Every Monday morning, in a lot of small businesses, someone opens the help desk, counts last week’s tickets, goes through the review sites, works out an average and types it all into an email. Then the same figures get typed again into the leadership scorecard. It’s slow and easy to get wrong.
We automated both of those reports inside a real UK equipment business with a busy customer service inbox. Here is what we did, so you can do the same.
What runs now, every Monday
Just after 8am, an AI assistant pulls last week’s figures from the help desk: replies sent and tickets resolved by each person, how long new customers waited for a first reply, and how many customers are waiting on us right now. It emails a summary to the customer service lead and checks in two KPIs on the leadership scorecard.
At about 9am, it reads the live review sites, counts new reviews by star rating, checks every figure against a second source, and writes the weekly reviews email in the layout the team already knew. Then it logs two review KPIs on the scorecard.
Nobody has to do anything, unless a check fails.
Step 1: agree what every number means
This was the most important step, and it had nothing to do with AI. “Last week” and “response time” can each be worked out several ways. The help desk’s built-in “last week” ran Sunday to Saturday, for example.
So we wrote the rules down. Last week means Monday 00:00 to Sunday 23:59, UK time. First response time is measured on tickets created last week, in clock hours, because a customer’s wait doesn’t stop at 5pm on a Friday. The AI agent that already answered some tickets is kept out of the team totals, because it replies in seconds and would flatter everyone.
Here is a prompt you can use to do the same for your own report:
“I want to automate a weekly report on [topic]. Interview me one question at a time until every figure in it has a written rule: what exactly counts, what is left out and why, which dates last week covers, which system it comes from, and how it should be shown. Then write a one-page definitions sheet in plain English, with a reason next to each rule.”
Step 2: read live, from two sources where you can
Early on, one week’s reviews email repeated the previous week’s figures. The page that had been read was an old cached copy. That one mistake shaped the whole design. Now every figure must come from the live sites during the run, never from search results, old emails or memory. The public review page is cross-checked against the review site’s business app, review by review.
Step 3: check before anything is sent
Before an email goes, the totals must add up, each person’s row must match the raw data, and the two sources must agree. Then a separate checker reads the finished email against the raw data. This is the prompt for that second check:
“You are checking a report written by someone else. Here is the draft email [paste] and the raw data it was built from [attach]. Compare every number, name and date in the email with the raw data. List anything that doesn’t match exactly, or anything the data doesn’t support. Do not rewrite the email. Reply ALL CHECKS PASSED only if you find nothing.”
Step 4: draft first, send later
Both reports ran as drafts until a person had checked a full week line by line. Only then were they allowed to send on their own, and only when every check passes. If anything is off, the email is saved as a draft with “Please check” at the top, and one short note explains why.
Step 5: log it to the scorecard without treading on anyone
Straight after each email, the run checks in that week’s KPIs on the leadership scorecard. It never overwrites a figure a person has already entered: it leaves it and reports both numbers. After checking in, it reads the value back to make sure it landed in the right week.
What the first week showed
The reviews report’s first live run went from start to finished draft in about five minutes, with no corrections to the figures. Both review sources matched on every review, and it turned up four follow-ups for the team, including two older one-star reviews that still needed attention.
The customer service summary showed a team that replies quickly on most tickets, but with a large backlog of customers waiting, much of it parked on the AI agent’s own account, plus open tickets with no owner. These are specific, fixable things, which is exactly what a weekly report is for.
Your checklist
1. Pick one weekly report that someone does by hand.
2. Write a rule for every figure in it, with the reason.
3. Connect the AI to your help desk and email, and sign it in to anything it needs to read.
4. Make sure it only reads: it never replies to, changes or closes anything.
5. Add checks: totals add up, sources agree, a second checker reads the email.
6. Run it as a draft until a full week needs no corrections.
7. Let it send only when every check passes, and hold a draft when they don’t.
8. Log the figures to your scorecard, without overwriting anyone.
9. Keep the instructions and a run log where the team can see them.
One safety point: use a business account for your AI assistant and check its data settings before sharing customer or financial details. Claude, Microsoft Copilot and ChatGPT can all do this kind of work; we used Claude.
Build it yourself, or have us build it
If you want to build your own, our guide Customer Service on Autopilot has every rule, every check and what happens when it fails, the problems we hit, 20 copy-ready prompts, and the full scheduled instructions for both reports.
If you’d rather have it done for you, our Customer Service Reporting Setup builds both reports inside your business, connected to your own help desk and review sites, and logged to your scorecard. It’s £795 + VAT, fixed price, done in two weeks.
We’re not IT people. We’ve sat in the MD’s chair and built the systems behind it, and we’ll talk to you in plain business terms. Not sure where to start? Take our free Growing Pains Check.
Leave a Reply