When AI assembles the report, reporting stops being a production task and becomes an intelligence task. The pulling, joining, formatting, and narrating are largely mechanical now; the analyst's value moves from making the chart to reading it. That is the whole shift, and it is a change of identity for anyone who was paid to build the deck.
I have been writing about reports for a long time, and the complaint has not changed: most reports are assembled, not understood. What has changed is that assembly is now cheap enough to give away, which finally forces the distinction. Let me take the two halves apart.
Reporting used to be assembly
For most of my career, the weekly report was a craft. You logged into four platforms, exported the numbers, reconciled the differences, built the charts, wrote the commentary, and shipped it by the deadline. The skill was in being fast and accurate under time pressure, and the measure of a good analyst was a clean report delivered on schedule.
Very little of that was analysis. It was assembly with a thin layer of comment on top, and everyone knew it — which is why the report so often went unread. It arrived, it was skimmed, and it changed nothing, because most of the effort had gone into making it rather than into what it said. I complained about exactly this years ago in don't waste time on reporting because people don't care about them, and the problem was never the format. It was that the work stopped at assembly.
What AI takes
The mechanical core of reporting is now largely automatable, and it is a bigger share than most analysts expect.
- Pulling and joining. Fetching numbers from ad platforms, analytics, and the CRM, and stitching them into one view, is a data-plumbing task a workflow does reliably once it is set up.
- Formatting and charting. Turning the joined numbers into the standard views is a template job, and a machine does it identically every week.
- Narrating the obvious. The first descriptive pass — spend up, conversions down, CPM rose — is language generation against the data, and it is accurate enough to save the analyst a lot of typing.
- Anomaly flagging. Spotting the number that moved outside its band is pattern-finding, and a model is good at it across more metrics than a person can watch.
That is a large fraction of the old job. What is left is the part that was always the point.
What stays human
If the machine produces the numbers and the description, the analyst owns the meaning.
The scarce and valuable act is interpretation: deciding what the movement means, connecting it to something the business already knows or should know, and making a recommendation that someone can act on. "Spend rose and conversions fell" is description. "Conversions fell because we shifted budget into a placement that looked efficient on last-click but delivers low-intent traffic, and here is what to do about it" is intelligence.
That second sentence is the job, and it is not assembly. It requires context the model does not have by default — what the client cares about this quarter, what changed last week, what the business already tried. This is the same point I made in 9 tips to get to good insights in your report: the value of a report is its insight, and insight has always been a human product. AI just makes the assembly around it disappear.
Assembly vs intelligence
| Dimension | Reporting as assembly | Reporting as intelligence |
|---|---|---|
| Core task | Pull, join, format, narrate | Interpret, connect, recommend |
| Value measure | Delivered on time | Changed a decision |
| Cost driver | Analyst hours | Analyst judgement and context |
| Machine share | Low (the old model) | High (the new model) |
| Human share | The whole job | The judgement and the "so what" |
| Reader experience | A deck to skim | A recommendation to act on |
| Failure mode | Late and unread | Fast and wrong, or fast and ignored |
Read the last row carefully. The AI-native report fails differently: it is rarely late, and it can still be wrong or ignored. Fast, accurate assembly of the wrong thing is not an improvement — it is the same mistake, delivered sooner.
The new risk: faster production of the wrong report
The temptation when assembly is free is to produce more of it: more metrics, more charts, more automated commentary. That is the same trap as the old 100-slide report, which I have argued against for years, only now the machine can fill it faster.
The discipline that replaces assembly time is editing. If the machine drafts the description, the analyst's first job is to cut it down to the few things that matter and add the interpretation. The report gets shorter, not longer, because the point is to make a decision easier, not to prove effort. Volume of output is now a warning sign, not a virtue.
How to redesign a report around intelligence
The redesign is concrete, and it is the same few moves as the rest of the lifecycle.
- Fix the data plumbing once. Join the sources in a workflow so the numbers arrive clean, and stop hand-exporting. Most reporting pain is upstream of the analysis.
- Automate the description, not the conclusion. Let the machine write the "what happened" and protect the "what it means" for the analyst.
- Lead with the recommendation. Reorder the report so the insight and the recommended action come first, and the supporting numbers follow.
- Set the anomaly threshold. Decide what counts as worth flagging, so the machine surfaces the signal rather than every wobble.
- Cut ruthlessly. If a metric does not change a decision, it goes. A shorter report gets read.
Do that and the weekly meeting changes character: it stops being a walkthrough of slides and becomes a conversation about what to do. That is the outcome worth optimising for. The Execution guide covers the operating side of this, and the for-team-leads track folds it into the wider rollout.
The meeting is the product
If the report is now generated, the meeting is where the value is realised, and it should be designed like it. The agenda is not a walkthrough; it is a small number of decisions, each with the interpretation and the supporting number already attached. When the assembly disappears, the weekly hour can finally be spent on the two or three calls that matter — where to move budget, which message to commit to, what to stop. A report that produces a good meeting has done its job; a beautiful deck that produces no decision has not.
What to measure
Measure the report by whether it changed anything, not by whether it shipped. Track how often a recommendation was acted on, how many metrics survived the cut, and how much of the analyst's week went to assembly versus interpretation. If the assembly share is still high, the workflow is not finished; if the interpretation share is low and nobody acted on the report, you have automated the wrong thing.
FAQ
Does AI replace reporting analysts?
It replaces the assembly half, which was most of the hours. It does not replace interpretation, context, or the recommendation. Analysts who move to the intelligence half become more valuable; those whose value was making the deck need a new definition of the role.
If the machine narrates the numbers, is the report still trustworthy?
Only if you still verify. Automated narration describes what the data says, which is not the same as what is true — the data may be wrong, the attribution may be misleading, the anomaly may be a tracking break. A human owns the verification, especially before a number reaches a client.
Should reports get longer now that assembly is cheap?
No — the opposite. When assembly is free, the discipline is editing. The machine can fill a hundred slides faster than a person; the analyst's job is to cut it to the few things that matter and add the interpretation. Shorter is usually better.
How do we get from assembly to intelligence?
Fix the data plumbing so the numbers arrive clean, automate the description but not the conclusion, lead the report with the recommendation, and measure whether the report changed a decision. The shift is mostly about what you stop doing, not what you add.
The short version
AI takes the assembly out of reporting — the pulling, joining, formatting, and narrating — and leaves the intelligence, which was always the point. The analyst's value moves from making the chart to reading it and recommending an action. The risk is producing more report faster instead of fewer, sharper insights. Redesign around the recommendation, measure whether anyone acted, and let the machine do the typing. The Execution guide places this in the wider lane.
I have made the mistake of shipping a beautiful report that changed nothing more than once, and it took me a long time to see that the beauty was the problem.
If your team has moved from assembly to intelligence, I would like to hear how you got the report shorter — that is usually the hardest part.
Cheers, Chandler