Concept mockup Not an official MPI publication. Independent concept artefact by Andrew Loughran in support of an application for the Senior Adviser (AI for Standards) role. Read the legal notice & disclaimer
Biosecurity Import & Export Standards Directorate Ko ngā Paerewa Whakauru me te Whakaputa Koiora
Concept artefact · v1.0 · May 2026
A working concept artefact

Four working AI tools for BIES, an operating model that keeps them safe, and a value case grounded in your actual workflow.

Built as a worked example of how the Biosecurity Import & Export Standards Directorate could responsibly use generative AI in real regulatory work. Every tool is live. Every value claim is defensible. Every safeguard is named.

Built for the BIES Directorate, MPI
Demonstrates safe, human-in-the-loop genAI in live regulatory workflow
Status working prototype, Anthropic Claude API
Author Andrew Loughran, applying for MPI26/19077920
The opportunity at a glance

Four bottlenecks in the standards lifecycle. Four tools. One control architecture.

The four tools below address the four most measurable time costs in the IHS revision and publication lifecycle. The same human-in-the-loop architecture supports all of them. Together they free an estimated 0.6 to 1.2 FTE of analyst time annually, returned to scientific and technical work that AI cannot do.

Working tools
4
Built and live in this concept. Each runs on the Anthropic Claude API.
Annual analyst hours freed
1,200to 2,400
Conservative central case across the four workflows. See the value model for assumptions.
Estimated dollar value
$110kto $230k
At the central public service hourly rate. Excludes second-order benefits to trade and importer compliance.
Time to first pilot
90days
First tool, with one IHS workstream, with full evaluation and kill conditions in place.
The four tools

Each addresses a discrete BIES workflow. Each is live. Each cites its sources.

Click into any tool. The sample tabs run pre-cached Claude output for instant testing. The "Bring your own" tab makes a live API call from your browser using your own Anthropic API key.

How the suite stays safe

Six controls. One architecture. Auditable, OIA-ready, and reversible.

A genAI assistant is only fit for BIES if a reasonable analyst would trust the output, a reasonable auditor could verify it, and a reasonable Director would sign their name to a standard built on it. The architecture below applies to every tool in the suite.

01 · CITATION DISCIPLINE

Every claim cites the source.

The model returns a verbatim quote and a source reference for every assertion. The analyst can verify in seconds. No claim survives without a citation.

02 · CONFIDENCE FLAGS

Uncertainty is visible, not buried.

Each output element carries an explicit high, medium or low confidence tag. Low confidence elements route automatically to verification before disposition.

03 · SME ROUTING

Scientific or legal questions are flagged out.

Where a theme involves scientific risk, legal interpretation, or technical judgement outside the model's scope, it is tagged for SME review. The model proposes nothing in those cases.

04 · NO AUTONOMY

The tool drafts. The analyst decides.

No auto-publish, no auto-disposition, no system action triggered by the model. Every output is a draft for analyst review and Director sign-out.

05 · OUT-OF-SCOPE HANDLING

Misrouted inputs surface, not vanish.

Submissions, queries or clauses that fall outside scope are listed separately with a recommended routing destination. They never quietly disappear from the analyst view.

06 · AUDIT TRAIL

Every prompt and response is logged.

In a production deployment, every model call, prompt version, and response is captured to MPI records. Reviewable by audit, by the Director, and by any submitter under the Official Information Act.

A complete operating model, decision rights matrix, four-tier risk classification, and pre-mortem are set out in the Governance page.

How to read this site

Four ways in, depending on what you want to see.

Want the workflow analysis first?

The Needs Analysis page maps the IHS revision lifecycle, identifies the four bottlenecks the suite addresses, and quantifies each with real BIES output data drawn from public sources.

Want to use a tool right now?

Open any of the four tool pages. Click "Run analysis" on a sample tab and you have a real Claude output in under a second. No setup. No API key.

Want the governance and value case?

The Governance page is the operating model. The Value Model is a live calculator, slide the assumptions and watch the numbers move.

Want to know who built it?

The About page sets out who I am, why this matters to me, and how my MPI Animal Exports and process improvement work prepared me for the role.

Want to see the full source data?

Every page cites the public source it draws on. The footer of each page links to the relevant MPI documents and the Biosecurity Act sections in play.

Just want to talk?

Email andrew@privateparadox.co.nz. Auckland, willing to relocate to Wellington for the role.