An agent on a messy record just automates the mess.

The pilots that failed last year did not fail on the model. They failed because quotes lived in one system, dispatch in a spreadsheet and complaints in three WhatsApp groups — so the agent confidently acted on data nobody trusted.

We have now sat in enough post-mortems of failed AI pilots to notice they rhyme. Nobody blames the model, and nobody should. What went wrong was upstream of it.

An agent is a very fast, very literal colleague who believes everything the record tells it. Point that at a business where the quote lives in one system, the dispatch note in a spreadsheet on someone's laptop, and the complaint in a WhatsApp group, and you have not automated the work. You have automated the disagreement.

The three failures we see most

None of those are intelligence problems. All three are handover problems, and they existed before anyone mentioned AI. The agent just made them visible at machine speed.

Fix the handover, then automate

The teams whose pilots worked did something unglamorous first: they got quote, order, dispatch, invoice and ticket onto one record, so that "what did we sell, ship and bill" had exactly one answer. Only then did they hand a queue over.

That order matters because a clean record makes automation boring, and boring is what you want. When the agent has one source of truth, its mistakes are small and obvious. When it has five, its mistakes are confident and expensive.

How to tell which one you are

Pick your last twenty complaints and try to answer, from the record alone, which order line each one refers to and who owns it. If you can, you are ready to automate a queue. If you find yourself opening a spreadsheet or asking a colleague, start there instead — and we will say so on the call rather than sell you an agent that will embarrass us both.