What AI summaries cannot do

AI-generated news briefings have become very polished. The major models now produce summaries that are grammatically clean, factually grounded, and broadly accurate about what has happened. For a reader who wants the shape of yesterday's news, they are useful. For a professional who wants to know what is actually happening in the market they operate in, they are something more dangerous than useless: they are confident, and wrong, in ways that are hard to detect.

The reason is structural. An AI summary is a synthesis operation. It reads a large number of sources, finds the claims they share, and surfaces the areas of agreement. The output is the consensus of what was published. That is simultaneously its value and its limitation.

Why consensus is the wrong output for business intelligence

The consensus is, by definition, already priced in. In markets, in competitive dynamics, in hiring decisions, in customer behaviour: the information that has been widely published and broadly understood is the information that every counterparty has also read. Acting on it is acting after the fact. The edge in any decision-making context is the signal that the consensus has not yet absorbed: the early indicator, the divergent data point, the development that was published but not understood, or that happened but was not yet published at all.

This is not an abstract point. It describes the concrete difference between a briefing that tells you what happened and a briefing that tells you what it means. The former can be generated automatically at low cost. The latter requires editorial judgment that is specifically calibrated to identify where the received view is incomplete, and why that gap matters for the decisions a specific reader is making.

What editorial judgment actually does

The function of a well-run business briefing is not to synthesise what was published. It is to apply a filter that removes what was published but does not matter, and then to apply a second judgment about what the remaining signal implies. The omission decisions matter as much as the inclusion decisions. But the harder function to replicate is the implication judgment: given what happened, what does it mean for the kinds of decisions that a UK founder, operator, or investor is making this week? Not in the abstract. In the specific sense of knowing that a CFO preparing a financing round cares about different signals than a founder choosing between markets, and that both of them care about different things than an equity analyst running a consumer-staples book.

An AI trained on publicly available text can produce a plausible approximation of this judgment. It cannot produce the thing itself, because the thing itself depends on contextual knowledge that is not in the training data: conversations with people who are not publishing what they know, familiarity with the actual decision calculus of the reader at the current moment, and pattern recognition across signals that individually look routine but together suggest something is shifting.

The data lag problem

There is a second structural limitation that specifically affects economic intelligence. The official statistics that describe the UK economy (ONS releases, Bank of England data, labour market figures) are published on a significant lag. When the ONS publishes a consumer spending figure, it is describing what happened six to eight weeks ago. By the time it is published, the AI summaries have absorbed it, the consensus has formed, and it is already historical.

The gap between what official statistics describe and what is actually happening now is not filled by reading more sources. It is filled by tracking the underlying signals continuously rather than waiting for the periodic release. CPIx does this across six component series (wages versus inflation, consumer credit, labour market conditions, retail demand, household savings, and energy costs), producing a read on consumer financial stress that describes the current state rather than the recent past. An AI summary of ONS data describes what was. A composite indicator maintained against live data describes what is. These are different products for different purposes, and conflating them is an expensive mistake.

What to do with this

The practical implication is not to stop using AI tools but to be precise about what they are useful for. AI summaries are good for orientation: the broad shape of what happened, stories you might have missed, questions worth investigating. They are a poor substitute for editorial judgment about what those events mean, and they are structurally incapable of being early.

The question worth asking of any information source is: when did this become available to everyone? If the answer is now, or yesterday, the information is already priced in. The sources worth investing time in are the ones that are early: ones that surface signal before it becomes consensus, that interpret before the interpretation is obvious, that build indicators from components rather than waiting for the aggregate to be released. That is what a well-run business intelligence briefing is for. It is not what an AI summary is for, and knowing the difference is, right now, an underpriced advantage.