> For the complete documentation index, see [llms.txt](https://docs.cortex.aic.io/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.cortex.aic.io/workflow-guides/products/ai-assisted-drafts.md).

# AI-Assisted Drafts

Create and review a model-assisted first draft from approved CSV evidence.

**Open this area:** In the CORTEX menu, select **Intelligence cycle → Products**.

Use a bounded, approved model operation on one exact CSV extraction, then check and explicitly save the resulting editable first draft.

**Who uses this:** Author, Analyst\
**Open in CORTEX:** `/products/new`\
**In-app guide:** `/help?topic=ai-drafts`

**Expected result:** A deliberately saved Product retains the first draft's origin and exact evidence without presenting model output as verified intelligence.

## Before you start

* Reporting, Product and Memory must be enabled, with generation, Product read/write, Memory and Direction access in the selected workspace.
* Basic title and section authoring appears first. Multi-source metadata preparation is optional and remains closed until you choose to open it.
* The source-set preview accepts 2–16 distinct Ready CSV material and extraction-version pairs. It is metadata preparation only and remains unverified for generation.
* The 2–4-source evaluation contract and protected v2 Product provenance path do not enable multi-source generation. No production signer or generation action is registered.
* The current generator still uses one Ready CSV linked to the selected requirement: 1–128 nonempty cells, each at most 2,000 characters, and at most 12,000 input characters including requirement context.
* The configured REPORT-DRAFT connection needs current security-owner approval in the matching system assurance baseline for this workspace, deployment, purpose and full handling. Hosted generation also requires no compartments or caveats.

## What to do

{% stepper %}
{% step %}

## Write the basic draft first

On Content, enter the Product title and work through the editable sections.

You can retain an internal draft without opening either optional AI or multi-source preparation.
{% endstep %}

{% step %}

## Open optional multi-source metadata preparation

Choose Prepare multi-source metadata only when you need it.

Find Ready CSV material, select 2–16 exact extraction versions and prepare the metadata preview. The panel distinguishes this metadata range from the separately evaluated 2–4-source offline input bound. It labels freshness Unknown because retained UTC acquisition and processing events are not a freshness assessment, and it remains Unverified for generation.

Closing the disclosure discards its unsaved selection and preview; reopening re-reads current authorised material. Changing requirement or workspace also clears it.

Review exact hashes, warning counts, handling and source-identity groups. Materials with the same Source ID are related material, not independent corroboration.
{% endstep %}

{% step %}

## Understand the protected future origin

The v2 verification path retains every ordered material, extraction, source, publication and custody hash.

A future first save must preserve the exact generated citation order and pass a fresh complete preflight within the one-hour receipt window. Later Product revisions may edit current citations; the original generated citation set and all-source access obligation stay pinned.
{% endstep %}

{% step %}

## Select and inspect the data

In a new Product editor choose the requirement, then open Draft from CSV evidence.

Find returns up to 25 matching visible records. A current selection outside refined results stays listed only after CORTEX re-reads its Ready CSV state and exact extraction under current access.

Choose the processed material and Extraction version, then follow Review source extraction to check the entire retained input.
{% endstep %}

{% step %}

## Select current AI approval

Under Approved AI use case, find the independently approved Generation purpose for this exact requirement, extraction and REPORT-DRAFT runtime, then select Use approval.

Refine the title if it is outside the first 20 visible search results. A selected approval can remain separately visible across a refined search only after a fresh authorised read confirms the same approved revision, protected scope and runtime configuration.

Revocation, revision or configuration drift, a failed search, changing the requirement, material or extraction, or losing current source access clears the choice and replacement confirmation.

You need explicit AI-use access.
{% endstep %}

{% step %}

## Confirm deliberate replacement

Read the provider notice and consider whether this data is appropriate for that approved processing boundary.

Confirm Replace the current draft text with a model-assisted first draft, then select Generate AI report draft.

Current AI, profile, processing and connection decisions are checked around the model call.
{% endstep %}

{% step %}

## Wait for an actual result

Generation is synchronous and has saved nothing.

Use Cancel generation if needed. A failure supplies no substitute model or invented report; inspect the safe error before deliberately retrying.
{% endstep %}

{% step %}

## Challenge every statement

Compare observations and cited cells with the extraction.

Reassess judgements, PHIA likelihood, separate confidence, implications, alternatives and limitations. Correct or remove unsupported prose and preserve required generated citations for the first save.
{% endstep %}

{% step %}

## Report a generated error

Use Report an AI error (new tab) beside the generated draft. Your draft stays open.

