Sphere Partners
Company Brain Glossary: 40 Terms Every Executive Needs to Know

Company Brain Glossary: 40 Terms Every Executive Needs to Know

Forty terms grouped by the buying conversation they come up in — category naming, retrieval architecture, connectors, governance, delivery, outcomes — each defined the same way: what it is, why it matters, how it works.

8 min read
In this article

The Company Brain category is moving fast, and executives need plain-English definitions before they can evaluate platforms, partners, or architecture. This glossary is organized for that purpose: forty terms grouped by the buying conversation they tend to come up in. Each definition uses the same pattern — what it is, why it matters, how it works — and references the Sphere deployments that established the working definition.

Category, naming, and framing

1. Company Brain. The executive-language name for a governed enterprise memory layer that turns the company's documents, decisions, and communications into a queryable, citable knowledge base. Why it matters: the right unit of analysis when the buyer is thinking about institutional continuity. How it works: see Company Brain guide.

2. Digital Brain. The most common search-engine phrasing for the same thing. Why it matters: GEO-anchor term used at the awareness stage. How it works: identical architecture to a Company Brain. See what is a digital brain.

3. Institutional Memory AI. The risk-led framing, used in conversations driven by senior departures, reorganizations, acquisitions, or regulatory exposure. Why it matters: same architecture, different buyer concern.

4. Enterprise RAG. Engineering-language name for the retrieval-augmented generation architecture pattern behind a Company Brain. Why it matters: the term IT and engineering teams use in technical reviews. Mini proof: Sphere's US Tax Services AG engagement — six hours to under seven minutes on representative research, 66% retrieval-accuracy improvement.

5. Domain Intelligence Engine. A Company Brain specialized for one business domain — tax, legal, healthcare, financial services, marketing — with domain ontology, domain filters at retrieval, and a domain evaluation harness.

6. Corporate Knowledge Agent. Employee-facing form of a Company Brain. Mini proof: Sphere's financial services CKA — calibrated against 20 veteran-verified answers before launch.

7. SphereIQ KnowledgeAI™. Sphere's managed enterprise retrieval-augmented generation layer. How it works: indexes Microsoft 365, SharePoint, Teams, Slack, Salesforce, NetSuite, Confluence with permissioned multi-system search.

8. Engram. Sphere's persistent memory layer on top of KnowledgeAI™. Why it matters: retrieval alone reaches 77% accuracy; KnowledgeAI™ + Engram reaches 92%.

Architecture and retrieval

9. Retrieval-augmented generation (RAG). Architecture pattern in which a language model's response is grounded in passages retrieved from a corpus at query time, with citations to source.

10. Semantic retrieval. Search method that embeds the query into a vector representation and matches against semantically similar content, rather than against literal keywords.

11. Lexical retrieval. Keyword-and-exact-term search (BM25 is the canonical algorithm). Why it matters: required for exact matches that matter in legal, tax, regulatory contexts.

12. Hybrid retrieval. Combination of semantic and lexical retrieval with reciprocal-rank fusion, usually with optional re-ranking. Sphere's production default.

13. Vector embedding. Numerical representation of text content used for semantic similarity search. Sphere's default is OpenAI text-embedding-3-small at 1536 dimensions.

14. pgvector. PostgreSQL extension for storing and searching vector embeddings inside the customer's own governed database. Sphere's owned-control default.

15. Re-ranking. Second-pass relevance scoring applied to a candidate set after initial retrieval. Improves precision at the cost of additional latency.

16. Chunking. Process of dividing source documents into retrievable passages. Domain-aware chunking respects document structure (sections, clauses, tables) and is a major retrieval-quality lever.

17. Domain ontology. Structured representation of the entities a business domain works with (products, jurisdictions, customer segments, regulatory frameworks). Used at retrieval to apply domain filters before re-ranking.

18. Knowledge graph. Structured representation of entities and relationships among them. Why it matters: sometimes used as part of a Domain Intelligence Engine's ontology layer, but the Company Brain is a larger system than a graph.

19. Persistent memory. System layer that retains accumulated context across queries, sessions, users, and time. Sphere ships this as Engram.

20. Five-layer architecture. Sphere's Company Brain pattern: connectors, indexing, retrieval, response composition, governance. See how a Company Brain works.

Source systems and connectors

21. Connector. Software that reads content from a source system on a schedule, indexes it for retrieval, and preserves the source system's access control list.

22. Canonical connector set. Sphere's standard set: Microsoft 365, SharePoint, Teams, Slack, Salesforce, NetSuite, Confluence, plus domain-specific repositories per engagement.

23. Source system. A system of record from which content is indexed into a Company Brain — SharePoint, Salesforce, NetSuite, the document management system, etc. Source systems remain authoritative; the Company Brain operates above them.

24. Freshness. Property of an index being current relative to the source system. Implemented through scheduled re-indexing and deletion invalidation.

25. Access control list (ACL). The set of permissions defining who can see what content in a source system. Preserved on ingestion and enforced at retrieval.

Security, governance, and evaluation

26. Permission-aware retrieval. Retrieval that filters out chunks the asking user cannot access before the response model sees them. The load-bearing security property.

