Enterprise Knowledge AI & Document Intelligence Answers from your own documents, with sources cited and permissions kept

We filter every query against the permissions held in each source system, so a user gets answers only from documents they could open there as of the last sync, which runs at an interval stated in the design. We read the access lists from each source, match them to users and groups in your identity provider, and apply them at query time, before any passage reaches the model rather than after the answer is written. Before go-live we test this with accounts from different departments, for example finance, HR and a user with no special access, asking the same questions and checking that nothing crosses between them. If a document is over-shared in the source today, the assistant inherits that, so we report the over-shared locations we find during ingestion.

We measure both on an evaluation set agreed with you and reduce them through the sources, the retrieval and the refusal rules; we do not promise they disappear. The set holds real questions from your staff in Arabic and English, with expected answers and source passages, plus questions the documents do not answer. We report the grounded-answer rate (answers supported by the cited passage), retrieval quality (whether the right passage came back) and refusals where no source exists, against thresholds agreed before the pilot. Fixes include removing superseded versions, splitting long documents more finely, tuning search for Arabic, and requiring every answer to carry a source or be declined. The set is re-run after each significant change, so a regression shows up as a number, not a complaint.

It depends on scan quality and the kind of handwriting, so we test on samples of your documents before agreeing the scope. Printed Arabic on clean scans usually reads well; low resolution, skewed pages, stamps over text, faded copies, complex tables and mixed Arabic and English lower the accuracy. Handwriting is the hardest case, especially free-flowing text, while short handwritten entries in fixed boxes do better. We measure accuracy per document type on your samples and agree which types are processed automatically, which go to review and which stay manual.

As current as the last sync from each source, at an interval agreed per source and stated in the design. Sources that change often, such as tickets, can sync more often than stable ones, such as approved policies. New documents, edits, deletions and permission changes all follow the same sync. Each answer shows the date or version of the document it used, so users can see how recent it is, and during data preparation we mark the current version where old and new versions sit side by side.

Usually some, because an assistant is only as reliable as the documents it retrieves from, but not a full clean-up before you start. At the start we assess the in-scope sources for duplicates, superseded versions, conflicting policies, documents without an owner or date, and scans without readable text. For the pilot we agree an authoritative source and an owner for each topic, remove or mark old versions, add basic metadata and run OCR on scans. We do not rewrite your content: your owners decide which version is current, and preparation beyond the pilot scope is planned as its own phase.

No, we do not train any model on your data. The assistant retrieves passages when a question is asked and passes them to the model as context, which does not change the model. If the model runs locally on a platform you control, questions and passages are not sent to a model provider; if a commercial model is used through its enterprise API, questions and passages reach that provider under its terms, and whether it may keep them or train on them depends on those terms and the settings switched on, which the design records. Personal data sent to a model outside the Kingdom is a transfer question under the Regulation on Personal Data Transfer outside the Kingdom. Your data stays yours; configurations, documentation and deliverables are handed over as set in the contract.

Wherever a wrong answer or extracted value would affect someone's rights, money or legal position. Examples are replies sent to customers or beneficiaries, interpretations of HR and legal rules, payment data taken from invoices, and contract terms relied on in a decision. We agree these cases with you, record them in the design and name a reviewer for each; internal look-ups, such as where a procedure is kept, usually need only the citation. The organisation-wide human oversight policy belongs to AI Governance & Regulatory Readiness; this service builds the review steps into the system itself.

Most systems that expose their documents and permissions through an API or a supported connector. That typically includes document management systems, intranet and collaboration sites, shared file servers and ticketing or service management systems, with your identity provider for users and groups. What decides it is whether we can read the system's permissions, not just its content; where we cannot, we limit that source to content everyone using the assistant may read, or leave it out. Extracted data can be passed to your systems after validation or review, for your staff to use, but an assistant that approves requests, updates records or sends replies on its own belongs to AI Agents & Workflow Automation.