Can AI write your client reports?

'AI reporting' usually means one of three things, and they fail in different ways. Turning a table of figures into sentences. Answering a question about the data in plain English. And noticing what changed without being asked.

The first is mostly a writing problem, and current models handle it well. The second and third are data problems, and no amount of writing ability fixes a data problem.

Drafting commentary from figures that are already right. Keeping the tone the same across a whole roster of clients. Answering a follow-up question in the meeting without anyone opening a dashboard. Summarising a long month into the three movements worth discussing.

Each of those saves a person from typing something they already know. None of them requires the model to know anything the data does not already say.

A model reading marketing data makes the same mistakes a new analyst would, without the instinct to stop and ask.

• Treating different things as one

  Two platforms report a column called conversions, defined differently, and the model adds them together because nothing tells it not to.

• Narrating a broken sync

  A day that never arrived looks exactly like a day when nothing sold. The model will explain the drop, and the explanation will be fluent and wrong.

• Inventing causes

  Asked why a figure moved, it offers a reason that fits, whether or not the data contains one. Correlation comes back phrased as cause.

• Reading a partial day as a full one

  Today's figures are incomplete until the day closes and the platforms finish counting. Compared with yesterday, they always look like a fall.

Modelled tables, with one name, one type, one grain and one currency for each metric, so it cannot add up two things that only share a label. A known freshness state for every source, so a late or failed sync is a fact the model can see rather than a gap it fills. Written definitions, so a conversion has one meaning it can be told. And the plan, so a fall can be read against what was expected.

With those in place the commentary improves without the model changing at all. Without them, a better model only writes more persuasive mistakes.

Let it draft and let a person send. The check is short when the data is clean, and it saves the one email nobody wants to write, which is the correction to a figure a client has already forwarded to their board.

And keep causes human. A model can say what moved and by how much. Why it moved is a claim about the world, and it should come from somebody who knows what happened in it.

The AI chatbot reads the same modelled tables as every client's reports, on your own LLM key, so you pay the model provider directly and its answers use the same definitions as the dashboards. Every source carries a fresh, late or failed state, and anomaly rules include data that simply never arrived, which is the case most often mistaken for a fall in performance.