aiseed.dev›AI-Native Software Development
3-04 / Series
3-04 № 04 · 2026

With the effort of outsourcing,
you can build it yourself.

The upstream judgment outsourcing cannot remove, and the de-responsibilization and hollowing-out it brings — for the same effort, you can build it yourself

2026.07.21 🌐 日本語

The SIer commission model has become structurally uneconomic. The previous chapter (3-03) showed that the premise on the office side — dependence on Microsoft — has inverted. This chapter takes the other side, the core (SIer) side. Even when you outsource, *the upstream judgment stays with the customer*. And on top of that, commissioning hollows out responsibility and capability. This chapter decomposes the commission process step by step and shows that structure. Its real focus is structure, not cost.

The Premise Has Inverted — You Can Now Hire the Excellent Engineer as AI

Commissioning an SIer — the system integrator a company hands its core systems to — used to be rational. Building software takes excellent software engineers across many fields: design, databases, frontend, infrastructure, security. Holding and retaining all of them in-house, company by company, was not realistic. Gathering the talent in one place, concentrating it in an SIer, was the more efficient path.

But that premise has inverted. *You can now hire the excellent engineer as AI.* And so the customer can become the builder. As 1-05 showed, you use OSS for the general part and build only the firm-specific logic with AI.

This chapter takes up the other side. Why is "commissioning an SIer makes life easier" now an illusion? We check it by decomposing the commission process step by step.

The Process Does Not Begin With Requirements

The process does not even begin with requirements. Before that, you have to think through how to improve the business operations themselves: the workflow, the approval steps, the paper and the rules, including the parts that have nothing to do with software. What should become a system is decided only after that. This is something only the customer understands, and it cannot be dumped on the SIer.

And improving those operations does not go well unless you understand what systems can do. Only once you can see what can be automated and where data connects do you know how to restructure the work.

The more upstream the judgment, the more it needs an understanding of systems. Which is exactly why that understanding must be held in-house.

From Upstream to Implementation, It Only Runs as a Loop

From upstream to implementation, the work does not go in a straight line. You build, you try, you restructure the operations from what you see, and you build again. It can only be run as a loop.

The SIer cannot do this. Every turn needs budget approval and a contract, so it usually runs to a year. That is also why almost no one develops business systems agile: the gates of budget and contract forbid the loop.

But with AI you can run this loop fast. Implementation comes back at once. What takes time is the testing. Testing cannot be fully automated. Whether it runs can be measured automatically, but whether it is what you actually wanted, and whether it fits the operations, only a human can confirm. That check is the customer's own judgment, the part you cannot send outside.

Even so, a turn is an order of magnitude shorter than the SIer's turn of a year. You can run it again and again. Only in-house building, with AI in your own hands, can do this.

Moving a Single SIer Engagement Takes This Process

On top of that, moving a single SIer engagement requires a process like this.

This is the proper form. In reality, though, people skip the process and settle for "we left it to the SIer." Requirements, acceptance, judgment — all handed over whole. It looks easier, but this is where de-responsibilization begins. As the next section shows, that is the deepest problem with commissioning.

flowchart TB subgraph Sier["SIer commission model — process steps"] direction TB S1["Requirements / RFP
(customer: weeks to months)"] S2["Vendor selection
(customer: weeks to months)"] S3["Contract negotiation
(customer + SIer: weeks)"] S4["Project management
(customer + SIer: full duration)"] S5["Acceptance / UAT
(customer: weeks)"] S6["Ops & maintenance handover
(customer + SIer: ongoing)"] S1 --> S2 --> S3 --> S4 --> S5 --> S6 end subgraph AI["AI-native process steps"] direction TB A1["Customer + AI on requirements
and design
(days to weeks)"] A2["AI implements
(at once)"] A3["AI + humans verify with tests
(not all automatable)"] A4["Evaluate and integrate
(continuous)"] A1 --> A2 --> A3 --> A4 A3 -->|fix and re-run| A2 end classDef good fill:#e8f5e9,stroke:#7a9a6d,color:#3a4d34 classDef bad fill:#fef3e7,stroke:#c89559,color:#5a3f1a class AI good class Sier bad

With the effort you spend on the outsourcing process, you can build it yourself.

Commissioning Erases Responsibility and Hollows Out Capability

The deepest problem with commissioning is not labor or cost. *Commissioning erases responsibility.* The moment the side that builds and the side that delegates split apart, "whose responsibility is this?" floats free. The delegating side says "we left it to the experts"; the delegated side says "we built to spec." No one owns the result whole. This is de-responsibilization.

And de-responsibilization breeds the hollowing-out of capability. What no one is responsible for, no one cultivates. Keep sending the work of building outside, and the capability never grows inside you. Eventually you lose even the eye to tell good from bad.

A Hollowed-Out Organization Cannot Stop Elementary Mistakes

The most expensive example is GitHub Copilot. Microsoft did not build the AI core itself; it handed it to OpenAI. The power to build sat with OpenAI while the product's responsibility sat with Microsoft, so responsibility split across two companies. And during that time, CEO Satya Nadella was thinning out his own basic research. Microsoft Research was one of the most prestigious basic-research labs in the industry, home to a Turing Award winner. Yet in his first year as CEO he closed the MSR Silicon Valley lab in 2014 (about 50 researchers), and in 2023 he disbanded the entire AI ethics team, prioritizing shipping OpenAI's models "to customers at very high speed." Long-term research was traded for speed.

An organization whose technical capability has hollowed out cannot stop elementary mistakes. Copilot was trained on public GitHub code. That code holds excellent work, but it also holds a vast amount of garbage. That what you train on decides the output is the most basic common sense of machine learning, yet the researchers who could see this and select good code were no longer at the center of the decision.

The result showed in dangerous code.

This is the mistake the world's largest software company made after scattering responsibility and hollowing out its own capability. But scattering responsibility does not make it disappear. The final decision rests with the CEO. The responsibility for this failure is Nadella's.

SIer commissioning has the same structure, only at a different scale. The customer says "we left it to the SIer"; the SIer says "we built to spec." Responsibility floats free, and the customer loses the ability to make use of software. The deeper the commissioning, the deeper the hollow.

So the answer is in-house building, the customer becoming the builder (1-05). Only taking judgment and capability back into your own hands stops the de-responsibilization and the hollowing-out.

The Disappearance of the SIer Model Is Inevitable

The cost of commissioning has reached parity with building in-house; commissioning erases responsibility and hollows out capability; and only in-house building can run the loop. Once you reach this point, *the disappearance of the SIer commission model is inevitable*.

Most of the work — the standard work AI can write — moves to the customer side. Even the specialized remainder (genuinely new technology, specialized regulation, scale-driven design) can no longer be sustained as a "multi-year operations commission." It turns into hourly consulting, or it is absorbed into the customer's in-house team (3-06). Neither is the SIer commission model itself.

Transition speed, Japan-specific dynamics (multi-tier subcontracting), and labor mobility are taken up in 3-07 and 3-09.

Summary

This chapter decomposed the premise on the core (SIer) side. Even when you outsource, the upstream judgment — improving the business operations themselves, and understanding the systems — stays with the customer. That judgment does not advance in a straight line; it advances only as a loop. And only in-house building, with AI in your own hands, can run that loop. On top of that, commissioning erases responsibility and hollows out capability. With the effort you spend on the outsourcing process, you can build it yourself.

But even when disappearance is inevitable, customers cannot always move at once. The existing commission relationship remains as lock-in. The next chapter takes up that lock-in (3-05).

Related articles