Skip to main content

AI Analyse ISAD

Draft ISAD(G) collection-level field values for a single node in your hierarchy — fond, sub-fond, series, or sub-series — from its contents and parent context. The AI proposes values; you review and refine before saving.

Available on Professional plans and above.

When to use it​

  • Bulk fond description — you've built a structure but haven't written descriptions yet
  • Serialisation refinement — tighten series-level scopes and contents after importing a hierarchy
  • Inheritance verification — check whether parent context is being inherited correctly at deeper levels
  • Sparse nodes — a series or sub-fond with little written description; the AI can infer it from its contents

What the AI reads​

When you click AI draft:

  1. This node's existing metadata — any ISAD(G) values already saved, so the AI refines rather than overwrites
  2. Parent context — the descriptions and dates of parent levels, so inherited values are consistent
  3. Descendant structure — how many sub-levels, series, and sub-series sit beneath this node
  4. Contained accessions — their titles, descriptions, date ranges, and AI-generated themes
  5. Item sample — metadata from a capped sample of items (up to 40) within contained accessions
  6. Your steer (optional) — free-text notes or an emphasis from you: a date range to highlight, known genre focus, or an angle to polish into the description

How to use it​

  1. Open any fond, sub-fond, series, or sub-series from the Fonds page.

  2. Near the top, you'll see an AI draft section with:

    • A text box for optional steer notes (e.g. "Emphasise the photographic content, focus on 1920s–1950s school records")
    • An AI draft button

  3. Click AI draft. Processing takes 20–40 seconds.

  4. The AI proposes values for the ISAD(G) fields below. Each proposed field shows:

    • The proposed value
    • Confidence — how strongly the evidence supports it (high / medium / low)
    • Rationale — one sentence on what the AI inferred and from which evidence

  5. Review each field. You can:

    • Accept it — tick the checkbox
    • Edit it — change the text before saving
    • Reject it — leave it unticked (the field stays as-is)

  6. Click Save changes to commit. Nothing is persisted until you explicitly save.

Reading the proposal​

LabelMeaning
High confidenceMultiple lines of evidence support this value strongly (e.g. consistent dates across accessions, matching subjects across items)
Medium confidenceSome evidence, but gaps or ambiguity. You should review.
Low confidenceSparse evidence or conflicting signals. Treat as a rough draft.

The AI never invents facts. If the evidence doesn't ground a field, it omits it entirely. An empty proposal is not an error — it's the signal: "I couldn't ground any fields; use your own judgment."

What it's good at​

  • Date inference — consolidating ranges from accessions and items into a series span
  • Genre and record-type identification — matching item types (correspondence, minutes, photographs) to series titles and scopes
  • Creator / accumulator refinement — distinguishing different donors or creators within a fond level
  • Multi-language hints — identifying dominant languages in the material (if items carry language tags)

What it's less good at​

  • Fine-grained administrative history — the AI infers structure from evidence, not from external knowledge. If a government department was renamed or abolished, you'll have to tell it via a steer note.
  • Provenance gaps — if the accession descriptions don't mention a key creator or context, the AI can't infer it.
  • Controversial or politically sensitive reclassification — these require human judgment and institutional policy, not AI drafting.

For these, use the steer box. The more specific you are, the better the proposal.

Steer notes — examples​

GoalSteer note
Emphasis on a dominant genre"These are predominantly correspondence; minimise the brief mentions of receipts and accounts"
Known date focus"The bulk of material is 1920s–1950s; the post-1970 items are sparse and ancillary"
Creator / donor context"All from the Smith family papers, but Series A is business records and Series B is personal correspondence"
Institutional structure"Pre-1980 records are organised by department; post-1980 by function"
Reparative narrative"This collection documents the history of [community]; centre their voices and reframe colonial-era records as contextual"

Tips​

  • Small steps. Analysing a few series at a time is more accurate than trying to analyse a whole 10-level structure in one go. Once you're happy with a series, move to its children.
  • Use existing metadata. If you've already written fond-level descriptions, the AI inherits them and refines series-level ones accordingly.
  • Consistency check. If two sibling series come back with contradictory information, that's a signal to add a steer note clarifying the boundary.
  • Re-run if you improve descriptions. Better accession-level descriptions feed better series-level proposals.

See also​