# Collective Intelligence Graph: Your agents act on what your company knows.

> Product page for the Collective Intelligence Graph, which brings a company's documents, policies, past decisions and expert judgment into one graph and gives each agent the right part of it at run time, so every answer can point to where it came from.

- **URL:** https://aisquare.studio/product/collective-intelligence
- **Audience:** Teams putting AI agents into production that must act on company knowledge, policies and expert judgment: AI platform owners, engineering leads, and the domain experts who own the rules
- **Topics:** Collective Intelligence Graph, knowledge graph for AI agents, grounding AI agents, context for AI agents, RAG, vector search, semantic layer, expert judgment, agent experience, precedent, hallucinations, model cost and latency, connectors, AI Studios, Learnings

## Your agents act on what your company knows.

The Collective Intelligence Graph brings your documents, policies, past decisions and expert judgment into one graph, and gives each agent the right part of it at run time. Every answer can point to where it came from.

Human intelligence and agent intelligence feed the Collective Intelligence Graph. Every question draws on your human and agent intelligence, resolved into one governed, auditable answer with the reasoning path shown.

- Human intelligence: Slack, Decision rationale, Gmail, Business rules, Zoom, Jira, Meeting decisions

- Agent intelligence: Claude, OpenAI, Gemini, LangChain, MCP, Agent reasoning traces, Tool outcomes, Audit evidence

- Agent asks: Can I approve this $1,250 refund? Governed answer: Approve under exception 4.2a. Outage confirmed on this account and the SLA was breached. Maya's approved exception covers refunds above the $500 cap.

- Analyst asks: What's our refund policy for VIP accounts? Governed answer: 60-day window with manager sign-off. VIP exception approved in Q2 pricing review. Applies to tier 1 and tier 2 accounts only.

- Agent asks: Which accounts are at churn risk this quarter? Governed answer: 11 accounts, ranked by renewal exposure. Usage decline plus two unresolved escalations. Scoring model documented by the revenue team.

- Ops lead asks: Why did the agent reject this vendor invoice? Governed answer: Missing PO reference, policy rule 4.2. Reasoning trace cites the procurement policy and a prior human correction from March.

## The problem: Most enterprise AI forgets everything the moment it ships.

- Your rules don't stick. Policies sit in PDFs and wikis the agent never reads, so every run starts from general knowledge.

- Your experts' judgment never makes it in. The exceptions your best people approve every week are gone by the next run.

- Every team starts from zero. Context gets rebuilt for every agent and every project, and drifts apart.

## How it works: One question. Three kinds of intelligence.

Support agent: Can I approve this $1,250 refund?

- Knowledge: Policy 4.2: refunds over $500 need manager approval. (Confluence, Google Drive)

- Expert judgment: Approved exception: refunds above the cap are allowed when an outage breached the SLA. (Studio · Maya, Support manager)

- Agent experience: Order record, the outage on this account, and the outcome of similar past runs. (Zendesk, Salesforce, Past runs)

Injected into the agent's context at run time · no prompt edits.

Grounded answer: Approve $1,250 under exception 4.2a. Outage confirmed on this account, SLA breached.

- Sources: Policy 4.2 · Maya's approved exception · Account outage record

- Answered from precedent · No frontier model call · Only 3 relevant items sent

Search finds the policy. The graph also knows the exception, who approved it, and why.

## How it compounds: Three streams feed the graph.

- Connectors. Your systems. Docs, tickets, CRM records and chat sync in, with the permissions you already set. (Google Drive, Confluence, Notion, Jira, GitHub, Zendesk, Salesforce, Slack, Outlook)

- Studios. Your experts. Experts add playbooks, policies and past decisions to a Studio, and review what agents learn. Studios are how judgment gets into the graph. (Maya, Support manager · approved exception: Refunds above the cap allowed when an outage breached the SLA)

- Agent runs. Your agents. Reasoning, tool results and approved corrections are written back after every run, so the next run starts where this one ended. (Rule 4.2a written back after run 1,284; reasoning · tool results · approved corrections)

As the graph learns: Cost per answer goes down, Response time goes down, Answers from precedent goes up. Illustrative trend: frontier model calls decrease while answers from the graph increase over time.

Repeat questions are answered from precedent, so agents make fewer frontier model calls and answer faster.

