For the last few years, we have been obsessed with one question: when does AI become intelligent enough to count as AGI?
It is an interesting question. I am increasingly convinced it is not the question most people should be worrying about.
Because while we debate definitions of intelligence, something much more practical is happening. AI is learning to use the same computer you use to earn your salary.
[Reported evidence] In its GPT-6 Astra announcement, OpenAI reports 72.6% on OSWorld 2.0, compared with 65.7% for GPT-5.6 Sol. Its latency simulations put Astra at roughly 40 minutes per task versus 75 minutes, about 47% less time. These are vendor-reported results under benchmark conditions, not an independent measurement of productivity in your office. A 72.6% benchmark score does not mean 72.6% of your job can be automated.
[Reported evidence] Anthropic positions Claude Sonnet around browser and computer use, while Google has integrated computer use into Gemini 3.5 Flash. These product announcements establish a shared direction. They do not prove equal capability or reliability across the same work.
I think the prize is much bigger than a better chatbot.
Intelligence was never the only bottleneck
For this argument, think of knowledge work as three connected activities: decide what should happen, execute the work, and verify the result.
[Inference, based on the observed progression from chat products to action-taking tools] Much of the early appeal of generative AI was help with the thinking around a task: explaining, summarising, brainstorming and recommending. But there was still an awkward gap between the answer and the outcome.
Someone still had to open the spreadsheet. Find the right file. Log into the CRM. Build the slides. Move information between systems. Change the website. Test the output. Send the email. Follow up.
In other words, AI could tell you what to do, but a human still had to translate intelligence into clicks.
Computer use targets that translation layer: operating an interface through actions such as clicking, typing and navigating. It is one route to execution. Direct software connections, often called APIs, and scripts are other routes. An effective workflow may combine them; there is no prize for using a mouse to do something a dependable integration already handles.
Why this feels different to me
Blender is a good example. I have always found tools like Blender fascinating, but practically inaccessible. The barrier is not imagination. I can describe what I want. The barrier is knowing how to navigate an enormous professional interface well enough to turn that idea into a finished 3D scene.
Imagine telling an agent: create a realistic 3D product scene, use this reference image, build the lighting around this mood, animate a slow camera movement, render three variations, compare them and keep the strongest one.
This is an illustrative workflow, not something I am claiming to have completed. The attraction is that I could spend more attention on the outcome and less on memorising the interface. I would still need to recognise bad geometry, judge the lighting and decide whether the final scene serves the brief. Access to production does not automatically supply taste.
That distinction is enormous.
The commercial that used to require a small production chain
Take the idea further. Imagine a small brand wants to launch a new shoe.
The brief could start as one sentence: create a 20-second commercial for this shoe aimed at urban runners in Dubai, using a premium night-time visual language.
A capable computer-using system could eventually research references in the browser, organize assets, generate a storyboard, work inside creative applications, produce or assemble scenes, create voice and music options, edit the footage, export different aspect ratios, prepare ad copy, upload draft campaigns and return performance-ready variants for human approval.
That is an illustrative end-to-end scenario, not a claim that GPT-6 can autonomously produce a flawless commercial across arbitrary software today. The steps still involve different tools, permissions, assets and review standards. A shoe that looks beautiful but changes shape between shots is not a production success. Someone must check product accuracy, usage rights, brand fit and whether the finished ad communicates anything worth paying to distribute.
Or consider the boring work, which may matter even more
The spectacular examples get attention. The mundane ones probably change companies faster.
Consider this illustrative request from a regional manager: “Find every store whose conversion has fallen for three consecutive weeks, compare the pattern with footfall and staffing, update the tracker, prepare a one-page summary for each store, and draft the follow-ups for Friday.”
That request currently crosses analytics, spreadsheets, email, calendars, dashboards and human coordination. Traditional automation can handle parts of it, but usually only after someone designs the workflow in advance.
The interesting possibility with computer-using agents is that the workflow can become more dynamic. The person describes the outcome; the agent figures out more of the path across the existing tools.
