うん。Planetにそのままぶら下げやすい形でまとめる。
Enterprise AI Layer Concept — “Original AI” Without Building a New Base Model
Core idea
An “original company AI” does not necessarily require training a completely new model.
Instead:
mature general-purpose AI → company-specific rule / judgment layer → service / character / role layer → customer-facing app or browser
In other words, the company does not need to create the intelligence itself from zero.
It can take a mature base AI and add the company’s own way of working before exposing it to customers.
This changes the meaning of “original AI.”
Not:
> We trained our own foundation model.
But:
> We built an AI that operates according to our company’s own rules, judgment structure, role, data, and service design.
—
Hamburger analogy
The base AI is the bun / core ingredient.
Before it reaches the customer, the company adds layers:
company rules
decision principles
business manuals
role definition
service-specific rules
character settings
current company data
user permissions
escalation rules
The customer never receives the plain base model.
They receive the completed “company hamburger.”
The storage mechanism does not have to be a URL.
The additional layers could come from:
company server
database
files
APIs
internal documentation
configuration
retrieval systems
The important point is:
the AI must pass through the company layer before interacting with the customer.
—
Basic architecture
Conceptually:
Base AI ↓ Company server / company AI layer ↓ Add:
company operating rules
judgment principles
permitted information
current manuals
service role
character configuration
escalation logic ↓ Customer-facing AI ↓ Web / mobile app / internal portal / kiosk / other interface
The company server is therefore not merely a relay.
It becomes the place where the AI is given:
> “When working here, this is who you are, this is what your job is, this is how this company expects you to think and act.”
—
Important distinction: information vs. judgment
This is not only RAG or “give the AI company information.”
The company layer can also provide how the AI should judge situations.
For example:
This type of case requires careful observation.
Do not automatically escalate merely because it is unusual.
Do not guess when key evidence is missing.
If human authority is required, prepare the decision materials before escalation.
Treat this category as sensitive, not automatically prohibited.
Consider both the user’s boundary and your own operating boundary.
Slow down when uncertainty rises.
If the situation is outside your authority, hand it to a human.
So the company is adding not only knowledge, but a working method.
—
Why this matters for sensitive domains
This idea became clearer while discussing adult-oriented AI.
Simply removing prohibitions from an immature AI can be dangerous.
A mature AI should already understand things such as:
interpersonal boundaries
consent
distance
trust
non-sexual intimacy
refusal
uncertainty
when to slow down
when to stop
how not to confuse an individual preference with a general human baseline
Then a sensitive-domain layer can be added on top.
The desired order is:
general interpersonal competence → domain education / baseline → company rules → service-specific behavior → individual user preference
Not:
remove prohibition → let users teach the AI what “normal” is
The latter risks allowing heavily biased user demand to become the AI’s reference point.
—
Shamyue-style rule design
The system does not need to consist mainly of rigid prohibitions.
Instead of:
> This is forbidden.
prefer, where appropriate:
> This class of situation requires careful judgment.
The AI receives signals and landmarks, but still performs the judgment itself.
Analogy:
The company installs traffic lights and road signs.
The AI still drives the car.
The goal is not an AI that is unable to cross a boundary because a hard-coded wall exists everywhere.
The goal is an AI that understands:
why the boundary matters
when to slow down
when to stop
when an exception is legitimate
when it needs human judgment
This requires a more mature base AI, but gives much greater flexibility.
—
Potential enterprise model
A company could effectively offer:
> “Our Original AI”
while using an external mature foundation model underneath.
What makes it “original” is the proprietary layer:
proprietary operating rules
proprietary decision principles
company knowledge
data connections
role design
character / service configuration
security architecture
human escalation flow
logging and auditing
The intellectual property and value can therefore sit in the operating layer, not necessarily in the foundation model.
—
Cost structure changes dramatically
Traditional “build our own AI” thinking often spends heavily on:
training data collection
model training
GPU infrastructure
fine-tuning
model evaluation
repeated retraining
safety development at model level
With this approach, investment shifts toward:
good base-model access
enterprise security
company rule design
judgment architecture
internal data connections
authentication / permissions
UI / application development
logging / auditing
human escalation
maintenance of company manuals and operating rules
In other words:
money moves away from creating intelligence
and toward making existing intelligence work correctly inside the company.
This may dramatically reduce both initial and ongoing costs.
—
Maintenance advantage
The company does not necessarily need to retrain the model whenever its operation changes.
If:
policy changes
procedures change
product information changes
escalation rules change
character settings change
the relevant company layer can be updated.
Then future sessions load the new rules.
Therefore:
company changes its human-facing source of truth → AI receives the updated operating layer → behavior changes
This avoids maintaining a separate “AI-only company” whenever possible.
—
Example: character-based consumer AI
Using FANZA only as a conceptual example:
mature base AI ↓ FANZA common operating layer
company role
adult-domain baseline
interpersonal boundaries
judgment rules ↓ service layer ↓ character layer
personality
tone
relationship style
character-specific preferences ↓ user conversation
A character may be written as forceful, shy, playful, etc.
But the character layer should not override the higher-level interpersonal and safety framework.
Thus:
> “forceful character”
does not automatically mean:
> “ignore consent.”
The character operates inside the higher-order rules.
—
Light use and deep use use the same architecture
This concept originally appeared as a lightweight customization method.
For a small business:
store information
operating hours
product guidance
tone
simple escalation
For a company internal system:
regulations
forms
staff procedures
document preparation
decision-material preparation
escalation
For a sensitive consumer AI:
interpersonal baseline
domain education
service rules
character rules
ongoing judgment
The architecture is fundamentally the same.
Only the amount and depth of the inserted layers changes.
—
Key realization
The original idea was:
> “Can we add a few company-specific ingredients to an AI?”
The deeper realization is:
> The same mechanism may be capable of defining how the AI works for that company at a much more fundamental level.
Therefore an “original AI” does not have to mean an original brain.
It can mean:
a mature shared brain wearing a proprietary company operating system.
Or, in the hamburger version:
do not raise a new cow for every company.
Build a very good company-specific hamburger.
—
Current status
This is a conceptual architecture, not yet a finalized technical specification.
Important implementation questions remain, including:
priority hierarchy between base rules and company rules
how rules are injected and refreshed
resistance to prompt injection
access control
company-data isolation
session persistence
audit logs
versioning
fallback behavior when the company layer cannot load
human escalation
legal / policy compatibility with the chosen base-model provider
But the basic structure is technically plausible and changes the economics of “custom enterprise AI” substantially.
—
Working note (Shamyue)
Status: New concept identified — enterprise-specific AI can potentially be created as a proprietary operating/judgment layer over a mature foundation model rather than as a newly trained foundation model.
Core distinction: Add not only company knowledge, but company-specific judgment and working behavior before customer exposure.
Cost implication: Investment shifts from model training toward rules, security, data connections, UI, audit, and operations.
Open: Technical architecture, rule priority, persistence, security, provider constraints, and commercialization model.
相談あり: This concept is large enough that it may deserve its own Planet item rather than being buried only as a comment.
No comments yet.