An AI answering for your company needs to know where the answer came from.

It is the layer that makes an AI agent answer for the company with authorised knowledge, stated limits and a trail. It orchestrates the model, queries the knowledge base, applies the agent’s rules, routes to a person where appropriate, and records every reply.

Not in the sense of a tree of canned questions. What exists is the layer behind the service: agent selection, context assembly, knowledge lookup, answer verification, human routing and recording. The chat is only the surface of that.

Synapse is the layer that sustains the answer: model, knowledge, rules, API and trail. Axiom is the business intelligence interface aimed at people working in the company. The boundary between the two is still being settled, so we prefer to present them as complementary.

Models from OpenAI, Anthropic and Google, chosen per agent. The key belongs to the organisation and is stored with strong encryption — model usage is billed to your provider account and is not included in the subscription. We do not promise to pick the "best model" on our own: routing follows the configured rules.

It is the technique of searching your authorised documents for the passages closest to the question and handing them to the model along with the instruction. That way the answer leans on your material, not only on what the model learned by itself. It reduces improvisation — it does not eliminate it.

The search returns the passage with the source identified, and what was consulted is recorded in the conversation along with the model used and the cost. How the source is shown to the end client is defined in the agent and channel configuration.

Yes. Documents, policies, manuals, product information and FAQs can be loaded into the knowledge base and updated without a new release of code. Loaded content is treated as data, not as instruction — and passages with patterns suggesting manipulation are flagged.

When information is missing, the topic is sensitive or it exceeds what the agent may handle, the conversation enters a human queue with the question, the urgency and the history. The person’s reply is applied back inside the same conversation, so the client does not have to start over.

Yes, a REST API at version 1, authenticated by a scoped key (messages, conversations and knowledge base, each with read and write separated). It accepts test and live keys, has header-based idempotency so messages are not duplicated, and a call limit per key.

Yes. The widget is embedded with a public key — the secret key never reaches the browser — and only works on the allowed origins. Product and company name, logo, colours, support contacts, legal links, language and widget copy are configurable. Availability per plan comes from the commercial table.

Only previously approved tools and within the configured scope, with protections against querying sensitive data and against internal addresses. The agent does not close contracts, move money, change systems on its own or act outside what was authorised.

Model keys and channel credentials are encrypted at rest, and communication with external systems is signed. Data is separated by organisation and the environment is dedicated to the contracted operation, with an append-only audit log and a defined retention period.