The Security form shows captured limitations and the exact generation reference; select an eligible Security owner and record observed facts with full handling.

Security reporting and AI read access are required. This does not approve the draft or bypass normal Product and release review.
{% endstep %}

{% step %}

## Save within the authorised scope

Add the required human processing declarations and use the normal Product save.

The protected first-save receipt lasts at most one hour, bounded by AI/processing validity. First adoption requires the same current AI approval. Saved provenance retains that exact revision; historical products remain subject to ordinary source and Product access.

Continue through ordinary review and release after saving.
{% endstep %}
{% endstepper %}

## Check the result

* Is the whole selected extraction relevant and within the bounds?
* Is this provider approved for this exact input and handling?
* Does every cited observation match its cell rather than merely sound plausible?
* Are likelihood and confidence assessed separately?
* Have I confirmed a real Product save?

## If you get stuck

<details>

<summary>Can AI scan all workspace requirements and other records?</summary>

The current production draft action reads the selected requirement and one exact Ready CSV extraction. It does not collect other workspace records automatically.

Choose the requirement first; the CSV and its extraction version must be linked to it, and current AI-use approval must cover that exact source. The multi-source preview prepares metadata only and cannot generate a report.

</details>

<details>

<summary>Can I generate from the multi-source preview?</summary>

No. It is labelled Unverified for generation and prepares an exact metadata set for the later multi-source workflow. The current Generate AI report draft action remains single-source and requires its existing exact AI approval.

</details>

<details>

<summary>Why can an older multi-source Product remain visible after its AI approval expires?</summary>

Historical read and export require current ordinary Product, Memory, source and custody access to every retained source. They preserve the original AI decision as provenance and do not reuse it as authority for a new generation.

A changed or denied source still hides the complete Product.

</details>

<details>

<summary>Why is my CSV absent or rejected?</summary>

Check Ready state, matching requirement, workspace and search refinement. Search shows at most 25 matching visible records.

A previously selected source is re-read before it can remain outside refined results; a failed search, denial, changed requirement, non-Ready state or missing extraction clears it.

The whole extraction must fit; create an appropriately scoped, separately authorised input through normal acquisition if needed rather than assuming silent sampling occurred.

</details>

<details>

<summary>Why did my selected AI approval disappear after a search?</summary>

Each search re-reads the selected approval and the current REPORT-DRAFT runtime.

The choice is cleared if access is denied, the approval is no longer Approved for Generation, its revision or protected scope changed, the runtime configuration changed, or the search failed.

Find and select a current exact approval before generating.

</details>

<details>

<summary>Why did generation or first save report an expired or changed draft?</summary>

The receipt, generating identity, workspace, source or current AI/processing approval may no longer match. Suspension, reassessment, expiry and changed configuration block new use.

Review current conditions and generate again before first save; manual provenance entry cannot substitute for the protected receipt.

</details>

<details>

<summary>Can a model failure switch automatically to a cloud service?</summary>

No. The configured provider is explicit and has no automatic fallback. Report service failure to the operator or continue manual authoring; Offline rejects hosted generation.

</details>

## Responsibility and limits

* The source-set preview contains metadata and no cell text. Its canonical digest identifies the exact previewed IDs, versions and hashes, but does not create or redefine a governed AI dataset manifest, approve processing or call a model.
* The analyst verifies meaning and inference. Valid JSON, known cell identifiers and model provenance do not prove accuracy or freedom from prompt injection.
* Do not weaken markings or remove caveats to fit a hosted provider. Use an approved local model or manual authoring when appropriate.

**Reference and scope:** Implementation guidance supporting analytical uncertainty and production in JDP 2-00 (4th ed.), 3.41–3.48 and Table 3.3. Model assistance and the eight-section assessment are CORTEX choices, not a doctrine-mandated capability.

## Related workflows

* [Author an intelligence Product and inspect its exports](/workflow-guides/products/product-drafts.md)
* [Record and manage a defensive security incident](/workflow-guides/assurance/security.md)
* [Govern AI use cases and exact data purposes](/workflow-guides/assurance/ai-governance.md)
* [Find evidence and select exact retained passages](/sources-and-evidence/memory.md)
* [Work through connection and deployment limits](/workflow-guides/assurance/offline.md)
* [Record UK processing paths and approved uses](/workflow-guides/assurance/uk-data-flows.md)