I have used AI in two very different contexts.
First, I used AI to build an OSS project largely by myself.
Second, I applied AI to brownfield development inside an organization.
In the second case, I did not use AI only for code generation.
I used AI across a much wider part of the development process:
- source code
- design documents
- implementation plans
- test specifications
- test cases
- release procedures
At first glance, this may sound as if AI can take over the entire development process.
But that was not the lesson I learned.
The more I used AI across these activities, the clearer the boundary became.
AI was very effective at generating drafts, connecting scattered information, translating context, and preparing artifacts for the next step.
However, AI could not be treated as the source of organizational responsibility.
When AI generated a design, the design still had to be checked against existing rules and constraints.
When AI created a test specification, the coverage still had to be judged against the real change intent and risk.
When AI prepared a release procedure, the procedure still had to fit the organization’s approval process, operational constraints, and rollback policy.
In other words, AI could help produce and transform work artifacts, but the structure that makes those artifacts valid had to remain outside AI.
That structure is the organizational backbone.
It is made of:
- rules
- workflows
- approvals
- systems
- controls
Through this experience, I arrived at a simple conclusion:
AI should not become the backbone of an organization.
AI works best as the nervous system that connects information to that backbone.
It should connect external ambiguity to internal deterministic operations, and then help shape internal outputs into external context.
Figure 1. AI should not replace the deterministic backbone of the organization. It should act as the nervous system that connects external states, human interpretation, deterministic operations, and external communication.
The common mistake: putting AI at the center
Many discussions about AI in organizations focus on workforce redesign.
They ask questions such as:
- How many people can one AI-augmented worker replace?
- Will organizations become flatter?
- Will middle management shrink?
- Will junior roles disappear?
- Will senior employees become managers of AI agents?
These are important questions.
But I think they come too late.
Before redesigning the workforce, we need to answer a more fundamental question:
Where should AI be placed in the control structure of the organization?
If we put AI at the center of decision-making, we create a serious problem.
AI can generate useful outputs, but it is not a stable source of organizational responsibility. It may produce plausible outputs without fully carrying the reasons, constraints, risks, or accountability that the organization requires.
This is especially dangerous in brownfield development.
Brownfield systems are not clean greenfield environments. They contain:
- historical decisions
- implicit constraints
- operational risks
- legacy interfaces
- undocumented dependencies
- organizational habits
- approval paths
- release constraints
- failure history
If AI is placed at the center without a deterministic backbone, it may generate work that looks correct but does not fit the real organization.
That is why AI should not be the backbone.
The backbone must be deterministic
In my model, the organizational backbone is deterministic.
By deterministic, I do not mean that everything is simple or mechanical.
I mean that the organization must have stable structures that define how work is accepted, checked, approved, executed, and audited.
The backbone includes:
- rules
- workflows
- approval processes
- systems
- controls
- quality criteria
- evidence
- responsibility boundaries
This backbone is where quality is guaranteed.
AI can support quality-related activities, but it should not be the final source of quality.
Quality must be anchored in the organization’s deterministic structure.
This is especially important when AI is used for planning, design, testing, and release procedures. If AI generates these artifacts without being connected to the organizational backbone, the outputs may be fast but unreliable.
The organization may get more content, but not necessarily more control.
AI as the nervous system
AI becomes valuable when it acts as a nervous system.
The outside world is ambiguous.
Customers do not always express requirements clearly.
Markets change.
Regulations change.
Incidents occur.
Field information is incomplete.
Requests arrive with missing assumptions.
Stakeholders speak from their own context.
This information cannot be passed directly into deterministic operations.
Humans first receive and interpret it.
Then AI can help transform it into forms that the organization can process:
- requirements
- assumptions
- design options
- task plans
- implementation guides
- test perspectives
- release steps
- stakeholder explanations
In the opposite direction, deterministic operations also produce outputs that are not automatically understandable to the outside world.
A release plan, a design decision, or a system constraint may need to be translated into the context of:
- users
- managers
- regulators
- partner teams
- field operators
- executives
AI can help reshape internal outputs into external context.
But humans remain the interface to the outside.
Humans receive, interpret, explain, negotiate, and take responsibility for communication.
AI connects and transforms.
Humans remain the responsible interface.
QCD: Quality, Cost, and Delivery Speed
This model also explains how AI affects QCD.
Q: Quality
Quality should be guaranteed by the deterministic backbone.
That means quality comes from:
- rules
- workflows
- approvals
- systems
- controls
- review criteria
- test policies
- evidence management
AI can help generate test cases, detect risks, summarize differences, or prepare review materials.
But AI itself should not be the final guarantee of quality.
The organization’s deterministic structure must remain responsible for Q.
