Skip to content

REF / INTELLIGENCE

AI document management that runs before anyone asks it anything

Ask a chat box about a repository nobody has organised and you get a confident answer drawn from files with no type, no dates and no retention position. EDMS by Sibasi does the unglamorous part first. Reading, classifying, extracting and labelling all happen as documents arrive. The conversational layer comes fifth.

See it read one of your documents Capability detail

REF / PIPELINE

Five stages, in this order, on every document

The order matters. Answering a question is only useful once the content underneath it has been read, typed, described and labelled, which is why that stage comes fifth.

  1. AI-01

    Read

    Optical character recognition on everything that arrives as an image.

    Scanned letters, photographed forms and image-only PDFs become searchable PDF/A with a text layer underneath. Handwritten entries on structured forms and checkbox fields are read too, which covers a good deal of what sits in a registry backlog.

    Output

    • Searchable PDF/A
    • Extracted text layer
    • Page-level confidence scores

  2. AI-02

    Classify

    The system decides what the document is before anyone opens it.

    Prebuilt models cover the document types every organisation handles — invoices, receipts, contracts, identity documents. Custom models are trained on your own material for the types that are specific to you: minute sheets, allotment letters, claim forms, tender returns, allocation letters.

    Output

    • Content type assigned
    • Filed to the correct library
    • Confidence threshold for review

  3. AI-03

    Extract

    The fields that matter come off the page as metadata.

    Counterparty, contract value, effective and expiry dates, invoice and LPO numbers, tax PIN, national ID, policy number, case reference. These become searchable, reportable, and available to workflows — without a data-entry clerk in the middle.

    Output

    • Populated metadata columns
    • Values available to workflow
    • Reportable in Power BI

  4. AI-04

    Govern

    Retention and sensitivity attach because of what the document is.

    A contract picks up a retention label of six years after expiry. A document containing a national ID number picks up a sensitivity label and the encryption that goes with it. Neither depends on a member of staff choosing correctly at the moment of filing.

    Output

    • Retention label applied
    • Sensitivity label applied
    • DLP policy in force

  5. AI-05

    Answer

    By now the content behind the answer has been described.

    Copilot draws on documents the asker is already permitted to see, and cites the source document behind each statement. Because classification and extraction have already run, the answers are grounded in described content rather than in a pile of untagged files.

    Output

    • Cited answers
    • Cross-document summaries
    • Draft correspondence from source material

REF / FOUNDATION

Built on Microsoft's AI stack, not on a model we host ourselves

This matters at procurement. The AI services are Microsoft's, covered by Microsoft's own commitments on data handling, residency and model training. Your security team can check those against Microsoft's documentation.

SharePoint Premium

Document processing, prebuilt and custom classification models, content assembly and taxonomy tagging.

Azure AI Document Intelligence

OCR, layout analysis, form and table extraction, including handwriting recognition.

Microsoft 365 Copilot

Natural-language search, summarisation and drafting, security-trimmed and cited.

Microsoft Purview

Sensitivity labelling, data loss prevention, retention labels and the compliance audit surface.

REF / APPLIED

Where it removes work somebody is doing by hand

Six things this changes in the first month, drawn from what clients had people doing manually before.

USE-01

Contract obligations that surface before they bite

Extracted expiry and renewal dates drive alerts months in advance. Ask which contracts renew this quarter, and which carry a termination notice period longer than sixty days.

USE-02

A backfile that becomes searchable, not just scanned

Digitising a registry produces images. Classification and extraction turn those images into records with a type, a date, a subject and a retention position.

USE-03

Audit and information requests answered from the system

Ask what the organisation holds on a subject, get a cited list scoped to what the requester is entitled to see, and export it as an evidence set.

USE-04

Invoice and claim processing without keying

Supplier invoices and claim forms are read, matched against a purchase order or policy, and routed for approval with the values already populated.

USE-05

Summaries of long documents in plain language

A 90-page consultancy report, a board pack or a tender return summarised against the questions you actually need answered, with each point traceable to its page.

USE-06

Unlabelled content found and brought under control

Scan an existing repository for documents holding personal data or contractual obligations that were never classified, and bring them under policy retrospectively.

REF / GUARDRAILS

The questions your security committee will ask

AI in a document repository is a governance question before it is a productivity one. These answers are written so they can go into a board paper as they stand.

It cannot see what the user cannot see

AI responses are security-trimmed to the requesting user's existing permissions. There is no privileged AI account that reads everything.

Your content does not train foundation models

Processing happens inside your Microsoft tenant boundary under Microsoft's commitment that customer content is not used to train its foundation models.

Confidence thresholds route to a human

Classification and extraction below a configured confidence level queue for review rather than being written silently into the record.

Every AI action is in the audit trail

Labels applied, values extracted and answers generated are recorded like any other action, with the model and version that produced them.

Answers carry citations

A Copilot response names the source documents behind it, so a claim can be checked against the record rather than taken on trust.

You choose where AI runs

Some AI capabilities require cloud services. Where policy prevents that, the deployment runs without them and the trade-off is documented at design stage.

REF / QUESTIONS

Asked most often about the AI layer

What does the AI in EDMS by Sibasi actually do?

Four things. It classifies incoming documents so it knows a contract from an invoice from a minute, and files them accordingly. It extracts the fields that matter — counterparty, dates, amounts, reference numbers — into searchable metadata, so nobody types them in. It applies retention labels based on what the document is, so retention no longer depends on someone remembering. And through Microsoft 365 Copilot it lets people ask questions in plain language across the documents they are permitted to see, with a citation to the source document for every answer.

Where is our data stored?

Wherever your policy requires. In a Microsoft 365 deployment, data resides in the Microsoft cloud region selected for your tenant. Where regulation or policy requires data to remain in country, the system can be deployed on SharePoint Server in your own data centre, or in a dedicated Azure region. Data residency is decided during the design phase and documented in the solution design, not assumed.

Is our content used to train AI models?

No. The AI capabilities run on Microsoft services within your own tenant boundary, and Microsoft commits that customer content processed by Microsoft 365 Copilot and its content-processing services is not used to train foundation models. Your documents remain subject to your own tenant permissions: the AI can only ever see what the requesting user is already entitled to see.

REF / NEXT STEP

Bring a document. We will show you what it reads off it.

A working demo against one of your own document types — a supplier invoice, a minute sheet, a claim form — in under an hour.

Book a demo Request a proposal