Demo by Andrew D Brown: a fictional company handbook, answered by keyword search. No LLM, and nothing you type leaves this page.

Demo by Andrew D Brown: a fictional company handbook, answered by keyword search. No LLM, and nothing you type leaves this page.

AI system

You are about to talk to a machine.

Ada is an artificial intelligence system, not a member of the People Team. Everything it says is generated by software from the Contoso Employee Handbook v2.0.

This notice is required by Article 50(1) of the EU AI Act and cannot be switched off.

The assistant

Ada

A limited-risk HR assistant for the fictional Contoso Ltd. Every answer is grounded in a single approved handbook, cites the clause it came from, and shows you exactly how it got there. Ask it something, then read the trace of how it got there.

A
Ada AI
Contoso HR Assistant · reading Handbook v2.0 (1 Sept 2026)
Retrieved from Handbook v2.0 · not legal advice · not a decision · a human is always available

Honest note about this demo

Ada runs entirely in your browser with no model behind it. Retrieval is real — a proper keyword-and-inverse-document-frequency search over 59 chunked clauses, with a confidence threshold that makes it refuse rather than guess. Generation is not: the answer is assembled from the retrieved clause rather than written by a language model. That is deliberate. It means every behaviour you see here — the disclosure, the citations, the refusals, the escalation, the retrieval trace — is a design pattern you can lift straight into Copilot Studio, a Claude Project, or a custom GPT, where a real model does the wording and these rules do the governing.

Building block 1 · system instructions

The instructions, out loud

This is Ada's entire system instruction. Nothing is hidden. Click any block to see what it does, which of the Five Pillars it belongs to, which law it satisfies, and a question that makes it fire.

Note the delimiters. Each ---BLOCK--- fences one category of information so the model's attention doesn't blur a rule into the data it is reading.

Paste it into Copilot Studio, a Claude Project, or a custom GPT.
Building block 2 · the knowledge base

Contoso Employee Handbook v2.0

One document. One version. Every clause carries a permanent identifier so the bot can cite it and you can check it. This is the whole of Ada's world — she knows nothing else.

Source of truth

One handbook, version-controlled. v1.0 is retired, not left in the folder next to v2.0 for the bot to get confused by.

Clean and dense

Every clause is a self-contained chunk: a heading, the question people actually ask, a short answer, a next action, the legal basis.

No stale numbers

Statutory rates change every April. The handbook signposts gov.uk instead of hard-coding an amount that will be wrong by next spring.

Building block 3 · capabilities & tools

What's switched off, and why

Most people building a bot turn everything on. A gold-standard bot turns things off on purpose — and can say why for each one. Every switch below is a risk decision, not a feature decision.

The one that matters most: no code interpreter

Give an HR bot data analysis and someone will, entirely reasonably, ask it to rank the team by absence days. At that moment the bot stops being a policy look-up tool and becomes a system used to evaluate people — Annex III, point 4 of the EU AI Act, high-risk. That brings a conformity assessment, a risk management system, logging, human-oversight design, registration in the EU database, and a much bigger DPIA.

Contoso's decision is the cheap one: don't cross the line. Ada informs people about policy. Humans decide about people.

The line Ada must not cross

Limited risk — where Ada sits
  • Explains what a policy says
  • Cites the clause
  • Tells you your next action
  • Hands you to a human
  • Duty: tell people it's an AI (Art. 50)
High risk — Annex III(4)
  • Screening or ranking job applicants
  • Evaluating performance or behaviour
  • Allocating tasks
  • Deciding promotion or termination
  • Duty: the full high-risk regime
Governance

Where the law lands on this bot

Every obligation below is matched to something you can actually click in this app. That is the test: not "we thought about the AI Act", but "here is the feature that discharges the duty".

Accurate in outline as at September 2026 and written for training, not as legal advice. The EU AI Act's Article 50 transparency duties applied from 2 August 2026; its high-risk regime for Annex III systems from the same date. The ICO's AI and automated decision-making Code is not yet in force — the enforceable statutory Code is expected in 2027 — so the honest claim is that Ada is designed toward its four stated pillars, not compliant with a code that does not yet exist.

EU AI Act

UK law


The defensible sentence

"Ada is a limited-risk AI system under the EU AI Act, disclosed to every user before first use under Article 50. It is deliberately kept outside Annex III(4) because it neither evaluates nor decides anything about a person. It is built to UK GDPR and the Data Protection Act 2018 with a completed DPIA, an Article 22 human-review route, Equality Act 2010 output monitoring, and security discipline under the Computer Misuse Act 1990 — and it is designed toward the ICO's stated AI Code direction ahead of the 2027 Code."

Say that, be able to point at the feature behind each clause, and you have done the job properly.

Your turn

Build one of these yourself

Ada is only three things: an instruction, a document, and a set of switches. Here is the order to do them in, and the trap at each step.