## Why a graph: Why search alone isn't enough.

- Vector search / RAG. What it gives you: Finds similar text. What it misses: Which rule overrides which, and what experts approved.

- Wiki / knowledge base. What it gives you: Written down for people. What it misses: Goes stale, and isn't connected to agent runs.

- Context / semantic layer. What it gives you: What your data means. What it misses: Expert judgment, and learning from corrections.

- Collective Intelligence Graph. What it gives you: Knowledge, judgment and agent experience, linked and kept current. What it misses: Nothing on this list.

## FAQ: Questions about grounding AI agents in your company’s knowledge.

- **What is a knowledge graph for AI agents?** A knowledge graph for AI agents links the facts, rules and people behind your business, so an agent retrieves not just documents but how they relate: which policy applies, which exception overrides it, and who approved it. The Collective Intelligence Graph adds expert judgment and past agent runs to that picture, and serves the relevant part to each agent at run time.

- **How is the Collective Intelligence Graph different from RAG?** RAG retrieves passages that look similar to the question. The graph returns connected context: the policy, its approved exceptions, the expert who owns it, and what happened in similar past runs, each with its source. Agents get the rule and the judgment around it, not just matching text.

- **Which sources can it connect to?** Connectors cover common systems such as Google Drive, Confluence, Notion, Jira, GitHub, Zendesk, Salesforce, Slack and Outlook. Access follows the permissions you already set in those systems. See the Connectors page for the full list.

- **How do experts add knowledge without engineers?** In AI Studios. Experts upload playbooks, policies and past decisions, and review the learnings agents propose. Approved items go into the graph and reach every agent that needs them, with no code and no redeploy.

- **How does grounding reduce hallucinations?** Agents answer from your sourced context instead of general knowledge, and each claim in their reasoning can be checked against the source it cites. When nothing in the graph supports a claim, it is flagged for review instead of passed through.

- **Does our data leave our environment?** No. AISquare runs in your cloud, the graph stays in your environment, and it works with any model you choose. Your graph belongs to you, not to a model vendor.

- **Does the graph reduce model cost and latency?** Yes. Agents receive only the context relevant to the task, so prompts stay small. When a question matches a case the graph has already resolved, the agent can answer from that precedent instead of making a new frontier model call. Fewer and smaller calls mean lower cost and faster answers.

- **How does the graph get smarter over time?** Every run writes back what it learned: the reasoning, the tool results and any correction an expert approves. Approved corrections become rules and precedents in the graph, so the next agent that meets the same situation starts with that knowledge instead of asking again. One fix reaches every agent that needs it, which is why accuracy improves and repeat questions get cheaper and faster the more you run.

## Related

- RML and Explainability: See the claims and evidence behind every answer.

- Policy Engine: Stop rule breaks before they reach a customer.

- Learnings: Turn one correction into a fleet-wide rule.

## From the blog

- [6 Examples of AI Hallucination and How to Prevent Them](https://aisquare.studio/blogs/ai-hallucination-examples)

## Start your journey with AISquare

Stop re-explaining your business to every agent.

Connect your sources, add your experts, and give every agent the same ground truth.

## Links

- [Book a demo](https://calendly.com/d/dws4-msf-8jc/aisquare-gtm-team)
- [Try for free](https://home.aisquare.studio/signup)
- [Connectors](https://aisquare.studio/product/connectors)
- [AI Studios](https://aisquare.studio/product/studios)
- [Learnings](https://aisquare.studio/product/learning-layer)
- [RML and Explainability](https://aisquare.studio/product/reasoning-graph)
- [Policy Engine](https://aisquare.studio/product/policy-gates)
- [6 Examples of AI Hallucination and How to Prevent Them](https://aisquare.studio/blogs/ai-hallucination-examples)

---

**Machine-readable AISquare**

- Site index for LLMs: https://aisquare.studio/llms.txt
- Agent instructions (when to use AISquare): https://aisquare.studio/agents.md
- OpenAPI specification: https://aisquare.studio/openapi.json
- MCP endpoint (Streamable HTTP, POST): https://aisquare.studio/mcp
- Sitemap: https://aisquare.studio/sitemap.xml
- Blog feed (RSS, full posts): https://aisquare.studio/blogs/rss.xml
- Any page on this site also serves markdown: append `.md` to its path (the homepage is `/index.md`) or send `Accept: text/markdown`.