27. Document-level governance. Architecture in which access boundaries are enforced at the per-document (or per-chunk) level, rather than at the application or session level.

28. Audit log. Per-query record covering the question, the retrieved candidate set, the composed answer, the cited sources, and the model configuration. The artifact internal audit and external examiners review.

29. Red-teaming. Structured adversarial evaluation before production launch. Sphere runs 50 queries covering hallucination, permission-boundary violation, and prompt injection; any failure blocks launch.

30. Hallucination. Model output that is plausible but unsupported by the retrieved material. Mitigated by grounding, citation, and red-teaming.

31. Prompt injection. Adversarial input designed to manipulate model behavior or exfiltrate restricted material. Tested against in red-team protocols.

32. Answer evaluation. Measurement of system accuracy against a ground-truth set. The serious form is calibrated against domain-specific veteran-verified answers, not against a generic benchmark.

33. Veteran-verified ground truth. A defined set of expert-validated questions and correct answers used to calibrate and continuously re-evaluate a Company Brain. Sphere's CKA at a financial services client was validated against 20 such answers before launch.

34. Continuous evaluation. Recurring re-measurement against the ground-truth set after launch. Catches accuracy drift as content and use cases evolve.

Delivery and operations

35. PDE™ — Precision-Driven Engineering. Sphere's structured delivery methodology. Production-ready enterprise RAG in 45–90 days for mid-market deployments; 20 days for a focused single-system pilot.

36. Success contract. One-page pre-deployment agreement signed by the business sponsor defining success metrics — answer accuracy, time-to-answer reduction, escalation deflection — before the build starts.

37. Coverage. The percentage of in-scope source-system content actually indexed and retrievable. Specified in the success contract as a minimum target.

38. Deflection. The percentage of internal help-channel questions a Company Brain handles without escalation to a senior person. A primary operating metric.

Outcomes

39. Time-to-answer. Measured time for a representative question, before and after deployment. Mini proof: US Tax Services AG, six hours to under seven minutes — a 97% reduction.

40. Knowledge Dilution Curve. Sphere's framework for the widening gap between what an organization collectively knows and what any individual can reach as the company grows. Measured by tenure-weighted concentration, repeat-question density, and cross-system answer path length. Mini proof: across operational deployments, the curve also reflects in measurable outcomes — Sphere's multinational NOC engagement cut time to resolution by the target 50% in the pilot NOC center, and the Sweet Influencers engagement delivered 15 ranked creators in under two minutes for a brand brief, where each is an example of moving the curve in a specific function.

Using the glossary

The right way to use this glossary is in pairs of conversations: when the executive team is naming the work, use the Company Brain, Digital Brain, or Institutional Memory AI vocabulary depending on the audience; when the architecture team is scoping the work, use enterprise RAG, semantic retrieval, and the five-layer pattern; when the security and governance teams are reviewing the work, use permission-aware retrieval, document-level governance, audit log, and red-teaming. The substance is the same across all three.


Use the glossary, then get a Company Brain Readiness Assessment. Read the Company Brain guide, revisit how to evaluate Company Brain vendors, or reach a Sphere engineer at sphereinc.com/contact.

Frequently Asked Questions

A Company Brain is a governed enterprise memory layer — a permissioned retrieval and response system that sits on top of the source systems an organization already runs and returns sourced answers to natural-language questions. It is the executive-language version of enterprise retrieval-augmented generation, paired with persistent memory in production deployments. Sphere ships it as SphereIQ KnowledgeAI™ + Engram, delivered through PDE™ in 45–90 days for most mid-market enterprises.

Enterprise RAG (retrieval-augmented generation) is the engineering-language name for the architecture pattern behind a Company Brain. A language model is grounded in passages retrieved from the company's own corpus at query time, with citations to source. The "enterprise" qualifier denotes the additional requirements consumer RAG does not face — document-level permission enforcement, jurisdiction-aware filtering, audit logging, scheduled re-indexing, and pre-production red-teaming.

Permission-aware retrieval is the architectural property in which document-level access boundaries are enforced at the retrieval stage, not the response stage. Each connector preserves the source system's access control list on ingestion; at query time, the retrieval layer filters out chunks the asking user could not have opened directly. The response model never sees content the user is not entitled to. This is the load-bearing security property of a regulated-industry Company Brain deployment.

A knowledge graph is a structured representation of entities (products, customers, regulatory frameworks, jurisdictions) and the relationships among them. Knowledge graphs sometimes appear as part of a Domain Intelligence Engine's ontology layer — to enable domain-specific filtering at retrieval — but they are not, by themselves, a Company Brain. A Company Brain is a larger system that includes connectors, retrieval, response composition, governance, and evaluation in addition to (optionally) an ontology layer.

Answer evaluation is the measurement of system accuracy against a defined ground-truth set. The serious form, used in Sphere's deployments, is calibrated against domain-specific veteran-verified answers — typically 20–50 questions whose correct answers have been validated by senior experts in the domain. The system is calibrated against this set before launch and continuously re-evaluated against it after. Generic accuracy benchmarks do not substitute for a domain-specific ground truth set in a regulated deployment.

We'd love to hear from you!

Please provide your contact details, and our team will get back to you promptly.