Every enquiry got an answer. None of them got recorded.

In mid-August we went looking for the enquiries that had come in through our business messaging line and found almost nothing. The AI assistant on that line had been answering people. It had not been keeping the conversations anywhere our team could see them.

We cannot tell you how many enquiries that was. That is the point of this post.

What happened

The assistant was built to reply to new contacts, ask a few questions and pass qualified enquiries to us. To keep a conversation, the code expected a particular kind of account to attach it to. Production did not have one. So for real people writing in from real phones, the assistant read the message, replied politely, and then wrote the conversation nowhere.

From the outside, everything looked fine. People got answers. From the inside, the leads screen was nearly empty, and an empty leads screen is easy to read as a quiet month.

When we found it, we checked whether the messages could be recovered. They could not. The messaging platform does not provide past conversations on request, and our server logs had already rotated. Whatever came in during that window is gone.

What we changed on 14 August

The fix was to move the record-keeping to the very first step. Every incoming message and every outgoing reply is now written to a simple log before anything else happens: before the system works out who the sender is, before the AI is called, before any rule decides what to do next. Nothing is allowed to skip that step, whatever the conditions.

On top of that log sits one inbox that lists every person who has messaged us, with the full conversation, so a person can read it and reply from the same screen. If the assistant is confused or slow, a human can see that and step in.

We recorded the failure and the fix in our company memory, with the reason, and the principle behind it is short enough to remember: record first, decide second. Any channel we add later starts from there.

The same week, a test conversation turned up a second fault. Partway through an exchange, the assistant would sometimes lose its place and reply with its opening greeting, as if the person had only just written in. The cause was a limit on how many steps the AI could take in one turn. When it ran out, it returned nothing, and the code filled the gap with the greeting. We rebuilt the conversation so that it keeps track of what it has already been told and what it still needs to ask, and so that a reply never falls back to the start.

Why this is an operations problem more than an AI one

Take the AI out of the story and it is a familiar one. A receptionist who takes calls and never writes down the numbers. A shared inbox where everyone assumes someone else replied. A website form that emails an address nobody checks. In each case the customer feels attended to, and the business has no record of the conversation, so there is no follow-up, no quote and no sale.

AI makes the failure quieter. A receptionist who stops writing things down still mentions the busy morning at lunch. An assistant that stops recording just keeps replying.

After the log

A complete log does not follow anyone up. It makes the follow-up possible. Somebody still has to read the inbox, own each new enquiry and act on it within a set time, which takes the same kind of written rule and named owner as any other follow-up.

It also brings a duty with it. A log of every customer message is personal data. Decide who can read it, how long you keep it, and keep it out of any AI tool that does not need it.

Test your own channels this week

Send a test enquiry through every route a customer can use: the website form, the main email address, the phone line, the messaging number, any chat widget. Use a name you will recognise. Then wait a day and ask the team where each one ended up. Any enquiry nobody can find was answered, or ignored, into a gap like ours.

Our revenue leakage check covers the enquiry stage along with the rest of the path to cash. Book 20 minutes at pgco.world/book.