The useful ambition is delegation across a whole outcome. But the quality of that delegation depends on the questions hidden inside the request. Does conversion use transactions or purchasing customers? Are the weeks comparable? Is staffing actually a plausible explanation, or simply the nearest available column? Clicking through the systems will not settle those questions.
Why OpenAI, Anthropic and Google will fight hard for your mouse
A model that answers your questions is useful. A model that can operate your software sits much closer to the economic output of your work.
If it can move between browser, spreadsheet, CRM, creative tools, internal systems and communication apps, the AI becomes something closer to an execution layer between human intent and software.
[Inference, based on these observed product announcements] I expect computer use to become one of the hardest-fought areas of frontier-model competition. The commercial attraction is plausible: the system that completes a workflow has more opportunities to become part of the next one. It can sit close to the files, preferences and review habits that make work repeatable. This is my interpretation, not a confirmed statement of the companies’ internal strategies.
Your competition may no longer be the person at your level
This is where the conversation becomes uncomfortable for people higher up the career ladder.
Titles have never perfectly contained professional competition. Still, seniority often comes with access to teams, budgets and specialists. Those resources let an experienced person turn an idea into something substantial. That execution advantage is the part worth watching.
AI does not erase experience. Judgment, relationships, pattern recognition, organizational context and accountability remain difficult to manufacture. But AI can compress some of the execution advantages that used to come with seniority.
A highly motivated junior employee may not suddenly acquire fifteen years of judgment. But if that person can research like a small team, analyze data, produce creative work, build simple software, operate business applications and repeatedly test ideas, the amount of output available to them changes.
[Inference, based on the observed expansion of coding and computer-use tools] The same possibility extends beyond employment: a founder can attempt a prototype, a marketer can explore a 3D scene, and an operator can test a workflow outside their original specialism. That widens the range of experiments they can attempt. It does not establish that they can deliver specialist quality without specialist review.
The boundaries do not disappear. They become more porous.
The new inequality may be agency
This is why I do not think the future simply belongs to whoever has the best AI model. If everyone eventually gets access to similar intelligence, access itself stops being the advantage.
Give 100 people the same model and you will not get the same 100 outcomes.
One person asks it to make a presentation. Another builds a system: research the problem, challenge the assumptions, analyze the evidence, execute the work, verify the output, document what happened, improve the process and run it again.
Same model. Very different leverage.
My bet is on people who combine expertise with agency: the ability to decide what should happen, organise an attempt and keep going until they have something worth evaluating. The advantage is not the volume of output. It is the ability to turn feedback into a better next attempt.
Persistence becomes strangely valuable again
AI is usually sold as a way to do less. I think the larger advantage may go to people who use it to attempt more.
Before AI, many ideas died at the sentence: I do not know how to do that.
I cannot code. I do not know Blender. I am not an analyst. I cannot automate that. I would need a designer. I would need a developer. I would need a team.
Those statements are becoming less final.
Person A uses AI to save ten hours and stops. Person B saves ten hours and reinvests them into another experiment. Then another. Then builds a reusable system around the experiments.
In this hypothetical comparison, repeated useful experiments create more opportunities to learn. Repetition without feedback can just produce a larger folder of unfinished work. The distinction matters for anyone tempted to confuse being constantly busy with making progress.
That is why I would reinvest some saved time into deliberate practice. Track which attempts were worth making, what failed and what became reusable. I explored the other half of this choice in outsourcing effort while keeping the thinking: some saved time should also become a better life, rather than another obligation.
Systems thinkers may become disproportionately powerful
My systems-thinking article asks readers to look beyond the visible task to the relationships that produce the result. Computer use makes that a useful discipline to practise: the request crosses tools, but the outcome depends on definitions, permissions, handoffs and feedback.
That is how you delegate well to humans. It is increasingly how you delegate well to machines.
The valuable person is not necessarily the one who remembers every menu inside Excel, Salesforce, SAP, Blender or whatever application comes next. It is the person who understands the system around the work.
