How to Build Your First AI Automation

The job we are going to build

Enquiries arrive. Somebody has to read each one, work out whether it is a sales question, a support problem, a billing query or somebody trying to sell you search engine optimisation, and then tell the right person. It takes two minutes and it happens all day. Nobody enjoys it, and whoever is doing it is usually doing it badly by four in the afternoon.

By the end of this lesson that job runs itself: every enquiry gets summarised, categorised and posted where the team can see it, with the original one click away. Nothing is sent to the customer. Nothing is deleted. The worst thing this automation can do is put a message in the wrong column of a board, and somebody will drag it back in a second.

Read it through once even if your job is different. Change the words enquiry and channel for whatever you have and the shape holds.

Step zero, which people skip

Collect ten real enquiries from the past month before you open any tool. Print them, paste them into a document, whatever suits. Pick them deliberately: six ordinary ones, two strange ones, one that is obviously spam, and one that arrived with almost nothing in it.

Now write down, by hand, what the right answer is for each. Category, one line summary, who should see it.

This takes fifteen minutes and it is the highest value quarter of an hour in the whole course. You now have an answer sheet. Without it you have no way to tell whether your automation is working, and you will end up judging it on whether the output looks confident, which it always does.

The four steps

Step one, the trigger. A new enquiry arrives. That might be a form submission, an email to a shared address, or a new row in a sheet. Whichever it is, this is the one thing that starts everything.

Step two, the AI step. Take the enquiry text and turn it into three fixed pieces of information: a category from a closed list, a short summary, and an urgency flag.

Step three, the post. Put the result somewhere the team already looks. A channel, a board, a shared sheet. Include the sender's name, the category, the summary and a link back to the original message.

Step four, the fallback. When the AI step returns SKIP, post it in a separate place labelled needs a human. Not the main channel, and definitely not nowhere.

That fourth step is what makes this a grown up automation rather than a demonstration. There is now a defined place for things it could not handle, which means the pile of unhandled things is visible rather than imaginary.

You are sorting incoming enquiries for a small team. Read the
enquiry below and reply in exactly four lines, nothing before or
after them.
Line 1: CATEGORY followed by one word from this list only.
SALES SUPPORT BILLING SPAM
Line 2: SUMMARY followed by one sentence under twenty words,
saying what the sender wants. No greeting, no adjectives.
Line 3: URGENT followed by YES or NO. YES only if the sender
states a deadline within seven days or reports that
something is currently not working.
Line 4: CONFIDENCE followed by HIGH or LOW. Use LOW whenever
two categories would both be defensible.
Rules:

Never invent a category outside the four listed.
Never repeat the sender's contact details in the summary.
If the enquiry is empty, unreadable, or is only a greeting
with no request, reply with the single word SKIP and stop.
ENQUIRY:

Look at what that instruction refuses to leave open. Four categories, not an invitation to invent one. A sentence limit rather than a request to be concise. A definition of urgent, because otherwise urgent means whatever the sender's tone suggests, and the loudest customers are not the most urgent ones. And a confidence line, which costs nothing and is the single most useful thing you can add to any AI step.

Test mode, then a handful, then alone

Do not switch this on. Run it once, by hand, on the first enquiry from your ten, and compare the result with what you wrote down. Then the second. Work through all ten.

You are not looking for perfection. You are looking for the pattern in the disagreements. Almost always it turns out you and the automation are using a word differently. Your billing category means anything mentioning money, and it is reading billing as invoice problems only. That is a one line fix to the instruction, and it is the only kind of fix worth making at this stage.

For each of your ten enquiries, fill one row.
ITEM  WHAT I SAID        WHAT IT SAID       MATCH?  NOTE
1     SALES, not urgent  SALES, not urgent  yes
2     BILLING            SUPPORT            no      it reads
billing as
invoices only
3     SPAM               SPAM               yes
...
Then answer three questions.

How many matched? Under eight, fix the instruction, not the tool.
Did every disagreement come from the same cause? Usually yes.
Did SKIP fire on the empty one? If not, your fallback is fiction.

Letting go

When eight or nine of ten match and you understand the ones that did not, switch it on. Then do one more thing, which takes a fortnight and no effort: keep opening the originals for a few days. Not all of them. Just enough to keep the connection between what the automation says and what actually arrived.

DAY 1        Switch on. Read every post against its original.
DAYS 2 to 3  Read half of them against their originals.
DAY 4        Check the needs a human pile. Is it empty or growing?
DAY 7        Count the LOW confidence ones. If more than one in
five, your categories overlap. Redraw them.
DAY 14       Ask the team one question: are you reading these,
or ignoring them? An ignored channel is a failed
automation, however accurate it is.
Then stop checking daily and move to the weekly sample in
lesson four. Do not skip that lesson. It is the one that
decides whether this thing is still right in six months.