An effective Enterprise Brain rests on five fundamental strata: connected Data, Memory through semantic search and graphs, Reasoning under models and rules, Agency deployed through superagents, and Governance guaranteed by audit. The hierarchy is non-negotiable; skipping any of these levels condemns the project.

In short

• The five layers are Data, Memory, Reasoning, Agency and Governance. Each has its technology, its deliverable and its internal owner.
• Layer 2 (Memory) is the one that turns scattered information into queryable knowledge. It is where the project is won or lost.
• Layer 5 (Governance) does not go at the end of the calendar: it is designed alongside layer 1.
• Superagents live in layer 4 and consume layers 1 to 3. Without them, they execute blind.
• Each layer has its own acceptance metric; the metric for the whole is the RAO (Operational Autonomy Ratio).
• The most expensive mistake we see in projects is connecting every source before defining the semantic model of the business.

Why is the Enterprise Brain organised in layers?

Because every problem in an enterprise AI rollout belongs to a different layer and is solved with different tools. When a company says "the AI makes things up", the fault is almost never in the model (layer 3): it is that the AI cannot find the right document (layer 2), or that the document was never connected (layer 1).

Separating into layers gives you three things: you can diagnose where it fails, replace one piece without rebuilding the system, and assign an internal owner to each level. This architecture is the basis of the Enterprise Brain pillar and of phases 2 to 5 of the NEXT-5 Roadmap.

What exactly does each layer contain?

Layer	Name	What it solves	Typical technology	Internal owner	Acceptance metric
1	Data	Reaching the sources where knowledge lives	ERP/CRM connectors, APIs, email, document manager, CDC	IT Manager	% of critical sources connected and synchronised
2	Memory	Turning scattered information into retrievable knowledge	Vector store, hybrid search, knowledge graph, semantic layer	Data or operations lead	Retrieval precision against a bank of real questions
3	Reasoning	Interpreting, comparing and deciding over that knowledge	Proprietary and open models, routing by task, business rules	AI architect	% of correct answers and cost per query
4	Agency	Running complete processes with tools	Superagents, orchestrator, tools with permissions, token budget	Business process owner	RAO per process
5	Governance	Controlling who can do what, with what trail	Roles, agent registry, logs, human approval, retention	Board + legal counsel	% of actions traced and audited

How do you build layer 1 (Data) without drowning?

The temptation is to connect everything. It is an expensive mistake: it multiplies the cost of ingestion, puts noise into memory and delays the first visible result by months.

The NEXTAI criterion is the opposite: connect the sources that feed the three to five decisions that move the most money. In most SMEs that means the ERP, the document manager and commercial email. The CRM comes later; the rest, when a process asks for it.

Recommended order of work:

1. Source inventory: system, owner, volume, rate of change and presence of personal data.
2. Classification by decision value, not by data volume.
3. Definition of the semantic model: what a client, an order, a contract, a cost and a margin are in your company.
4. Connection of the first 3 sources in read-only mode.
5. Incremental synchronisation, not full dumps.
6. Validity marking: which document is in force and which has been superseded.

Point 3 is almost always skipped and it is the one that causes the most trouble later. If "active client" means one thing in the ERP and another in the CRM, the AI will answer the same question with two different figures and the project will lose credibility in the first demo.

What does layer 2 (Memory) do and why is it the critical one?

The Memory layer is the one that answers "what do we know about this?". It combines two complementary mechanisms:

Semantic search. Documents are chunked, turned into vectors and retrieved by meaning, not by exact word. This is where you decide the chunking (preserving titles, sections and hierarchy), the hybrid strategy (semantic plus keyword) and the reranking of results. The practical detail is in Enterprise RAG: an implementation guide.

Knowledge graph. It models the relationships of the business: a client belongs to a group, the group has a framework contract, the contract applies a discount to a product family, the product has a cost and a margin. That is what makes it possible to answer questions no single document contains. Expanded in the enterprise knowledge graph.

The four mistakes most often repeated in this layer:

1. Fixed-size chunking that destroys tables and clauses.
2. Ingesting old versions without marking validity.
3. Not inheriting permissions from the source system, so the AI leaks to people it should not.
4. Not building an evaluation bank of real questions before opening the system.

From our experience on NEXTAI projects, an evaluation bank of 80 to 150 questions with known answers is the single instrument that most reduces the risk of abandonment: without it, nobody knows whether the system got better or worse after a change.

How do you choose the model in layer 3 (Reasoning)?

There is no one model for everything. Layer 3 routes each task to the right model according to three variables: difficulty of the reasoning, sensitivity of the data and cost per query.

