How Syllabium uses AI

Authority Before Action

Every AI request is a defined action on named content. It passes twelve checks in a fixed order, waits for a person where it must, and is recorded.

Syllabium is an enterprise learning management system (LMS) for regulated industries, made by Athantic AI LLC. This page describes how Syllabium governs the AI inside the platform: what each AI agent may do, when its work waits for a person, and how each action is recorded.

The Assistant’s Four Modes

Each author chooses how far the assistant may go. Suggest is the default.

Mode

Off

The assistant is switched off.

Mode

Ask

It answers, and writes nothing.

Mode · default

Suggest

It shows each change before and after, and the person applies it.

Mode

Act

It writes into the author’s own draft, recorded, with undo.

Authority, Capability by Capability

For each AI agent and each capability, an administrator sets one of four outcomes. Unset means refused.

Outcome

Allowed

An explicit grant for a named agent and capability.

Outcome

Human review

The action waits for a person. Safety-critical work by an AI agent does too.

Outcome

Admin review

A different administrator must decide.

Outcome

Blocked

The refusal and its check are recorded.

Governed AI

AI Works Here Under Authority

AI agents act through the same recorded route as people. Administrators set what each one may do.

  • Every AI action passes one route, with an emergency stop at its front.
  • An author sets the assistant to Off, Ask, Suggest or Act.
  • AI work on safety-critical content waits for a person.
  • Every model call is logged with its model, tokens and cost.

Syllabium uses Claude models, through Amazon Bedrock and Anthropic's API.

The Controls at a Glance

Each control Syllabium applies to AI, and what it does to an AI action.

ControlWhat it does
Authority by capabilityAn administrator grants each AI agent each capability, and anything not granted is refused
Twelve checksRun in a fixed order before any AI action; a refusal names the check that refused it
Human reviewThe action waits for a person
Admin reviewA different administrator decides
Safety-critical ruleAI work on content flagged safety-critical always waits for a person
LimitsEach AI agent has limits on actions and tokens an hour and on spend a day and a month, and a cap across all agents applies at each level
Emergency stopRead first by every AI action; it halts all AI work or a part of it
Action logEvery attempt and every model call, in tamper-evident records kept seven years
Identity ruleA person’s identity is never sent to a model

AI review before publish

The AI Agents in Syllabium

Each AI agent, what it does, and how a person stays in charge of it.

AI workWhat it doesHow a person stays in charge
The authoring assistantWorks beside authors in the studio, in Off, Ask, Suggest or Act, on text onlyThe author chooses the mode, and publishes or discards every draft
Reordering proposalsProposes a new order for courses, modules, lessons and slidesEach proposed change waits for a person by default
Engagement checksReads each learner’s overdue lessons, missed reflections and days unseen once a day, and raises a suggested intervention in fixed words, without calling a modelA staff member approves, rejects or resolves each intervention
Notes on reflectionsDrafts a note on a learner’s reflection for the supervisorIt runs only when the enrollment consent names it and a person turns it on, and the learner never sees the note
Question checksA verifier checks each assessment question against the slides it cites, and a separate critic checks the verifierA rejected safety-critical question goes to a reviewer
The learner’s assistantExplains a slide to a learner, and offers a debriefEvery request passes the same twelve checks and is recorded
Quality checksRun the platform’s built-in checks, and write a short summary of each runVerdicts are decided by rules, never by a model, and a failed run stays failed until a named person acknowledges it with a reason

AI drafting of questions, alt text and captions is planned, and so are further AI roles, among them a tutor, a grader, a translator and a proctor.

AI agents in trainingBlock-level LLMs

Questions About Governed AI

What governed AI means in Syllabium, and how it works.

What Does Governed AI Mean in Syllabium?

In Syllabium, governed AI means that every AI action inside the platform is checked against the authority an administrator has set, waits for a person where a rule requires it, and is recorded. It is how Syllabium governs its own AI, not a course about AI governance.

Organizations author and teach their own courses on Syllabium, and each organization owns its content.

Who Grants an AI Agent Its Authority?

