several departments. In that setting, he argued, a manager's real job is not to produce more motion. It is to answer harder questions. What is the actual goal? What absolutely cannot go wrong? Which risks can the company accept? What counts as a complete result? Who makes the final call when something unexpected happens? Who takes responsibility when things fail?
Those tasks, he wrote, are real management. They are difficult, and they do not fit neatly into a polished slide deck. A manager who cannot write prompts still has time to learn. A manager who cannot define goals, make judgments, or take responsibility when something breaks will not be saved by any agent.
He went even further, saying that the biggest shock AI delivers to managers may have little to do with whether the tools are new. Instead, AI exposes incompetence that used to stay hidden inside process. In the past, it was easy to blame weak execution, poor information, or not enough staff. But if the agent has already found the information, drafted the plan, and lowered the cost of execution, then the decision that still never gets made becomes much harder to excuse.
He also applied that standard to himself. He said he cannot ask everyone else to create actively while insisting every detail match his own preferences. He cannot say he values outcomes and then judge people by overtime hours, reply speed, or how busy they look. He also cannot push vague goals downward and then cover his own management failure with a line about learning to use AI. In his view, those are all signs of managerial weakness.
That is why he believes the shape of an AI-era organization often depends on what the company already was. A company that does not trust people will use agents to make surveillance more detailed. A boss who likes control will use agents to issue commands faster. Only a company that respects people will truly put agents in the hands of ordinary employees as tools. AI does not automatically produce advanced management, he wrote. It reveals what was already there.
New employee training is getting harder, not easier
The sixth part of the essay deals with newcomers, which he said is one area where his company is still struggling and still experimenting. The basic problem is simple: once a new employee arrives, the old training path no longer works cleanly.
He gave several examples. A junior content worker used to begin with research, headline revisions, case collection, and first drafts. A junior business employee might start by organizing client files, sitting in on meetings, writing notes, and revising proposals. A junior legal employee would begin with basic contracts and clauses. These jobs were tedious and sometimes painful, but many people developed professional instinct through that slow, repetitive work.
Now an agent can hand a new employee an 80-point answer in minutes. In the short term, that feels great. Someone in their first week may produce work that once took six months to reach. The company sees lower training costs. The new hire feels stronger right away.
But he argued that better output does not mean real growth. If a newcomer never goes through those foundations, they never get the chance to understand why the agent produced a certain answer. When AI returns something that looks correct but is actually wrong, that person may not even know enough to doubt it.
He wrote that he is not most afraid of new employees failing to use AI. What worries him more is a new employee who can only use AI and never develops judgment of their own. Newcomers are always the weakest group in an organization, he said. They are often unsure what they are allowed to ask, which rules are outdated, and whether a leader saying "handle it yourself" means real decision-making authority or not. The agent gives them output, but not necessarily the ability to bear consequences.
If the company only looks at results, he said, a new hire may be pushed along by that 80-point answer until one day they make a serious mistake in a critical place, at which point the company turns around and asks why they did not understand the issue. For that reason, he believes the current shift is not good for the growth of new talent.
He said he has been thinking about how to redesign the path for new employees. That does not mean forcing them back into meaningless drudge work or pushing them into a manual era on purpose. It means spending more time asking them to explain why the agent made a choice, compare different options, meet clients directly, see the consequences of mistakes, make a judgment while someone experienced is still there to support them, and put their own name on the final output.
A newcomer, he wrote, needs more than speed. They need a safe way to make several mistakes. They need correction from someone who truly understands the work. They also need to know that one day they can become the person who supports others. Giving a newcomer an agent that can produce an 80-point answer is easy. Giving them a path to become highly skilled is the real management challenge.
AI should remove friction, not remove people
The final main section returns to the essay's central claim: the more AI the company uses, the more traditional it looks. His explanation is that once AI accelerates everything that can be accelerated, the parts that are difficult and cannot be sped up become visible at last.
He framed that contrast plainly. Data can be organized automatically; trust cannot. Contracts can be generated quickly; responsibility cannot. Genuine communication between people cannot be replaced. Client information can certainly be analyzed, but whether a client still trusts you when bad news arrives is not something pure analysis can calculate, he wrote. These are old, slow, and unspectacular matters, but the survival of a small company often depends on them.
He said his company is not one of the businesses protected by a technical moat. It is a small company trying to carve out a piece of business in this era, support its people, and find a way to survive. The company has stayed alive, he wrote, because clients still hand over budgets, creators still entrust their careers to the team, partners are still willing to work together on offline events and even variety shows, and colleagues still believe in the idea of "connecting everything in the AI era."
For that reason, he increasingly believes the real value of data, agents, and automation inside an organization is not to remove people from the middle. Tagging creators is not about reaching a point where the company no longer needs to know them. It is about helping a talent manager understand what that creator has experienced and what they need now before the meeting begins. Having the business team use agents to sort client materials is not about never meeting clients again. It is about making sure the time in front of the client is not wasted on homework that should have been done already.
He said the same logic applies to storing meetings, documents, and SOPs. The point is not to make the organization depend only on systems. It is to ensure that someone newly joining the company does not have to lower their head and ask everyone for help, or repeat every mistake that earlier employees already made. His summary was short: AI should remove unnecessary friction between people, then give one person more time to actually stand in front of another person.
That, he wrote, is the company's real nature. It is not a super company running by itself in the cloud with a few dozen digital employees. It is a group of ordinary people using the most advanced tools available in this era to do very traditional things well: serve a client well, support a creator well, train a newcomer well, keep a promise, and make life a little better for the people working together.
In that sense, becoming more traditional as AI use rises is not a step backward. It is what happens when technology strips away the outer layer of efficiency and leaves the most basic part of work in view again.
His closing conclusion
Near the end, the author looked back on the past two and a half years and wrote that the company has been through many life-or-death moments and many wrong decisions. It is still alive, and it is moving toward a bigger office, but he said he would not claim to have found the correct answer.
He said he does not know whether full-company AI adoption fits every firm. He does not know whether a "thin middle platform, thick front line" structure is the best organizational form for the AI era. He does not know whether the growth path his company is trying to design for newcomers will really work. But there is one point he now feels increasingly sure about: organizational change in the AI era may appear to be about tools, workflows, data, and efficiency, but the real test is still how a company treats people.
He broke that down into a set of practical questions. Is the company willing to give creative power to the front line? Is it willing to give the time saved by agents back to clients, creators, employees, and life outside work? Is it willing to give newcomers room to make mistakes and grow before they are fully mature? And when things go wrong, is it willing to step forward rather than hide behind process and reporting?
He ended by quoting Antoine de Saint-Exupery from Wind, Sand and Stars: 「The greatness of a profession may first lie in its power to join men together. There is only one true luxury, and that is human relationships.」 He wrote that the line may once have sounded merely romantic, but after two and a half years of running a company, surviving moments when the company might not have made it, and watching the people around him gradually become more numerous, it now feels true.
He also said the company may later launch special columns to share how colleagues across finance, legal, HR, operations, business development, and other roles are using AI in their day-to-day work, adding that he still finds those practices instructive when he watches them.