# AI Data Unification for Small Business — One Knowledge Layer Every Agent Reasons Over > Source: https://ibl.ai/service/ai-data-unification/small-business Unify the tools you already run — accounting, CRM, e-commerce, scheduling — into one knowledge graph your AI agents reason over, no IT team required. You own all the code and data. For small businesses, the records that AI agents need are scattered across your accounting, CRM, e-commerce, scheduling, and email tools. AI Data Unification maps them into one living knowledge graph — an ontology agents reason over — exposed over the Model Context Protocol (MCP) without extracting a single record to the cloud. You own all the code and the data. The unified layer runs inside your network, behind your firewall and identity provider, and every agent — ibl.ai, Claude, Cursor, or your own apps — queries it scoped to the caller's role. ## What This Is ### AI Data Unification connects QuickBooks, HubSpot, Salesforce, Shopify, Stripe, Google Workspace, and Slack into a single ontology — a structured representation of your organization that agents can reason over. It has two layers: a semantic layer (the nouns — entities, attributes, relationships) and an operational layer (the verbs — actions, functions, permissions). Together they become the single source of truth every agent reads from. ### Unify once, reuse everywhere Model an entity once and the tenth agent costs a fraction of the first, because it inherits the same definitions, the same data, and the same governance. The implementation is open source at github.com/iblai/ontology — MIT-licensed, Python 3.11+, 163 passing tests. ibl.ai is family-owned and operated from New York, NY: a long-term partner, not a vendor that ships a license and disappears. ## The Two Layers ### The Semantic Layer — the nouns Entity types model the real-world objects in your organization — customers, orders, invoices, products, and appointments. Attributes capture their characteristics, and relationships define how they connect. A customer places an order; an order contains products; an invoice is an action with rules. This is the shared vocabulary your domain experts already use, made machine-readable so agents reason over the same reality your people do. ### The Operational Layer — the verbs Actions define the permissible changes an agent can make. Functions encode logic. Permissions govern who — human or agent — can do what. Agents inherit the same permissions as the users they serve, so they can only act within your existing rules. Every action is captured as data, so audit trails are built in, not bolted on. ## Five Core Principles ### Shared understanding at scale Teams operate from unified definitions instead of siloed, per-system data. Local decisions gain global context because every agent and every department reads from the same graph. ### Human-readable by design The ontology is expressed in familiar business terminology, not a technical schema. Domain experts can explore and validate it directly, bridging technical and non-technical teams. ### Build once, reuse everywhere A properly modeled entity serves every application that needs it. The cost of launching the tenth agent is a fraction of launching the first, because the knowledge is already there. ### Capture decisions as data Actions are recorded and become queryable knowledge. Audit trails and decision history are a property of the system, not an afterthought you instrument later. ### Ground AI in organizational reality Agents operate on the same knowledge humans use. Grounding answers in your real entities and relationships removes the translation layer where hallucinations creep in. ## How the Unified Layer Works ### Inbound connections Databases connect through the Google MCP Toolbox; REST-based systems connect through custom MCP servers, defined in tools.yaml. For small businesses, that means QuickBooks, HubSpot, Salesforce, Shopify, Stripe, Google Workspace, and Slack. Source credentials never leave your network. Each inbound connector has isolated credential scope, with secrets encrypted at rest. ### Knowledge materialization Data is synced into three layers: text memories in Markdown, a Postgres cache, and vector embeddings for semantic search. The unified graph stays fresh on a schedule you control. ### Outbound — one MCP server The unified layer exposes itself as a single MCP server over HTTPS, behind your firewall and identity provider. Any authorized runtime — ibl.ai, Claude, Cursor, or your own apps — connects to it, and results are scoped to the caller's identity. ## Security & Governance ### Read-only enforcement Version 1 is read-only by design. Before provisioning, a seven-test safety suite verifies that every write attempt is denied; if any write succeeds, provisioning halts and emits remediation SQL. Agents can read your systems — they cannot mutate them. ### Identity & access Every MCP request carries the user's Microsoft Entra ID JWT. The gateway validates it and resolves the caller's role from roles.yaml, so an agent sees exactly what that user is allowed to see — row-level scoping included, mapped to your SOC 2 obligations. ### Permissions inherited from your people Schema and data permissions are governed centrally. Agents inherit the same permissions as the users they serve, so access control is consistent whether a human or an agent is asking. ## Built-In Service Catalog ### Pre-configured connectors for Small Business The catalog ships defaults for QuickBooks, HubSpot, Salesforce, Shopify, Stripe, Google Workspace, Slack, and Zendesk. Seed any of them with one command (ontology service add --from ), then test and provision from the CLI. ### Provision from the CLI A single CLI covers the whole lifecycle — service discovery and testing, schema analysis, role validation, scheduled sync, and Docker Compose deployment. Stand the layer up, test it, and ship it from your terminal. ## Full Ownership ### Your code & data The entire unified knowledge layer is open source. Records stay in the systems you already run — the layer queries them in place; it never copies them to ibl.ai or any third party. "You will have your data, and we may not even have access to it." ### Your infrastructure The MCP server, the Postgres cache, and the vector index all run inside your perimeter. Deploy it in the cloud you choose, in your VPC, on-premise, or fully air-gapped. ### Built on Agentic OS AI Data Unification is the knowledge foundation the rest of the ibl.ai platform reasons over. Pair it with Agentic OS to deploy agents that act on the unified graph — all owned and self-hosted by you. --- *[View on ibl.ai](https://ibl.ai/service/ai-data-unification/small-business)*