Spaces

Weekly chat review checklist for central bank teams

This checklist gives a central bank team a short weekly routine for reviewing the public conversations an AI agent handled. It focuses on accuracy, restraint and what the questions reveal about your published pages.

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

Before you start

A central bank cannot hand its voice to a tool and look away. A short weekly read of real conversations shows whether the agent kept to published facts, stayed out of policy and passed things on promptly. It also shows what the public actually wants to know.

Set aside thirty minutes at the same time each week and open a sample of chats, not just the odd ones. Tick each point as you go and write one line of action for every miss. Rotate the reviewer so more than one person understands how the agent behaves.

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

[ ] Read at least twenty chats, including some that went well.
[ ] Check every answer against the published source it relied on.
[ ] Look for any hint of opinion on rates, banks or the currency.
[ ] Count how many chats were handed to a person, and why.
[ ] Measure how long handed-over chats waited for a human reply.
[ ] List questions the agent could not answer.
[ ] Check that dates, hours and closures are still correct.
[ ] Note any reports of scams or suspect notes and confirm they were escalated.
[ ] Look at the tone of replies to upset or confused people.
[ ] Update the instructions and record what you changed.

What the weekly pattern tells you

After a few weeks you will see which publications people cannot find, which forms confuse them and which exhibition dates cause repeat questions. That is useful feedback for your web and publications teams, collected without a survey.

Share a short summary with the teams who own those pages. Keep names and personal details out of it, and quote the question in general terms.

Mistakes to avoid in the review

A review loses value when it becomes a formality.

  • Reading only the first message of each chat.
  • Fixing a wrong answer without finding why it happened.
  • Skipping weeks during busy periods such as a major announcement.
  • Keeping the findings to one person.

FAQ

How many chats should we read each week?

Around twenty is enough for most teams to spot patterns, and more in the week after any change to the instructions or a big public announcement.

Who should do the weekly review?

Someone from communications who knows your published material, ideally with a second reader from the counter or public service team who hears real questions in person.

What if we find a wrong answer?

Correct the instructions first, then contact the person if the error could matter to them. A short, honest follow-up from a staff member is better than saying nothing.

More free tools for central banks and monetary authorities

Read the guide: Answer public enquiries to a central bank clearly and carefully.