AI System Registry

Answers questions about the organisation's own supply-chain analysis — alerts, entities, dependencies, metrics, documents — by calling a fixed set of read-only investigation tools and citing what they returned.

Inputs

• The user's question and the conversation so far.
• Optionally a snapshot of the widget the question was asked from.
• Read-only tool results over the tenant's own v2 analysis: overview, alert search and detail, entity, graph neighbours, blast radius, root cause, facts, metrics, document search, lineage and trace.
• The tenant's Company context block, within its configured token budget.

Output

A streamed answer with the list of tools it called, citation chips (alert, entity, lineage, document page/line, widget) and optional inline charts. A citation to an id no tool returned is dropped, not shown. Labelled AI-generated per Art. 50.

Human oversight

Read-only and advisory: the tools cannot acknowledge, resolve or change anything, and a person acts on alerts in the Risk register. At most eight tool steps run before the model must answer, and the user can stop it at any time.

Training & data

Third-party foundation model (Anthropic Claude) through the DefenceOS engine. Questions and tool results are not used to train the model.

Limitations

• It can only see what the analysis holds; gaps in the uploaded data are gaps in the answer.
• It may misread or over-generalise tool results. Every cited item can be opened and checked.
• Non-determinism: the same question can be answered differently on different occasions.
• With AI unavailable or egress disabled, no answer is produced and the panel says so.

Answers questions about the documents the user has loaded in Ingestion, citing the exact passages it used.

Inputs

• The text of the documents in the chosen context (included files, the selected file, all files or a picked set) and any extracted entities.
• The user's question and the conversation so far; optionally the offline supply-chain report.

Output

A streamed answer with citation chips that jump to the cited line of the source document. Labelled AI-generated per Art. 50.

Human oversight

Advisory. The chat changes no document, extraction or analysis; a person decides what, if anything, to do with an answer.

Training & data

Third-party foundation model (Anthropic Claude) through the engine. Document text is not used to train the model.

Limitations

• It answers from the text it was given; OCR errors and unreadable pages carry into the answer.
• Long documents may be truncated to fit the model's context.
• Non-determinism in wording.

Proposes the edits for a change the user asks for in a contract's draft, lists a contract's key terms, obligations, risks, dates, unusual clauses and supplier-risk angle, and answers questions about the contract.

Inputs

• The text of one uploaded contract (its working draft), read by the engine from the organisation's own store — the browser sends only the contract id.
• The change the user asks for, or the user's question and the contract's discussion so far.

Output

For a requested change: exact find-and-replace edits, applied to the draft only where the quoted text occurs exactly once, with a summary and the edits not applied. For insights: structured lists citing clauses. For questions: an answer citing clauses. Labelled AI-generated per Art. 50.

Human oversight

Nothing the assistant does is final: edits land in the draft, which the user reviews in the redline, recreates as a version and sends for approval; a named approver approves or rejects it, and only the approved version downloads. Insights and answers change nothing.

Training & data

Third-party foundation model (Anthropic Claude) through the engine. Contract text is not used to train the model.

Limitations

• Analysis, not legal advice; it can miss or misread a clause, and the redline and approval are the checks.
• It answers from the text it was given; OCR errors in a scanned contract carry into the answer. Very long contracts are cut to fit the model's context.
• Non-determinism: the same request can give different edits or wording.
• With AI unavailable or egress disabled, no change, insight or answer is produced and the page says so.

Reformats a document's extracted text as readable Markdown (headings, lists, tables) without changing its content.

Inputs

• The document's extracted text, in chunks of about ten thousand characters, with page markers kept.

Output

A Markdown rendition the user can view, edit, save, regenerate or discard. A stored result is reused without a new model call. Labelled AI-generated per Art. 50.

Human oversight

The original text stays the source of truth; the Markdown is a view the user can edit or discard, and it is marked stale when the text changes.

Training & data

Third-party foundation model (Anthropic Claude) through the engine. Text is not used to train the model.

Limitations

• It can drop, merge or reorder content, especially in complex tables. Compare with the original before relying on it.
• It does not correct OCR errors in the source text.

