Agentic AI · Technology
Other AI reads your files. TGM understands your business.
Most AI reads your documents like a smart stranger skimming a filing cabinet: fast, but with no idea how the pieces fit. TGM builds a map of how your company actually works — customers, people, money, operations, and what drives what — and updates it automatically as new files and decisions arrive. Small, specialized AI models trained only on your company do the sorting; the big AI just writes the final answer.
The stack — and where everyone else stops
Six layers between a company and the AI it uses. Four of them are ours, end to end.
TheGreyMatter.ai owns layers 02–05 and calls any frontier model with a scoped slice — never the whole corpus. Models are bought, not built, and swap without a migration.
Where people work
Claude · ChatGPT · Microsoft 365 Copilot · Gemini · Amazon Quick · TGM apps
MCP connection
Governed, read-only MCP server — role-scoped, every read logged. The AI calls TGM.
Orchestration
20 agentic applications, all live — every one an agent — on a fleet of 32 production agents running continuously.
Business context
9 Vectors ontology · typed causal graph · permission-scoped memory · the firm’s playbooks.
Small language models
TGM-Nano (1B) and tuned SLMs classify, retrieve and route; each entity’s SLM from its own corpus.
Your company
Documents · systems of record · people, meetings and decisions.
- Assistants
- Claude, ChatGPT and Copilot reason brilliantly over what they are handed — and stop at the session.
- Search and graph
- Microsoft Graph and RAG index documents. They describe what is there; they do not model why.
- Deal platforms
- Hebbia, Rogo and data rooms organize per deal — and stop at close.
The ontology is the asset
Anyone can stand up a graph database. The value is the typology encoded in the schema.
| Tier | Vectors | Behavior |
|---|---|---|
| Assets — what it has | Market · People · Finance | Behave as stocks |
| Processes — what it does | Strategy · Operations · Execution | Behave as flows |
| Structures — what holds it | Expectations · Governance · Entity | Behave as constraints |
- Vectors
- 9
- Sub-Vectors
- 37
- Diagnostic themes
- 571
- Ontology nodes
- 742
- Typed relationships
- 2,250
- Nodes reached in 1–3 hops
- 608
Running in production on a Neo4j schema at v16. The framework-driven method is described in US 9,489,419 B2 — filed 2013, before enterprise LLMs existed. Vector similarity answers “what resembles this?” — useful, and it never compounds. Typed causal edges answer “why did the number move — and what do we do?”, ranked by the framework.
Why it is betterBring your own framework — a diligence model, scorecard or playbook maps onto the three tiers. A proprietary method becomes a reason to buy, not a reason to pass. The platform is the typology, not our framework; the 9 Vectors ontology is the reference implementation that ships in the box.
Small language models, formed per company
Tuned SLMs classify, retrieve and route; the frontier model only generates.
Classification accuracy · same 9 Vectors task
Ingest
Connected drives and systems are crawled continuously; every file lands on a Vector, Sub-Vector and theme with nobody opening an app.
Form
Each entity’s SLM is built from its own corpus alone. No customer data ever enters shared model weights.
Route
TGM-Nano plus three tuned task models (file detection, association, analysis) serve production traffic behind an LLMRouter TGM owns.
Generate
The frontier model receives a scoped, permission-filtered slice — never the whole corpus — and writes the answer.
- Accuracy
- 30 points above frontier models on domain classification — a 1B model beating models hundreds of times larger.
- Cost
- Own inference and fewer tokens per question. Spend follows companies, not seats, and stays predictable.
- Independence
- Engines swap without migration. A customer’s accumulated context survives a change of model supplier.
What already exists
Production evidence measured on 15 September 2026; delivery state stated per capability.
Causal business model
Typed edges support prescriptive traversal across functions, beyond document retrieval alone.
Agent layer
32 agents run production logic for diligence, finance, operations, boards, contracts, documents and evidence checks — across nine areas of the business.
Enterprise controls
Single sign-on and SCIM at the door. Workspace scope, approvals, audit and tenant isolation are enforced in the data path.
Delivery state Organizational memory · live Agent adapter · live Orchestration loop · live
Agents and applications
The agent catalog and the production application set, verified in the deployed container images.
Agents by area — diligence 4 · finance 5 · strategy and market 4 · people 2 · operations 2 · governance and legal 2 · boards and exit 2 · documents and data quality 6 · platform 5.
