01
Memory that accumulates.
Each conversation enriches the next. Your context, your tone, your clients are already there from the first word. Time invested yesterday saves you time today.
Enterprise AI architecture · my vision
My vision
of AI in the enterprise.
Learns. Remembers. Belongs to you. After 10 years designing and deploying monitoring and automation systems in B2B, I see AI exactly as a system — inputs, rules, actions, memory. Here's the architecture I propose to my clients to turn it into a lasting advantage.
Three properties
Well-designed,
AI accumulates.
Three properties turn AI from a point-in-time tool into lasting infrastructure: memory, coherence, ownership. That's exactly what I design for my clients.
02
Team coherence.
The whole team speaks the same language. Common vocabulary, shared standards, control over what goes out. Coherence by design.
03
Ownership of your intelligence.
Your conversations live in your files, with you. You switch tools whenever you want — everything stays. Your company intelligence belongs to you.
AI in the enterprise is less a question of tool than of structure. Memory that accumulates. Rules that stick. Investment that compounds.
Louis-Charles Cuierrier
The ecosystem
Three spaces,
one system.
Here's how I structure the architecture for my clients. Each space has a precise role: intelligence is created in the first, built in the second, and operates in the third.
Claude Chat
Think
Describe your company, plan your agents, design your workflows. An AI advisor guides you and produces structured files.
→
Claude Code
Build
Drop in the files, say "execute". The architecture builds itself. Agents are born. Skills install.
→
Cowork
Operate
Your agents work daily. They ingest your documents, run your tasks, and grow more capable over time.
Everything is connected through Markdown files in Obsidian.
Readable by a human, executable by any AI.
The tools
An open
ecosystem.
Each component has a precise role. None is proprietary. All are replaceable.
Claude
The AI that powers the agents. Used in Chat (planning), Code (building), Cowork (daily operations).
Obsidian
Markdown editor that serves as navigation interface for the vaults. Visualizes connections.
Cowork
The desktop agent that operates daily. Connected to the vault, runs tasks, ingests documents.
Claude Code
Terminal tool for building and automating. Creates vaults, runs workflows. 1M tokens of context.
MCP & Connectors
Bridges to your existing tools. Gmail, Calendar, CRM, SharePoint. The agents read and act.
Git
Complete history of each vault. Every change is tracked. Rollback possible at any time.
The foundation
The company
brain.
At the center of the architecture, a single vault holds everything that defines your company. Your identity, products, values, procedures, tone of voice.
Every AI agent in your organization reads this vault. They inherit the context, standards and rules. You write it once, everyone benefits.
No agent can modify the company brain. Read-only. The source of truth stays under your control.
The test: if a new employee read this vault, would they understand the company in 30 minutes?
company-vault/
├── identity/
│ ├── mission-vision.md
│ ├── values.md
│ └── brand-voice.md
├── products/
├── positioning/
├── procedures/
│ └── responsible-ai/
├── people/
└── glossary.md
The specialists
One agent per role.
Each an expert.
Each agent has its own vault, its own memory, its own skills. It learns from your documents, your emails, your procedures.
Two types of memory
Declarative memory
The wiki — what the agent knows
Entities, concepts, source summaries, cross-analyses. Automatically built by ingesting your documents. Cross-referenced, interlinked, kept up to date.
Procedural memory
Workflows — what the agent knows how to do
Step-by-step procedures with a clear split between what the AI does, what it prepares for validation, and what the human keeps. Each workflow is also an executable skill.
agent-sales/
├── CLAUDE.md
├── INSTRUCTION.md
├── config.md
├── company-context.md ← inherited
├── index.md
├── log.md
├── gaps.md ← learning
├── raw/ ← your documents
├── wiki/
│ ├── entities/
│ ├── concepts/
│ ├── sources/
│ └── synthesis/
└── procedures/
├── workflows/ ← skills
└── templates/
Learning
An agent who doesn't know…
learns.
The golden rule: an agent never does what it doesn't know. It fabricates nothing. It asks, documents the gap, and waits to be taught. Then it knows forever.
