Building AI Tools for Policy & Advocacy Teams

You're not starting from zero

Before the how-to, a snapshot of where the network actually is. Ten AmChams across nine countries told us how they use AI today — and what they want next.

Everyone has started — but 8 in 10 work without a shared method. The opportunity isn't adopting AI; it's getting structured and learning from each other. That's what this guide is for.
Source: AmChams in Europe — AI Adoption Survey, May–June 2026 · 10 respondents, 9 countries · anonymised.

Where AI actually is today

Forget the hype cycle. For a policy team, three shifts matter — and none of them require you to be technical.

From chatbot to agent

Three levels of using AI. Most people stop at level one. The leap that changes how a policy team works happens at level three.

What is agentic AI?

A chatbot responds. An agent works toward a goal — it breaks a task into steps, uses tools, checks its own output, and only comes back when there is something to show. An agentic system is several of these agents organised like a small team.

Think of it as a junior policy team that never sleeps: a researcher, a writer, a fact-checker and an analyst — each given clear instructions once, then reused on every project.

The architecture, without the jargon

It is not a single app. It is a set of plain-text instructions that turn Claude into a policy team. Every layer is a file a non-technical person can read and edit.

What it does every day

Work that used to take days of manual effort, delivered in minutes — with every figure traced to a source.

What it's made of

Three simple building blocks. None of them is code you have to read — they are instructions, recipes and timers that turn one assistant into a working team.

Each block is just a short instruction file in plain language. Add one and the system grows — no code touched.

Where your team meets it

No terminal, no setup for the people who use it. The system runs in the background and you reach it through tools your team already knows — including your phone.

Same brain, different doors. A non-technical colleague uses it from a familiar chat app — without opening a folder or a line of code.

See it in action

Four outputs we generated before this session — nothing typed live. Each one is a plain file your team can open, edit and trust.

Pre-generated on purpose — no waiting, no live failures. Open any of them and every figure traces back to its source.

How your AmCham can start

You don't need our full system to begin. Pick one platform, give it your documents, and you have results the same afternoon. Then grow.

Either path costs around $20–30 / month per person to begin. The platform matters less than the decision to start.

Doing it safely

An agentic system acts on your behalf. A handful of guardrails keep it trustworthy — and keep your members' trust.

Your first agentic system, in two prompts

Open Claude Cowork or Codex, point it at a folder of your documents, and paste these. Hit the Copy button — they are yours to keep.

You are my AmCham policy assistant. Read every file in
this folder and tell me what you found. Then create a
file HOUSE-RULES.md defining: who we are, our tone
(formal, concise, sourced), our priority topics, and one
rule you must never break — never state a figure without
citing its source. Ask me 3 questions before you start.

Create a reusable Skill called "position-paper". Given a
topic and our position, it should produce a structured
draft: context, our stance, key arguments, and a sources
list. Save it as a skill so I can call it again by typing
/position-paper. Show me the skill file before saving.

Make it run on its own — and stay honest

Add a teammate and let the system work while you sleep. Then a few simple habits to keep it trustworthy.

Set up a scheduled task that runs every weekday at 08:00.
Each run: check for new EU and national legislation on
[our priority topic], list only what changed since the
last run with a source link for each item, and save it to
a dated file in /radar. Keep it read-only — draft, never
send. Notify me only when something actually changed.

Create a specialist agent called "fact-checker". Its only
job: read a draft and flag every claim that lacks a source
or looks invented, with a suggested fix for each. It must
not rewrite the draft — only review it. I will call it
after every document before anything goes out.

• Give it your real documents. Context beats clever wording — feed it the files you actually use.
• Start read-only. Let it read and draft; grant "send" or "delete" only once you trust it.
• Iterate. Your first prompt is a draft, not a contract — refine it in plain language.
• Keep secrets out. No passwords, API keys or members' personal data in prompts.