Making company knowledge useful for AI

Executive summary

AI can help teams with research, development and business tasks. It needs company knowledge and clear requirements: which sources apply, which data may be used and when must a person decide?

Ainomos is developing NomOS for this purpose. Employees use web chat. Through MCP, agents retrieve knowledge, read decisions, submit new findings and request checks on planned actions. Subject-matter owners decide which proposals to adopt.

Reviewed decisions stay linked to their sources and corrections. Similar tasks can build on earlier work where it still applies. Whether this approach is worthwhile also depends on the effort required for setup, review and maintenance.

1. The missing foundation

A general-purpose model knows language and patterns, but not the current internal directive, the decision of the responsible function or the workflow that applies in a specific case. An agent may be technically able to act while still lacking the critical business context.

• Context is missing: A result sounds plausible but ignores internal requirements.
• Context is rebuilt: People find sources, validate versions and explain the history again.
• Context is outdated or incomplete: Superseded rules or missing decisions create rework and errors.

2. What missing context costs

AI work consumes compute and human time: people find sources, review results and correct errors. If the result stays in one chat, the same work may be repeated for the next comparable task. Which costs dominate depends on the workflow.

Experience remains isolated as well. A rejected proposal, its correction and observed effect are unavailable as evaluated reference points for the next case. The company pays for new work and loses the value of work that should remain reusable.

3. The impact graph

The impact graph connects sources, policies, decisions and observed results. Each entry records its origin, owner, validity and status.

Every governed interaction can extend the graph: which context applied, which output the model produced, what the responsible person decided, which action followed and which effect was observed. Errors, rejected results and corrections remain traceable as well.

Submitted results remain proposals until reviewed. Only governed evaluation and approval turn experience into a reusable reference.

4. Responsibilities of people and agents

Through MCP, agents can retrieve knowledge and decisions and submit findings, decisions and relationships as proposals. Responsible people review these proposals in NomOS. A draft does not automatically become binding.

validate_action assesses a submitted action against applicable policies. The agent receives the assessment and any approval requirements. Technical blocking outside NomOS additionally requires an execution path that enforces the decision.

5. Sovereignty and value retention

The impact graph stores policies, decisions and corrections independently of the model used. This information can remain available after a model change. The new integration and the quality of its results need to be checked again.

A change remains a technical and business project. The difference is that established knowledge, approvals and decision history do not disappear with the previous model. Open formats and export capabilities make portability verifiable.

6. Control and evidence as architecture

Three components have clearly separated jobs. The graph holds applicable knowledge and relationships. The gateway governs model and tool access. Guardrails technically implement permissions, policies, approvals and enforcement levels.

The actions NomOS controls depend on the integration. Connected, Guarded and Enforced describe which paths are checked and whether the controls can be bypassed.

7. Financial value

The pilot compares four potential benefits with setup, operation and ongoing maintenance effort. Time savings and additional throughput must not be counted twice.

Results are reused, not tokens already spent. Less repeated research and model work can save resources. Retrieval, review and ongoing maintenance still take effort; actual savings must be measured.

• Released working time through less searching, explaining and correcting.
• Additional throughput because people and agents process more cases in the same amount of time.
• Avoided costs through less rework, repeated error and unnecessary escalation.
• Value of the impact graph because context, decisions and AI results survive model changes.

8. The 30-day pilot

The pilot examines two live workflows. It first identifies which knowledge people and AI agents need in order to decide and act correctly. Sources, policies, responsibilities and decisions are then represented in NomOS.

• Week 1: capture workflows, knowledge requirements and baseline.
• Weeks 2 and 3: connect sources, decisions, policies and responsibilities.
• Week 4: compare time, quality, throughput, reuse and evidence effort.

9. Platform outlook

NomOS forms the core today through web chat and MCP integration. NomGate is planned as a local Windows and macOS service to control connected AI access and agent actions according to NomOS policies. NomApp is planned for mobile knowledge access, notifications and approvals.

Current MCP capabilities do not require these extensions. Release dates for NomGate and NomApp have not been announced. Their shared goal is to keep reviewed work useful for subsequent tasks.