In Syllabium, an administrator grants each AI agent its authority, capability by capability, and anything not granted is refused. A person activates or retires each AI role with a recorded reason, and an AI agent never approves.

A grant can be limited to one library or one block. Capability grants for people and for AI agents sit side by side, each with a reason. The assistant does not appear on the pages that govern it.

Can Organizations Set Their Own AI Rules?

Yes. AI settings in Syllabium exist at the platform, tenant and organization levels, and a lower level can only be stricter. Tenant owners and organization administrators set the rules for their own level.

How Are AI Agents Identified?

Each AI agent in Syllabium acts through its own account, with no password and no sign-in, and is never impersonated. No person can hold an AI role.

Where Does AI Wait for a Person?

AI work in Syllabium waits for a person wherever an administrator has set Human review or Admin review, and always on content flagged safety-critical. The authoring assistant does not publish content, so a person decides what is published.

Above ordinary publishing, an AI agent’s action always waits for a person.

AI review before publish

How Is Safety-Critical Content Protected?

In Syllabium, a person flags content as safety-critical, and no setting gives an AI agent a plain “allow” on it: its work there waits for a person. The rule is enforced in the code and in the database.

What Is Recorded About Each AI Action?

Syllabium records every attempted AI action with the check that decided it, and every model call with its model, tokens and cost. The record is written whether the action was allowed, refused or failed, and it is kept seven years in tamper-evident records.

The chained records hold no prompt, answer or person’s words. A refusal names the check that refused it and never reaches a model.

Evidence

How Is AI Spending Limited?

Each AI agent in Syllabium has limits on actions and tokens an hour and on spend a day and a month. Every call’s cost is estimated before it is made and held to a cap, and a cap across all agents applies at each level.

A warning email goes out as spend nears a limit, and at the limit the work is refused.

How Is AI Stopped in an Emergency?

Syllabium has an emergency stop that every AI action reads first. It can halt all AI work, one agent, one kind of work or one type of content, and while it is set every AI page shows a red banner.

Setting a stop takes two steps.

What Data Does the AI Receive?

Syllabium never sends a person’s identity to an AI model: a rule in the database refuses identity data for every provider. Each AI request is a defined action on named content, and the platform assembles what the model is sent.

The authoring assistant works on text only, and sends the text of the content it works on. A model provider can be limited to course and platform text, so that it receives no learner text. A reflection reaches a model only under the enrollment consent, as the learner wrote it, and without the learner’s name, email, user id or cohort attached.

Can You Bring Your Own Model?

Yes. Syllabium embeds LLMs in each block of a lesson and runs them on your own model or ours (BYOM), under the same checks and records as every other AI action.

Each call follows the organization’s model route, or else its tenant’s, or else the platform’s, and without a route no call is made. A tenant can use its own AI account.

Your own modelsBlock-level LLMs

What Is Agentic RAG in Syllabium?

Agentic RAG in Syllabium answers from each field’s own library: it draws only on what the asker’s organization may read, and it cites each claim to its paragraph and version. When the library does not hold enough evidence, the answer says so.

Agentic RAG

Who Can See What the AI Did?

Staff in Syllabium see each AI agent as a named actor: its work carries the mark “(AI)” and the agent’s role. Syllabium also shows staff AI control pages, an AI activity tab, marks on content the assistant changed, and a report of how people decide AI proposals.

In the studio, authors can search a library by author, people or AI roles, and the content tree marks AI activity.

How Are Learners Told About AI?

Before any learner page, each learner in Syllabium accepts the documents in force, among them the notice of how AI is used. Each acceptance is recorded.

How Is the AI Itself Tested?

Syllabium tests its AI with test sets that carry planted errors and must clear set pass marks. Its built-in checks also confirm that each of the twelve checks refuses when it should.

Is Each Use of AI Assessed?

Yes. Each use of AI in Syllabium has its own impact assessment, built to ISO/IEC 42001.

ISO/IEC 42001Standards

Get Started

Enroll your organization online, see the pricing, or talk to us. Syllabium is made and operated by Athantic AI LLC.

Contact us