Turns what a supply-chain lead describes in conversation into proposed Company context items (priorities, critical suppliers, policies) that a person reviews and saves.

Inputs

• The user's message, the conversation so far, and optionally text parsed from an attached file (pre-filled into the box, never sent unseen).
• The existing Company context items.

Output

A reply plus proposals grouped by section — new items or updates with a line diff — each flagged when it contains a number the user did not state. Nothing is saved until a person chooses Save. Labelled AI-generated per Art. 50.

Human oversight

Every proposal needs an explicit Save (or edit, merge or discard) by the user; items keep a version history. Without AI, an offline keyword classifier proposes items instead, and says so.

Training & data

Third-party foundation model (Anthropic Claude) through the engine. Conversations are not used to train the model.

Limitations

• It can misfile an item into the wrong section or paraphrase it imprecisely.
• Numbers it introduces are flagged as unverified; check them before saving.

Reviews an analysed document against the organisation's Company context and lists issues, each backed by a verified quotation from the document.

Inputs

• The document's text and the Company context block.

Output

A list of issues with severity and verified quotes; a quote that cannot be found in the document is discarded. Results are stored and reused until the document or the context changes. Labelled AI-generated per Art. 50.

Human oversight

Advisory. Review issues raise no alert and change no data; a person decides whether to act.

Training & data

Third-party foundation model (Anthropic Claude) through the engine. Documents are not used to train the model.

Limitations

• It can miss issues and can over-flag. Absence of an issue is not a compliance statement.
• It only knows the Company context it was given.

Writes one document from facts the engine gathered deterministically from the organisation's own data — a response plan, a report, a letter to a supplier, a request for quotation or a board roll-up — citing the alert, row and entity ids the facts carry.

Inputs

• The facts JSON of the run: numbers, names, dates and ids gathered without AI from the tenant's analysis, each with its citation marker.
• The playbook's drafting task and the account's UI language code.

Output

A streamed Markdown draft with citation chips (alert, row, entity). A citation to an id the facts do not contain is dropped and counted as unverified. Labelled AI-generated per Art. 50; a run whose draft came from the deterministic template is labelled as such and is not model output.

Human oversight

Advisory. The draft changes nothing; the playbook's proposed actions are created only when a person clicks Accept, and a person sends any letter. When AI is unavailable or fails, the run says so and continues with the deterministic template, so a run always ends with a draft.

Training & data

Third-party foundation model (Anthropic Claude) through the DefenceOS engine's AiService (egress switch, audit). Facts and drafts are not used to train the model.

Limitations

• It writes only from the facts it was given; a gap in the uploaded data is a gap in the draft.
• It may phrase a fact misleadingly even though the number is right. Every citation can be opened and checked.
• Non-determinism: the same facts can give a differently worded draft.
• The body of outreach and RFQ letters stays English (facts, part names and alert texts are stored in English); only greeting and closing follow the UI language.

Answers a visitor's description of a supply chain issue with two to four possible ways through, each labelled with the weakest evidence it relies on (declared, recorded, observed or corroborated). It has no access to any organisation's data.

Inputs

• The visitor's messages in this conversation (at most the last 20 turns), sent from the browser with each question.
• Up to three files the visitor attaches to a message (PDF, Word, Excel, CSV, text, Markdown or JSON; 10 MB each), sent with that message only.

Output

A streamed answer: a framing sentence, numbered solutions shown as cards with a confidence badge, and the weakest link. Labelled AI-generated per Art. 50.

Human oversight

Advisory and stateless: nothing is stored — the conversation lives in the browser tab only and is gone on reload, and files are processed in memory and discarded. The answer changes no data; the visitor decides what to do with it. The visitor can stop an answer at any time.

Training & data

Third-party foundation model (Anthropic Claude) called by the Skansar server; the browser never holds a model credential. Messages and files are not used to train the model.

Limitations

• It knows only what the visitor writes or attaches; it cannot see any organisation's records.
• It may misread an attachment; long files are truncated to fit the model's context.
• Non-determinism: the same question can be answered differently on different occasions.
• Rate-limited per visitor; when the service is busy or unavailable the screen says so and offers Retry.