The 20 applications in production
- TheGreyMatter.ai
- 9Vectors.ai
- DueDiligence9.ai
- DDQ9
- ExitReady.ai
- Integration9.ai
- Operate9.ai
- Forecast9.ai
- Snapshot9.ai
- Board9.ai
- Contracts9.ai
- Pipeline9.ai
- NPS9.ai
- Culture9.ai
- OrgDesign9.ai
- Interview9.ai
- Measurement13.ai
- SupplyChain9.ai
- SWOTAnalysis9.ai
- BenchmarkedOutcomes.ai
The fleet works without being asked
A document that lands on a connected drive does not wait for someone to open an application.
| What happens on its own | What it takes, and what comes next | |
|---|---|---|
| A document lands. | SharePoint, OneDrive, the systems already in place | A connected source. Where an API or drive connection exists, this is automatic. |
| The fleet sees it. | Nobody uploads it, opens an app or writes a prompt | Otherwise, add the file. With no connector, the document is added once and the fleet takes it from there. |
| It is classified. | Into the right Vector, Sub-Vector and theme | Live today. Continuous ingestion, classification and graph growth. |
| The graph grows. | The model of the company is fuller than it was that morning | Also live. End-to-end process orchestration across systems. |
Why it mattersThe difference between a tool somebody must remember to open and a team that has already done the work. This creates automated ECM and business context and orchestration.
What the layer does for any frontier model
Tuned small models classify and retrieve; the frontier model only generates.
| Capability | What the LLM gets | How TheGreyMatter.ai does it | Status |
|---|---|---|---|
| Business context | The company’s structure, not just its text | Typed causal graph — 742 nodes, 2,250 edges, 9 Vectors | LIVE |
| Orchestration | A controller that runs the work, not a chat window | AgentOrchestrator, CorpusFirstHelper, LLMRouter; 32 agents | LIVE |
| Persistent memory | Context that outlives the session, the user and the model | Durable organizational memory across sessions and users | LIVE |
| Connected conversations | What one chat learns, the next one knows | Cross-user memory — the tenth employee’s work reaches the eleventh | LIVE |
| Fewer tokens | A scoped slice, not the whole corpus | Tuned SLMs classify and retrieve; the LLM only generates | LIVE |
| Security | Nothing it should not see ever reaches it | Permission compiled into the query; five isolation sites | LIVE |
| No vendor lock-in | No lock-in to any one frontier vendor | Four tuned SLMs behind a router TGM owns | LIVE |
| Bring your own cloud | Its own container and tenancy; the corpus belongs to the entity | Its own container, graph and memory, in its own cloud — and it travels with the company | LIVE |
| Predictable cost | Spend that does not scale per seat | Own inference and own models | LIVE |
Not instead of the AI you already have — above it. You keep your seats and add the layer; no rip-and-replace.
How TGM works with Claude and any model provider
Two directions across one boundary: we call the model, and the model can call us.
TGM calls the model
- Generation only. We call the provider’s API for the language step.
- Tuned SLMs do the rest. Classification, retrieval, association.
- Fewer tokens per question. A scoped slice, never the whole corpus.
- No vendor lock-in. The router is ours; engines swap without migration.
The model calls TGM — over MCP, live
- A governed, read-only surface reachable from inside Claude or ChatGPT.
- TGM is the server, not the client. We keep the orchestration controller.
- No new tool to learn. The user stays in the assistant they already use.
- Scoped at the query. Restricted nodes never enter model context.
Why a provider caresEvery deployment puts context-rich usage of their model into more enterprise workflows — and nothing is ripped out. The customer keeps their seats and adds the layer.
Security by construction, not by policy
Competing portfolio companies can share one deployment without sharing one fact.
Permission compiled into the query
- Deny by default. Invisible unless the role grants it — nothing is reachable by omission.
- Enforced inside the traversal. Restricted nodes never enter the result set, so they never enter model context.
- Whole-path integrity. A causal chain returns only if every node on it is visible.
- No negative disclosure. Two roles get two coherent answers; neither reveals the other exists.
- Audited per principal. Every path returned is logged to the egress ledger against the asker.
Isolation and deployment
- Five enforced isolation sites. Omitting tenant scope raises a TypeError. It fails loudly; it never leaks.
- A container per entity. Own graph, memory, corpus and SLM — it leaves with the company at exit.
- Walled deal workspaces. Nothing crosses deals or the firm’s SLM; dead deals are purged.
- Lift and shift. The whole platform, tuned models included, runs in the customer’s cloud.
- SSO and SCIM. Your identity provider, MFA and policies. We never store a password.
Five roles with access audit and break-glass. Own SLM per entity — no data in shared weights.
Why it is betterFiltering after retrieval is the common failure — the model has already reasoned over what it hides. TGM never lets it see. That is what opens regulated buyers.
Sign in the way you already do
Single sign-on and SCIM make TGM one more approved app behind the login your company already runs.
No new passwords
People sign in with the corporate credentials they already use. Their MFA and conditional access apply automatically. We never see or store a password.
Instant offboarding
Disable someone in the company directory and SCIM removes their TGM access and ends active sessions. No orphaned accounts with portfolio data.
Automatic onboarding
Add someone to the right directory group and they are provisioned into TGM with the right role. No tickets, invitations or seat-by-seat setup.