01
The agent blocks
Unknown task or missing context. The agent fabricates nothing. It stops.
02
It documents the gap
Precise entry in gaps.md.
What's missing, why, what would unblock it.
03
You teach it
A procedure, a document, an example. 5 minutes. The knowledge is in the wiki forever.
The gaps.md file is the best indicator of an agent's maturity.
Gaps that close mean an agent that's growing.
The skills
One workflow,
one skill.
Each task your agent knows how to do is documented in a file. That file is both a procedure (readable by a human) and an executable skill (readable by the AI). Same file, two uses.
i.
AI does it
The agent runs this step alone. Reading data, searching the wiki, calculations, formatting. Risk-free tasks.
ii.
AI prepares, you validate
The agent does the work, then stops and presents it to you. You verify before continuing. The default mode for most steps.
iii.
Human does it
This step needs your judgment, your client relationship, or your signature. The agent skips and tells you what to do.
Workflow example
Process a purchase order received by email
Trigger: an email identified as a purchase order lands in the inbox
| # | Step | Who does it? | Detail |
|---|---|---|---|
| 1 | Read the email and extract information | AI | Client, products, quantities, requested date |
| 2 | Check client history in the wiki | AI | Existing client? Special conditions? |
| 3 | Validate prices against approved list | Prepares | The agent calculates, the human confirms |
| 4 | Generate the PO in the template | AI | Official template filled automatically |
| 5 | Verify payment conditions | Prepares | Standard conditions applied, exceptions flagged |
| 6 | Approve and sign the PO | Human | The owner approves the final document |
| 7 | Send the confirmation to the client | Human | The agent prepares the email, the human sends |
| 8 | Record in the CRM | Prepares | The agent prepares the entry, the human confirms |
Governance
Responsible
by design.
Responsible AI is defined up front, before the first agent. Six questionnaires capture your principles ; every agent inherits and respects them automatically.
Principle 1
Reliability
The agent cites its sources. When it doesn't know, it says so. Zero fabrication.
Principle 2
Privacy
Sensitive data identified and protected. Clear rules on what AI can see.
Principle 3
Fairness
Equitable, auditable decisions. Bias identified and corrected at the source.
Principle 4
Transparency
The user knows when AI is involved. Sources traceable. Nothing hidden.
Principle 5
Accountability
Each agent has a human owner. AI doesn't make final decisions.
Principle 6
Inclusivity
AI serves everyone. All languages, all technical levels.
These principles live in the company brain. Every agent inherits them automatically. Your leadership keeps full control.
Your company intelligence lives in files you own. Not in a platform that bills you per seat. In folders on your server. Portable to any AI tool.
The difference
Why not just
ChatGPT?
A chat AI is a conversation. A vault architecture is a system. The difference isn't in the tool — it's in the structure.
| Classic chat AI | Vault architecture | |
|---|---|---|
| Memory | Limited and unstable across sessions | Persistent, structured, in files you own |
| Team coherence | Every person on their own | Single source of truth for all agents |
| Learning | Starts from zero | Grows with each document ingested and question asked |
| Portability | Locked to one vendor | Portable to any LLM, today or tomorrow |
| Governance | None, or after the fact | Built in by design with 6 responsible AI principles |
| Investment | Cost grows with usage | Value grows with usage |
My method
How I get you
into action.
A structured path I've validated with my clients. Short, targeted, in phases. Concrete results from the first week.
Phase 1
3 half-days
Base training
Your team understands Claude, Cowork, Obsidian, Claude Code. Foundations laid. You see the complete vision.
Phase 2
1 day
Discovery and mapping
We analyze your company together. We identify the 2-3 roles that would benefit most from a dedicated agent.
Phase 3
2-3 weeks
Building the architecture
The company brain builds. Agents are born. Workflows install. Responsible AI principles in place.
Phase 4
Ongoing
Autonomy and evolution
You run the system day-to-day. I stay available for major evolutions. The system grows with you.
Next step
If this approach
resonates, let's chat.
The first call is about understanding your context and seeing whether this architecture fits. No preachy pitch — just an honest conversation about what would make sense.