C: Cost
AI improves cost by reducing friction in the nervous system.
It reduces the cost of:
- collecting information
- summarizing context
- translating between technical and business language
- preparing documents
- identifying affected areas
- generating test perspectives
- creating release procedures
- adapting explanations to different audiences
The cost reduction does not come only from “writing code faster.”
It comes from reducing rework, duplication, coordination overhead, and information loss.
D: Delivery Speed
AI improves delivery speed by accelerating information flow.
When external information can be transformed into internal execution artifacts faster, the organization can move faster.
When internal decisions can be shaped for external communication faster, stakeholders can understand and act faster.
AI improves delivery speed because it shortens the distance between:
- request and requirement
- requirement and plan
- plan and implementation
- implementation and test
- test and release
- release and explanation
In short:
Q is guaranteed by the backbone.
C and D are improved by the nervous system.
Why senior engineers often benefit more from AI
This model also explains something I have observed in practice.
Senior engineers often use AI more effectively than junior engineers.
This is not because seniors know more prompts.
It is because seniors can provide more context.
A senior engineer can give AI information such as:
- why a feature exists
- which constraints are real
- what kind of failure is likely
- which design choice is risky
- where hidden dependencies may exist
- what the review will focus on
- what operations will care about
- what should not be changed
- what must be explained to stakeholders
The more useful context a human can provide, the greater the effect of AI on cost and delivery speed.
AI amplifies the context given to it.
If the context is shallow, the output remains shallow.
If the context is rich, AI can produce outputs much closer to real execution.
This is why senior engineers often get better results from AI.
But this should not remain an individual advantage.
The next step is to externalize senior context into organizational knowledge.
That means documenting:
- domain knowledge
- system constraints
- design rules
- review criteria
- release policies
- failure history
- escalation conditions
- quality gates
Once this context is externalized, junior members can also use AI more effectively.
In other words:
AI skill is not only prompt skill.
It is context transfer skill.
Responsibility matters
There is another reason why AI should not be placed as the backbone.
Responsibility.
If AI becomes the center of organizational decision-making, responsibility becomes blurry.
Who is responsible when AI makes a wrong design assumption?
Who is responsible when AI creates a release procedure that misses an operational constraint?
Who is responsible when AI-generated test cases fail to cover a critical risk?
The answer should not be “the AI.”
The organization must preserve responsibility through deterministic structures.
That means:
- humans remain responsible interfaces
- approvals remain explicit
- quality gates remain defined
- evidence remains recorded
- systems and controls remain authoritative
AI can support the flow of information, but responsibility must remain attached to human and organizational structures.
This is why I describe AI as the nervous system, not the backbone.
A nervous system carries signals.
It does not replace the skeleton.
What becomes lighter in the organization
This model does not start from the goal of reducing headcount.
However, it does make parts of the organization lighter.
What becomes lighter is the information relay layer.
Organizations often spend a large amount of effort on:
- translating external requests into internal tasks
- translating internal decisions into external explanations
- preparing repeated documents
- summarizing meetings
- converting technical details into stakeholder language
- collecting scattered context
- aligning different teams
- reformatting the same information for different audiences
AI can reduce this burden.
As the nervous system improves, fewer humans are needed only to relay, reformat, and restate information.
But this does not mean removing the backbone.
The organization becomes lighter because the nervous system becomes more capable, not because controls disappear.
This distinction is important.
A lightweight organization without a backbone is fragile.
A lightweight organization with a strong deterministic backbone and an AI nervous system can be both faster and safer.
Practical implications
If you want to apply this model, do not start by asking:
Which AI tool should we introduce?
Start by asking:
What is our deterministic backbone?
Then identify:
- what rules govern the work
- what workflows must be followed
- who approves what
- what systems are authoritative
- what controls must not be bypassed
- what evidence must be recorded
- where human responsibility must remain
After that, define where AI should act as the nervous system.
For example:
- intake of external requests
- extraction of assumptions
- brownfield impact analysis
- design draft generation
- test specification generation
- release procedure creation
- stakeholder communication
- post-work retrospective summaries
This makes AI useful without making it uncontrolled.
Final takeaway
My experience with AI in OSS development and brownfield organizational development led me to this model:
AI should not become the backbone of the organization.
The backbone must remain deterministic:
- rules
- workflows
- approvals
- systems
- controls
AI should become the nervous system:
- connecting information
- transforming context
- reducing rework
- accelerating delivery
- shaping communication
Quality is guaranteed by the backbone.
Cost and delivery speed are improved by the nervous system.
That is how AI can make organizations faster and cheaper without making them irresponsible.
Don’t make AI your backbone.
Make it your nervous system.