Spaces

First-week checklist for a central bank trying an AI agent

This checklist walks a monetary authority through the first week of letting an AI agent help with chats, starting with drafts that staff release. It suits teams that need to move slowly and keep control.

Runs in your browser. Nothing you enter is sent anywhere or saved.

Before you start

A central bank cannot afford a rushed launch. The first week is about proving that the agent says only what you want it to say, and that awkward messages end up with the right person. Think of it as a trial in which staff stay fully in charge of every reply.

Spread these steps over five working days and tick each one off. If a step feels uncomfortable, stay on it longer. There is no prize for speed, and a slow start with strong checks is exactly what a public institution should do.

Tick each one as you go. 0 of 10 done.

[ ] Agree who owns the trial and who may stop it at any time.
[ ] Write down the public facts the agent may repeat.
[ ] List the topics it must not answer and the desk for each.
[ ] Set the agent to draft replies that a person releases.
[ ] Test it with ten awkward messages you wrote yourselves.
[ ] Check that every hand-over reaches the right desk with a clear summary.
[ ] Read every chat at the end of each day.
[ ] Fix the notes whenever a reply was wrong or too wordy.
[ ] Tell front-desk and press staff what the agent will and will not do.
[ ] Hold a short review on day five and decide what happens next.

A sensible shape for the week

Use the first two days for writing the notes and the do-not-answer list. Use the third day for testing with made-up messages. Run drafts with real chats on the fourth day, with a person releasing each reply, and use the fifth for review.

Only after several weeks of good drafts should you consider letting the agent send the plainest facts, like opening hours, on its own. Many institutions decide to keep a person in the loop for good, and that is a fair choice.

Warning signs in week one

Stop and fix the notes if you see any of these in the drafts.

  • The agent explains what a decision means for households.
  • It names a bank, a company or a person.
  • It agrees to something on behalf of the institution.
  • It treats a claimed ministry contact as confirmed.

FAQ

Do we need to tell the public we are using an AI agent?

It is good practice to be open about it. A short line in your greeting, saying that an AI agent helps with simple questions and staff handle the rest, builds trust and sets fair expectations.

How many chats will we see in the first week?

Probably few, because members of the public rarely write to a central bank. That is helpful in a trial, since you can read every one and learn what people really ask.

Can the agent look up tour bookings or request status?

Only if your own agent has access to those systems. Spaces carries the messages. Anything beyond answering from your notes depends on the tools you connect yourselves.

More free tools for central banks and monetary authorities

Read the guide: A realistic role for an AI agent at a monetary authority.