Skip to content

The AI layer

The work incumbents leave to you

Buying an eQMS is easy. Getting fifteen years of Word files into it — classified, de-duplicated, mapped to clauses, with the gaps identified — is the part that stalls projects for a year. That is what the AI layer is for.

The boundary

Read everything. Write to one place.

The AI layer sits outside the validation boundary so it can improve every week without triggering revalidation of the quality system. That is only safe because of what it is allowed to touch.

The AI layer reads everything and writes nothing except proposals. There is no other path in — not for an administrator, not for us.

What it does

Four jobs, in the order they pay off

01

Corpus migration

The reason projects stall — handled first

Your existing corpus is ingested, classified and organised into document families with proposed numbering, types and owners. Everything arrives as a proposal your quality manager works through — in bulk, with the obvious cases pre-decided — rather than as thousands of files someone has to retype.

  • Scanned PDFs and drawings read with layout-aware OCR
  • Duplicate and near-duplicate clustering into version families
  • Document type and owner extracted from approval tables
  • Orphans and superseded copies flagged, not silently imported

02

Clause mapping & gap analysis

ISO 9001:2015 · API Spec Q1

The standards live in the system as structured data, which means your corpus can be walked against them. The output is the list you would otherwise pay a consultant to produce before a surveillance audit: which clauses are covered, by which document and revision, and which are not covered at all.

  • Standard clauses held as data, not hard-coded
  • Your documents mapped clause by clause
  • Gaps produced as a worklist, not a score
  • New standards are a data load, not a new product

03

Change-impact analysis

“What else does this break?”

Revising a core procedure quietly invalidates other documents, retraining obligations and in-flight records. Doqumus answers that question before you approve the change, by traversing the link graph and explaining what it found.

  • Dependent procedures, forms and work instructions
  • Training obligations that will be reopened
  • Open records referencing the current revision
  • Traversal of the link graph, explained in plain language

04

Grounded search & drafting

Every answer cites a controlled revision

Ask what the current procedure says about a calibration interval and get the answer with a citation you can click through to the effective revision. Ask for a first draft of a new work instruction and get a proposal that still has to be reviewed, approved and signed like anything else.

  • Plain-language questions over the controlled corpus
  • Citations to revision and clause, with quoted text
  • Drafts start from your house style and structure
  • Drafts enter as proposals — never as documents

Governance

Four things that make this defensible in an audit

“We use AI responsibly” is a sentence. These are mechanisms.

One door, structurally

AI output lands in a proposals store and can go nowhere else. Promotion requires a human e-signature, and the promotion event permanently records that the content was AI-drafted and who approved it.

Visibly uncontrolled

Every proposal is marked UNCONTROLLED on screen, in exports and in print. Nothing that came out of a model can be mistaken for an effective revision.

Citations or nothing

Grounded answers cite the controlled revision and clause they came from, down to the quoted text. An answer that cannot cite its source is not returned as an answer.

Recorded and evaluated

Every proposal records the model, prompt version and parameters. A golden document set with known-correct outputs is re-run on every prompt or model change — that harness is the AI layer’s evidence.

The off switch

Turn the AI layer off and you still have a quality system

Whether AI is enabled at all, which model backend it uses and which region it runs in are per-tenant settings. Some customers will start with it off, migrate manually and switch it on later. The core works either way — degrading cleanly is a feature, not a compromise.

Start by finding out what you actually have.

A free program you run on your own machine counts what is really on your shared drive: how many documents, how many are copies of each other, how much is scanned paper. It sends us nothing. Reading it together is the step after that, when you want it.

Get the free scanTalk to us