Something strange happens when AI makes small software projects much faster.
A small online service can be almost finished in two days, while the outside world is still saying: “Under review,” “Identity verification in progress,” and “Please allow several business days.”
That reverses the old order.
Before, building the product could easily take longer than getting an account or application approved. Now the product may be ready while the approvals are still sitting in someone else’s queue.
This is not a claim that every service can be built in two days. It is one small web-service example where AI handled a large share of research, coding, document drafting, and revisions.
But the pattern is worth watching:
As AI gets faster, the slowest part of launching a service may become human and organizational waiting time.
1. “Several business days” can be longer than the coding
Shipping a service takes more than code.
There may be identity checks, payment-account verification, affiliate or advertising review, domain verification, access requests for outside services, government filings, and requests for additional documents.
None of these procedures is unusual.
What changes is the comparison.
AI can draft something in minutes, revise it in tens of minutes, and push a large chunk of implementation forward in hours. Meanwhile the approval page says, “We usually reply within three business days.”
Code moves in hours. Approval moves in business days.
That difference in time scale becomes the new waiting room.
2. “Wait for approval, then build” gets reversed
The old cautious sequence was simple: get approved first, then invest time in building.
If the core product can be built quickly, the order changes.
Apply first. Build while waiting. Keep outside integrations replaceable. Insert IDs, keys, or account settings after approval. Switch providers if the application fails.
In other words, overlap approval time with development time.
Approval is no longer dead time. It becomes a parallel track.
The important operational habit is not refreshing the status page every five minutes. It is keeping a simple list: applied, under review, more information needed, approved.
3. AI can build government documents from something close to a blank page
One of the more surprising examples came from a government filing that required a network architecture diagram.
You might expect the official form to contain a polished diagram template.
Japan’s official Form 3 for this filing does not. The e-Gov PDF is a single A4 page with the title “Network Configuration Diagram” and six notes explaining what the applicant must show, such as communication flow and outside network providers.[1]
There is no ready-made set of boxes and arrows.
The applicant has to decide how to show users, the internet, cloud services, databases, notifications, and the direction of data flow.
AI can read those requirements, extract the important parts, and turn them into a document that a reviewer can understand at a glance.
That is different from “filling in a form.”
It is closer to reading the rules and designing the actual document from scratch.
4. The machine creates; the human reviews
Then comes the part that feels upside down.
AI builds the diagram. AI drafts the explanation. AI looks for missing pieces. AI revises the output.
Then a human checks it.
Does this match the real service? Is the description accurate enough to submit? Can someone take responsibility for the final version?
The historical workflow was human creation followed by human review.
The emerging workflow is increasingly AI creation followed by human review.
Anthropic released Claude Opus 5.5 on September 22, 2026 and markets it for long-running coding and knowledge-work tasks. In Anthropic’s own measurements, it generates output more than 30% faster than Opus 5 and costs about 40% less on typical workloads. The company also shows examples involving spreadsheets, analysis, and executive presentations.[2]
Those are vendor-reported results, not proof that every office will see the same gains.
But they are a useful sign of how far the “AI creates, human checks” pattern is expanding.
5. Does that mean office jobs disappear?
It is tempting to jump from this to “office jobs are over.”
The evidence is more nuanced.
The International Labour Organization’s 2025 update found that clerical occupations remained the group most exposed to generative AI. Data entry clerks, typists, accounting and bookkeeping clerks, and administrative secretaries are among the roles with particularly high exposure.[3]
But the ILO does not conclude that most of those occupations simply vanish.
About one in four workers worldwide is in an occupation with some degree of generative-AI exposure, yet the ILO says that because human input remains necessary, job transformation is more likely than wholesale replacement for most work.[3]
The first thing to disappear may not be the job title. It may be the repetitive tasks inside the job:
drafting standard text, formatting documents, copying information, building comparison tables, producing first drafts, making everything look consistent.
What remains is more likely to include:
checking facts, handling exceptions, coordinating with people, taking responsibility, approving a final action, and stopping the system when the AI is wrong.
Before “office work” disappears, office work may shift from creation toward verification and judgment.
6. The new bottlenecks are outside the model
If AI speeds up code, writing, and document production, the next constraints move outside the AI.
Reviews. Identity verification. Contracts. Banks and payment systems. API access. Search traffic. Advertiser approval. User response. Trust. Legal responsibility.
Doubling model speed does not automatically double the speed of those things.
A reviewer still needs time. Identity checks still have rules. You cannot know user demand before users arrive. Trust does not compound overnight.
So the competitive skill changes.
It becomes less about “How fast can I type the code?” and more about:
What should I apply for first? What can I build while waiting? Where must a human sign off?
7. When building becomes cheap, you can choose the winners later
Software projects used to be expensive enough that people tried very hard to pick the winner before building.
That makes sense when every idea costs months.
But if a small experiment can be built quickly, the strategy changes.
Build the interesting idea. Get it to a testable state. Put it in front of reality. Stop the dead ones. Grow the ones that show signs of life.
Do not predict the winner perfectly before building. Build cheaply, then keep the winners.
This does not mean flooding the world with low-quality junk.
It means lowering the cost of each experiment so real feedback arrives sooner.
As creation becomes cheaper, the scarce resources become the idea, distribution, user response, trust, and the judgment to decide what deserves more investment.
8. Faster building does not mean skipping the safety gate
There is an important limit.
If a failed prototype can be rolled back, moving quickly is reasonable.
But speed is not an excuse to skip review when a project touches personal data, payments, safety, regulation, shared assets, someone else’s reputation, or content that is hard to retract.
Software can be rolled back faster than trust.
The faster AI gets at creating things, the more important it becomes to decide where humans must deliberately slow the process down.
Speed and lack of review are not the same thing.
9. Free time becomes option value
When the cost of making things falls, spare time changes meaning.
It is not merely empty time.
It becomes the ability to test an idea the moment it appears.
You can work from home. You can do a little from somewhere else. You can capture an idea while moving around. And you can choose not to work on it today.
If the building process itself is enjoyable, that matters even more.
Revenue and approvals may take time, but if experimentation feels like play, the cost of continuing is much lower.
For an individual builder in the AI era, being willing to keep experimenting because it is fun can become a genuine advantage.
10. The faster AI gets, the more humans move toward deciding
AI can now push a small service from idea to working form very quickly.
It can even turn sparse administrative requirements into a readable document.
Meanwhile, identity checks, reviews, approvals, and accountability still run on human and organizational time.
That pushes the center of human work toward:
What should we build? What must be verified? Where should we stop? Which experiments deserve to survive?
Office jobs are not disappearing overnight.
But the relative value of humans manually producing every first draft, table, diagram, and formatted document is falling.
And for people building services, the next competition is already visible:
Pick better ideas. Parallelize outside waiting time. Keep human review where consequences are hard to reverse. Stay in the game long enough to find the winners.
Welcome to the world where the code is finished before the “three business days” are over.
References (3)
- e-Gov, Telecommunications Business Act Enforcement Regulations, Form 3: Network Configuration Diagram
https://laws.e-gov.go.jp/data/MinisterialOrdinance/360M50001000025/622223_2/pict/2FH00000068245.pdf - Anthropic, “Claude Opus 5.5”, September 22, 2026
https://www.anthropic.com/claude-opus-5-5 - International Labour Organization, “Generative AI and jobs: A 2025 update”, May 20, 2025
https://www.ilo.org/publications/generative-ai-and-jobs-2025-update