What outcome are we trying to create? What information is required? Which systems need to interact? What could go wrong? Which decisions require human judgment? How will we know the result is correct?
The more execution AI absorbs, the more those questions matter.
Try one complete workflow before buying the bigger promise
Return to the regional manager’s request. Here is how I would turn that hypothetical brief into a bounded experiment, using sample or approved data and keeping communications as drafts.
| Part of the brief | What to make explicit | Evidence to review |
|---|---|---|
| Outcome | Identify stores with three consecutive weekly conversion declines | A shortlist linked to the underlying weekly figures |
| Context | Agree on the conversion formula, reporting weeks and comparable stores | Definitions and data dates alongside the analysis |
| Boundaries | Read approved exports; edit a copy of the tracker; draft follow-ups | Original files preserved and no messages sent |
| Execution | Compare trends, prepare summaries and record missing information | A reviewable tracker and one-page store notes |
| Verification | Recalculate one flagged store and one unflagged store independently | Matching calculations and an explanation of any differences |
| Feedback | Record corrections, review time and unresolved exceptions | A decision to repeat, revise or abandon the workflow |
Notice that “compare staffing” is not permission to conclude that staffing caused the decline. The summary should distinguish a pattern from an explanation and leave the manager with questions worth investigating. A neat document is only useful if the reasoning survives inspection.
The experiment should end with an accepted result and a record of what it cost to obtain. If checking every step takes as long as doing the work yourself, you have learned something valuable. Narrow the task, improve the inputs or choose a simpler automation. Full-scale automation should be earned through repeated evidence, not inferred from one successful demonstration.
This connects to the argument in everyone can build software now: making something run and making something useful are separate achievements.
The career ladder starts looking more like a graph
The traditional career ladder assumes capability accumulates in a relatively predictable sequence: junior, senior, manager, director, executive. Experience still matters, but AI gives people more ways to move sideways into capabilities that previously required formal specialization.
[Inference, based on the observed expansion of tools available to individual builders] I expect some boundaries around titles to weaken. A junior colleague with strong judgment and a well-designed workflow may produce a result that previously needed help from several disciplines. That is a possibility to prepare for, not a forecast that every junior employee will outperform a director.
That should create some FOMO for senior leaders, but not panic. Senior people still possess something extremely valuable: accumulated context. The mistake would be assuming context alone remains enough while execution capacity becomes dramatically cheaper.
If you are senior, use the advantage you already have
The opportunity for experienced people is actually enormous. You already have judgment, relationships, organizational knowledge and pattern recognition. AI can multiply those advantages if you learn to delegate to it properly.
The opportunity cost of waiting is practice. A new subscription can be bought later; the lessons from briefing badly, checking carefully and learning where delegation breaks down have to be earned. Senior people have an advantage here if they bring their experience into the experiment.
I would worry less about knowing the newest prompt and more about having no firsthand experience of directing a complete piece of work. You cannot judge the usefulness of delegation solely from a launch video.
Pick one recurring workflow that ends in a reviewable draft. Write down its inputs, boundaries and acceptance criteria. Try it on a small example. Count briefing time, waiting time, corrections and review, as well as execution. Keep the workflow only if the accepted result justifies the total effort.
You do not need AGI for this to become disruptive
We can spend years arguing about exactly when artificial general intelligence arrives and whose definition counts. You do not have to settle that debate before investigating whether an agent can help with a useful piece of work.
For most careers, another question may prove more immediate: what happens when intelligence can operate a computer?
If execution becomes cheaper and more dependable, the practical question is what you will choose to attempt with it. More output is one answer. Better decisions, faster learning and work that previously felt inaccessible are more interesting answers to me.
The advantage then moves toward people who know what they want, understand systems, exercise judgment and keep experimenting long after everyone else has stopped playing with the new tool.
You do not need AGI for that world.
Evidence note: Product claims above are attributed to their vendors and were checked on 7 September 2026. The Blender, commercial-production and retail examples are illustrative, not reports of completed deployments. Career implications are the author’s interpretation.