Audit-ready access
Sign-ins land in the customer’s own identity logs, so security teams verify access from their side instead of taking our word for it.
Why it mattersA 500-person firm doesn’t create 500 accounts or trust us with 500 passwords. IT approves one app it already knows how to govern.
Invented before the AI boom
Filed for patent in 2013 — and today every one of TGM’s 20 agentic applications runs on that method across the TGM-Nano platform.
Turn a company’s raw information into a structured picture of the business, organized by a framework and weighted by how much each source can be trusted. That method was filed for patent in 2013 and granted in 2016 — years before ChatGPT. The patent’s own drawing labels raw company data “Gray Matter”, which is where the name comes from.
| What the patent describes | What runs in TGM today | Status |
|---|---|---|
| A business framework with components and sub-components | 9 Vectors, 37 Sub-Vectors, 571 themes | LIVE |
| Data from many sources transformed into that framework | Continuous ingestion and classification | LIVE |
| Each source weighted by how much it counts (claim 12) | Weighting and scoring in the retrieval path; 143 tests | LIVE |
| Drill-down from the picture to the underlying source | Answers cite the source documents | LIVE |
| Questions tagged to framework components drive what data is pulled | All 20 agentic applications across the TGM-Nano platform | LIVE |
- Patent
- US 9,489,419 B2 · filed 18 September 2013 · granted 8 November 2016 · in force through 30 September 2033
- Ownership
- Sole named inventor: Edwin Miller. Owned by (I)Sage Management, LLC and licensed to TheGreyMatter.ai (9Vectors, Inc.) under a perpetual signed license agreement.
- Claim 12 alignment
- The claim’s data-weighing limitation maps to the weighting and scoring in the live retrieval path. Four suites — 143 test functions — cover the weighting behavior as shipped.
Engineering reading of claim scope, not a legal opinion.
What it is — and what it is notProof the idea is original and was examined, not a reaction to the AI boom. It protects the framework-driven method; what keeps competitors out over time is each customer’s compounding context. The firm is not renting a demo: the method underneath was examined, granted, and is held under license for the life of the patent.
We build it the way we ship it
A second fleet writes and reviews the platform itself — agentic engineering is our practice, not our pitch.
How the platform gets built
- Agents code and review continuously, with QA in the loop.
- Conformance is machine-checked against the spec in each deployed image.
- A small team, a large surface — 20 applications, 32 agents, one ontology.
- The same fleet pattern we ship to customers builds the product.
What that buys you
- Pace. The platform improves on the same cadence your graph does.
- Precision. Every claim carries a delivery state, because CI decides it.
- Proof. We operate agentic systems at scale daily, not in a lab.
- No demo gap. What is described here is what is running.
Why it mattersPlenty of companies sell AI. Fewer are built by it. The fleet that writes and reviews our code is the same pattern we put inside your business.
Why a general-purpose LLM cannot do this
Both of them read. Only one of them accumulates.
| A general-purpose LLM | TheGreyMatter.ai | |
|---|---|---|
| Structure | Reads flat text. No model of how the parts of a business cause each other. | Reads into a structure. Every input lands on a Vector, Sub-Vector and theme. |
| Memory | Starts over. Each session is isolated; nothing carries forward. | Grows with every input. Documents, spreadsheets, transcripts, chat, board packs — structured or not. |
| Connection | Is not connected. No durable link to the company’s files, systems or history. | Stays connected. Governed connectors keep the picture current as the company changes. |
| Classification | Guesses the category. 66.5% on 9 Vectors classification. | Classifies reliably. 96.5% on the same task. |
The gap is not intelligence; it is structure and memory.
The CTO’s scorecard
Where we lead, each row with a delivery state — and where we deliberately concede.
| Capability | General-purpose assistant | TheGreyMatter.ai | Status |
|---|---|---|---|
| Model of the company | Reads flat text | Typed causal graph — 742 nodes, 2,250 typed edges | LIVE |
| Reasoning about why | Descriptive retrieval | Multi-hop causal traversal in production | LIVE |
| Memory | Per user, per session | Organizational, role-scoped, enforced in the query | LIVE |
| Agents | General subagents, on request | 20 agentic applications, all live, on 32 production agents | LIVE |
| Models | Tied to one vendor’s stack | Four tuned models behind a router we own | LIVE |
| Isolation | Single-tenant boundary | Five enforced sites; container and SLM per entity | LIVE |
| Cost shape | Per seat plus usage at API rates | Own inference; a scoped slice, not the corpus | LIVE |
| Framework-driven method | General-purpose by design | Patented method (filed 2013, granted 2016) run by all 20 agentic applications; perpetual license to TGM | LIVE |
Where we concede, stated plainlyFrontier reasoning (we buy it), connector breadth, Compliance API and content-level eDiscovery, developer ecosystem, per-seat scale. Compete where speed compounds; concede where scale compounds.
Designed and built by Edwin Miller