Type of task	Model profile	Dominant criterion
Classify, extract fields, label	Small or open model	Cost and latency
Draft an answer with sources	Mid-range general model	Writing quality
Compare contracts, decide exceptions	Advanced reasoning model	Precision
Particularly sensitive data	Open model on own infrastructure	Data sovereignty

On top of that layer sit the business rules, which should not be delegated to the model. If the maximum discount without approval is 12%, that is a deterministic rule, not a probabilistic interpretation. The boundary between what the code decides and what the model decides is documented explicitly.

What changes when layer 4 (Agency) is switched on?

The risk changes. Up to layer 3 the system reads and answers; from layer 4 onwards it writes into real systems: it creates quotes, updates the CRM, sends emails, posts entries.

A Superagent is an agent that owns a complete business process, with persistent memory, execution tools, explicit limits of authority and its own KPIs. Its deployment record includes, at minimum:

1. The process it owns and its boundaries.
2. Permitted tools and the systems it writes to.
3. Limits of authority: maximum amount, maximum discount, daily volume.
4. Cases that force human approval.
5. Monthly consumption budget.
6. KPIs, including the target RAO (Operational Autonomy Ratio).
7. A named human owner.
8. A withdrawal plan if it stops adding value.

The NEXTAI deployment rule: no Superagent gets write permission until its process has run in suggestion mode for a measured period. Usually between 2 and 4 weeks, with human review of 100% of the outputs.

Why can layer 5 (Governance) not go last?

Because permissions are inherited from layer 1. If ingestion did not preserve who could see each document, there is no way to apply access control afterwards without re-ingesting everything.

Governance of an Enterprise Brain covers four fronts: permissions by role applied to every answer, full traceability (which source supports each statement and who consulted it), human approval for sensitive actions, and an agent registry with its record and history. The full framework is in AI governance. On obligations under the EU AI Act and the GDPR, the design should be validated with your own legal counsel: this article informs, it does not advise.

How does this architecture look in a real company?

Realistic case: an insurance broker in southern Europe. 65 employees, €9M in commission income, 14,000 live policies, a sector-specific ERP and 20 years of documentation in shared folders.

Problem: answering a cover query meant locating the policy, the particular conditions, the endorsements and the previous correspondence. Between 20 and 40 minutes per query, with an unbearable peak in January.

Built layer by layer:

• Layer 1: the sector ERP, the document manager and the claims mailbox. Three sources, not eleven.
• Layer 2: chunking that respects the structure of the policy conditions, validity marking per endorsement, and a policyholder–policy–cover–claim graph.
• Layer 3: a small model to extract covers, an advanced model to interpret exclusions.
• Layer 4: a support Superagent that drafts the reply with the clauses cited.
• Layer 5: permissions by portfolio, traceability of every citation, and mandatory human approval before sending.

Result, from our experience on NEXTAI projects with equivalent profiles: response time from 20–40 minutes to 5–8 minutes, and a RAO of 55% on standard queries after the third month, with 100% of answers about exclusions reviewed by a person.

How NEXTAI does it

NEXTAI implements the five layers in the order of the NEXT-5 Roadmap: Mapping for layers 1 and 2, construction of the Enterprise Brain with its own evaluation bank, and only then activation of Superagents over processes with an identified owner.

Each layer is delivered documented and handover-ready, with its internal owner and its metric. The whole is measured with the RAO process by process, and the company's overall position with the IMAN (AI-Native Maturity Index). We do not deliver closed architectures: if you want to change supplier tomorrow, you can.

Frequently asked questions

Can the layers be built in parallel? Layers 1 and 2 can be, partly, and layer 5 has to be designed from the start because permissions are inherited from ingestion. What you cannot do is switch on layer 4 without having validated layer 2 with an evaluation bank. A Superagent on an unevaluated memory just executes mistakes faster.

Do you need a knowledge graph or is a vector store enough? A vector store alone answers documentary questions ("what does the contract say?"). Relational questions ("which clients in the group have an expired contract and an outstanding balance?") need a graph. Most companies need both: the graph supplies the relationships and the vector store the content.

How many sources should you connect at the start? Three, in most cases: the ERP, the document manager and the email of the area involved. Expanding comes later, driven by specific processes. Connecting eleven sources at once multiplies the cost and delays the first result without improving answer quality.

What internal profile does the company need to sustain this architecture? A data or operations lead who spends part of their day curating context, and a process owner for each Superagent. You do not need a data science team. Without those two roles, the architecture degrades within months.

The success of this architecture is assessed through source coverage, retrieval precision, cost per query and process execution rate. While RAO sets the pulse of the business, the other indicators dissect its operational behaviour. Find the missing links in your infrastructure with our free Digital Audit.