How a Company Brain Works: The Technology Behind Organizational AI Memory

A Company Brain connects to the systems where institutional knowledge already lives, indexes that knowledge by meaning rather than by filename, and returns sourced answers with the company's permission boundaries intact. The architecture is not experimental. It is the same five-layer pattern Sphere has shipped into regulated professional services, financial services, and multinational operations clients, with measurable production results. This article walks through each layer in plain language, with the operating evidence attached.

How does a Company Brain work?

A Company Brain has five layers. They run as a single system at query time, but they are built and governed independently.

1. Connectors that read content from the source systems where institutional knowledge already lives, on a schedule, with the source system's permission model preserved.
2. Indexing that converts the content into a hybrid representation — vector embeddings for semantic search and lexical indexes for keyword precision — enriched with domain-aware metadata.
3. Retrieval that, given a user question, finds the most relevant chunks across the index using semantic similarity, lexical match, and (in regulated deployments) jurisdiction or role-based filters.
4. Response composition in which a large language model is constrained to the retrieved material and produces a cited answer in plain language, with links back to each source document.
5. Governance, freshness, and evaluation — the layer that keeps the system safe, current, and accountable: document-level permission enforcement at query time, scheduled re-indexing as source content changes, audit logging of every query, and continuous evaluation against a set of veteran-verified answers.

The architectural promise is simple: the user asks a question in plain language, and the system returns a sourced answer drawn from across the company's own corpus, without the user ever having had to know where the information was stored. The engineering work that makes that promise reliable is mostly in layers 2, 3, and 5.

What systems does a Company Brain connect to?

The canonical connector set in Sphere's SphereIQ KnowledgeAI™ deployments covers the source systems most enterprises actually run:

• Microsoft 365 — Outlook, OneDrive, Office documents
• SharePoint — document libraries with the existing permission model preserved
• Teams — channels, chat history, shared files
• Slack — public channels, private channels (with consent), DMs (when scoped by policy)
• Salesforce — accounts, opportunities, notes, attachments
• NetSuite — records, attachments, customer history
• Confluence — spaces, pages, attachments
• Enterprise-specific repositories — domain document stores, ticketing systems, internal wikis

Each connector runs on a schedule, respects the source system's access control list (ACL), and updates the retrieval index as the underlying content changes. The permission boundary is not flattened on ingestion; it is carried through to the retrieval layer so that a query never returns chunks the asking user could not have opened directly.

At US Tax Services AG, the connector set spanned SharePoint, Outlook, Teams, PDF archives, and personal advisor files — exactly the sprawl regulated professional services firms typically operate. The same five-layer architecture absorbed all of it, with role-based access and jurisdiction filters added at the retrieval layer for the firm's regulatory profile.

How does a Company Brain retrieve answers?

The retrieval layer is the part most teams underestimate. Three operating mechanics are worth naming.

Hybrid retrieval. Pure vector search is good at finding semantically similar passages but can miss exact-term matches that matter in legal, tax, or regulatory contexts. Pure keyword search is the opposite. Sphere's deployments combine both — vector embeddings (1536-dim, with OpenAI's text-embedding-3-small as the default) and lexical retrieval (BM25 with reciprocal-rank fusion) — and then apply optional re-ranking before the response model sees the result. The accuracy effect is measurable. At US Tax Services AG, retrieval accuracy on the firm's internal benchmark improved by 66%; research time on representative client questions dropped from six hours to seven minutes — a 97% reduction.

Domain-aware filtering. A Company Brain in a regulated industry needs more than relevance — it needs jurisdiction, role, and date filters that constrain retrieval before re-ranking. The wrong tax memo from the wrong jurisdiction is worse than no answer. Sphere's deployments expose these filters as part of the query, not as a post-hoc check.

Permission enforcement at retrieval, not response. Document-level access is checked at the retrieval stage, so chunks the user cannot see never reach the response model. This is the load-bearing security property of the system, and it is the reason document-level governance has to be a first-class concern in the architecture rather than a feature retrofitted later.

The response model — typically GPT-4o or Claude Sonnet, selected per deployment — is then constrained to the retrieved material. It composes a sourced answer in plain language, with citations linking to each underlying document. The model is the smallest part of the work. The retrieval pipeline in front of it is most of it.

For a deeper retrieval discussion see the Enterprise RAG pillar guide.

How does a Company Brain stay current and secure?

A retrieval system that ages out within a quarter is a liability. Three mechanisms keep a production Company Brain current and safe.

Freshness. Connectors re-index on a schedule the deployment defines — typically hourly for high-velocity sources (Slack, Teams) and daily for slower sources (SharePoint, NetSuite). When a document is deleted in the source system, the corresponding index entries are invalidated, not orphaned. The system's notion of "current" is anchored to the source systems, not to a snapshot at deployment.

Governance. Every query, every returned document, and every model response is logged for audit. Administrators see what the system is being asked, what it is returning, and where the returns come from. In regulated industries this layer is non-optional — it is what makes the system reviewable by internal audit and acceptable to external regulators.

Pre-production red-teaming. Before any Sphere deployment goes live, the system runs through a structured adversarial evaluation: 50 red-team queries designed to trigger hallucination, permission-boundary violations, and prompt injection. Any failure blocks production launch. This is non-negotiable in Sphere's PDE™ delivery framework.

Continuous evaluation against veteran-verified answers. Once live, the system's responses are compared against a defined set of expert-validated ground truth. At a financial services client, the Corporate Knowledge Agent was calibrated against twenty veteran-verified answers before launch and continuously re-evaluated against the same set after. The veterans defined the truth; the AI was held accountable to it.

The combined effect — fresh content, audited queries, red-teamed before go-live, evaluated against expert ground truth — is why a properly built Company Brain is not experimental. It is a production system with the same operating discipline as any other piece of enterprise infrastructure.

What does this look like in practice?

The clearest operating example is the multinational NOC engagement. Sphere built automated intake, classification, ticket routing, topology tracking, and knowledge-base integration on top of the client's existing ticketing and runbook systems. The same five-layer pattern absorbed it: connectors into tickets and runbooks, hybrid retrieval over the indexed content, jurisdiction-aware filters for which network and which region, and audited responses for the on-call team. Incident response time dropped 50% — because the runbook, the prior-incident context, and the system-specific notes were addressable from one query instead of three on-call escalations.

In each Sphere deployment the architecture is the same; the corpus, the connectors, and the domain filters change. The pattern is repeatable because it has been built five layers at a time, with the governance and evaluation work treated as first-class engineering rather than as a deployment afterthought.

The architecture is well-characterized

The hard part of a Company Brain is not the model. It is the governance, the evaluation discipline, and the implementation work that turns the architecture into a production system. Sphere ships this as SphereIQ KnowledgeAI™, paired with Engram for persistent memory, and delivered through PDE™ (Precision-Driven Engineering) — typically 45–90 days to production, with a 20-day path for single-system deployments.

See how Sphere builds a governed Company Brain. Read the Company Brain guide, revisit what a digital brain is, or reach a Sphere engineer at sphereinc.com/contact.