Consultation response workflow
When governments propose a policy, they invite the public, businesses, and civil-society groups to send written feedback — a process called a consultation. Analysts then have to read hundreds of submissions and turn them into something decision-makers can use. This tool demonstrates that workflow.
The analyzer demonstrates a workflow for turning consultation input into policy-relevant themes, trade-offs, evidence gaps, 'You Said / We Did' language (the standard format governments use to show the public how their feedback changed the policy), and briefing outputs.
Sample workflow snapshot
This demo uses pre-written example outputs instead of calling an AI live, so nothing you type ever leaves your browser. It shows exactly what the production system would produce — see 'Production architecture' below for how the live version would work.
Sample responses
9
Example projects
3
Briefing sections
12
Your data
Stays in your browser
Example: a consultation on AI-generated content labelling receives sample responses from civil society, startups, and platforms. The tool organizes themes, identifies evidence gaps, and prepares draft briefing material for review.
How it works
The tool is designed as a transparent workflow. It keeps the path from input text to briefing output visible so policy reviewers can see what requires validation.
Create a consultation project and add fictional or sample stakeholder submissions with stakeholder type, position, source notes, and response text.
The structured prototype workflow groups repeated issues such as transparency, implementation burden, safeguards, public trust, and evidence needs.
Response summaries flag missing metadata, unsupported claims, limited evidence, and questions that require targeted follow-up.
The analysis separates areas of agreement and disagreement from minority concerns and politically sensitive trade-offs.
Briefing cards outline possible routes, likely supporters, likely objections, delivery complexity, risks, and evidence required.
The sample briefing produces structured sections, including stakeholder summary, theme analysis, risks, options, and 'You Said / We Did' language.
Outputs are preliminary and must be reviewed by qualified policy officials before they inform any decision, publication, or engagement.
Production architecture
In plain terms: this page shows canned outputs; the live version would use an AI model at two controlled points, with logging, human review, and strict data-handling rules around it.
This prototype pre-computes its outputs to stay free, private, and stable for public review. The production system it simulates is specified below.
Intake - submissions ingested with stakeholder metadata. No submission text leaves the environment without a data-processing agreement covering the model provider.
LLM first pass - per-submission stance coding and theme extraction, using a pinned model version, with prompts and outputs logged for audit.
Rule-based aggregation - theme frequencies and evidence-gap counts computed by inspectable code between model passes.
LLM drafting pass - briefing sections drafted from the aggregated structure, never directly from raw submissions.
Human review gate - model-coded stance divergences from a stakeholder's self-declared position would be routed for individual review before publication. No output should be exported or used in decision-making without named sign-off.
Illustrative estimate at July 2026 public API pricing, assuming 2-4 page submissions and a Claude Sonnet-class production model. Estimates cover model inference only, before engineering, hosting, logging, privacy review, human review, and QA. Pricing estimates are illustrative and based on publicly available frontier-model API pricing assumptions as of July 2026; exact costs vary by provider, model, token volume, retries, and implementation design.
| Per submission (stance coding + theme extraction) | ~$0.02-0.04 |
| Per briefing draft | ~$0.10-0.20 |
| Full 500-response consultation, end to end | ~$15-30 |
A worked comparison of rule-based and LLM-assisted analysis is planned for a future version of this prototype.
Use the seeded sample data for public review, or create fictional local projects in the browser. Downloads and exports run locally from the static page.
Do not enter confidential, personal, sensitive, proprietary, or real stakeholder submissions into this portfolio tool. Use sample or fictional text only. Outputs require qualified policy review.
Summarize where groups stand, what they support, what they oppose, and what they want decision-makers to consider.
Identify repeated themes, missing evidence, weak arguments, implementation concerns, and follow-up questions.
Separate areas of broad agreement from contested issues, minority concerns, and public-policy trade-offs.
Generate structured summaries, policy options, risks, next steps, and 'You Said / We Did' draft language for review.
This structured workflow is intentionally conservative. It demonstrates organization and briefing structure, not automated decision-making.
Metrics for theme frequency, missing metadata, missing evidence, and implementation barriers.
Theme and stakeholder summaries stay linked to the sample response register for review.
Draft outputs support policy options, risks, follow-up planning, and briefing pack exports.
Disclosure
Model
Planned production use: first-pass stance coding and theme extraction, plus interface development assistance.
Human
Workflow design, data-risk boundaries, and review rules for any divergence between self-declared position and model coding.
Why
See 'Production architecture' above for where a model would sit in the live pipeline and where accountable review gates would apply.