Every Monday morning, in many companies, the same ritual repeats: someone exports figures from three or four tools, pastes them into a spreadsheet, checks the formulas have not broken, builds a few charts, then writes a short commentary for management. Half a day of work, sometimes more, for a document read in five minutes.

This work is valuable, but most of its time is not spent analyzing: it is spent collecting, copying and formatting. That is exactly what can be automated. This article describes how we build reporting that updates itself, with commentary written by AI, without ever sacrificing the reliability of the numbers.

The problem is not the chart, it is the collection

Most companies already have a visualization tool. The bottleneck is upstream: data is scattered across the CRM, the accounting tool, the e-commerce platform, customer support, and each export has its own format. As long as collection stays manual, reporting stays fragile: a shifted column, a forgotten filter, a wrongly selected period, and the figure presented to the committee is wrong.

Automating reporting therefore starts with automating collection, not with adding AI.

Step 1: connect the sources

Each data source is connected directly, through an API or a scheduled extraction, to a shared storage space. No more manual exports, no more copy-paste. Data is retrieved at a fixed time, in a stable format, with history kept. If a source does not respond or returns incomplete data, the system flags it instead of producing a half-empty report.

Step 2: define a single source of truth

Before calculating anything, you need to agree on definitions. What is an “active customer”? Is revenue counted at order, at invoicing or at payment? These questions seem obvious, yet two departments of the same company often answer them differently.

Each indicator gets a written definition, approved by the business, and translated into a single calculation rule. That rule, and only that rule, feeds every report. No more meeting debates about which spreadsheet is right.

Step 3: compute in code, never with the model

This is the golden rule of automated reporting: numbers are computed by code, deterministically. Totals, changes, averages, comparisons with the previous period or the target: everything is produced by verifiable queries and calculations. A language model is not meant to add up columns or compute a percentage: it can get it wrong, and nothing would guarantee the same question gives the same result twice.

Code computes the numbers. AI explains what they say. People decide what to do with them.

Step 4: let AI write the commentary

Once the numbers are computed, AI steps in where it adds real value: writing. It is given the already computed indicators, their changes, the targets and some context, and it produces readable commentary: what is growing, what is declining, the most significant gaps, and possible explanations when the data makes them identifiable.

For example, if sales drop in one region while the number of quotes is stable, the commentary can note that the conversion rate has fallen in that area. The model does not guess: it puts into words gaps the code has already highlighted. It is instructed to cite only the figures it was given, and to flag when a change has no obvious explanation in the data.

Step 5: flag anomalies

Good reporting does not just describe: it draws attention to what is out of the ordinary. Simple rules detect unusual changes, crossed thresholds, missing or inconsistent data: a sudden drop in an indicator, a value outside its usual range, a source that returned nothing. These alerts appear at the top of the report, so readers know immediately where to look.

They also act as a safeguard: a data anomaly caught before sending avoids presenting a wrong figure to management.

Step 6: review before sending, then schedule

The report is generated automatically, but it is not sent automatically to management, at least at first. A responsible person reviews it, checks the alerts, adjusts or completes the commentary if needed, then approves sending. This review takes a few minutes instead of half a day of production.

Scheduling does the rest: a daily report for operational teams, weekly for managers, monthly for leadership, each with its own level of detail. Over time, some low-stakes internal reports can go out without review, once confidence in the system is established.

A concrete example

A services company produces a weekly sales report for its leadership committee. The data comes from the CRM, the invoicing tool and a hand-maintained target tracking spreadsheet. The sales manager spends a good part of every Monday on it.

After automation, the target spreadsheet is replaced by a table maintained in the reporting tool, the other two sources are connected directly, and indicator definitions are formalized with leadership. Every Monday at dawn, the report is generated with its charts, a few paragraphs of commentary and a list of alerts. The manager reviews it, rewords a sentence, adds field information the data does not contain, and sends it before the meeting starts. The time freed up goes to what matters: preparing the decisions to make.

Mistakes to avoid

  • Asking the model to calculate: this is the most frequent source of errors, and the hardest to spot.
  • Automating vague definitions: if indicators are not clearly defined, automation will just produce disputed numbers faster.
  • Sending without review from day one: trust is built gradually, report after report.
  • Piling up indicators: a forty-chart report does not get read. Better a dozen useful indicators, well explained.

Our recommendation

Start with a single report, the one that takes the most time and is most awaited. Formalize its indicators, connect its sources, automate the calculation, then add generated commentary and alerts. Keep a human review until the system has proven itself. Within a few weeks, you turn a weekly chore into a reliable steering tool, and your teams spend their time analyzing rather than copying.