In short
- The key rule: figures are calculated by code from your data, never by the language model. The model writes the text around verified numbers and marks what it cannot support.
- Every statement in the draft points to a figure or a source, so that a reviewer can check it quickly.
- Three levels fit different needs: scheduled reports from a template, AI‑written commentary on exceptions, and an agent that builds ad-hoc reports on request.
- The same pattern works for management reports, project status reports, technical and inspection reports and customer reports.
The problem
Most reports follow the same structure every period, and still someone builds them by hand. Figures are copied from systems into a spreadsheet, a chart is rebuilt, commentary is written from memory, and the result is checked only when there is time. Typical errors are numbers from the wrong period, a formula that was overwritten, and commentary that does not match the figures. The report arrives late, and the people who could analyse it are busy producing it.
The workflow in seven steps
- Specification: the report is defined once: structure, sources, figures, thresholds that count as exceptions, tone and layout.
- Data collection: the system pulls data from your sources, such as the ERP, CRM, accounting software, spreadsheets, databases and forms, through exports or interfaces.
- Data checks: rules look for missing values, duplicates and outliers before anything is reported. Problems are listed, not hidden.
- Calculation: totals, ratios, changes and rankings are calculated by code or in the database. The language model does no arithmetic.
- Drafting: the AI writes the commentary around the calculated figures, in your style, and ties every statement to a figure or a source. It marks statements it cannot support.
- Review: a person reads the draft next to the figures, corrects it and approves. Corrections feed the next run.
- Distribution and log: the report goes out as a document, an email or a page in your tools, and the log keeps data version, rules and approval.
For documents built from drawings, photos and site lists, the example on the AI automation page shows the same pattern with a technical report.
Three levels of automation
Scheduled reports from a template
A fixed report with fixed figures, produced on a schedule. Business intelligence tools already do this well, and a template is often enough.
AI‑written commentary and exceptions
The figures are produced as before, and the AI adds what takes time: a summary of what changed, the largest movements with likely causes found in your data, and a list of exceptions that need attention. This level suits management and project reports.
An agent for ad-hoc reports
People ask a question in plain language, for example “how did the three biggest customers develop this quarter?”, and the agent builds the report from the same governed data, with the same checks and sources. Access rights decide what each person may see. More on agents under AI agents for business.
Systems and interfaces
Reports draw on spreadsheets, databases and the systems of record, for example the ERP, the CRM and accounting software, and are produced as Word or PDF documents, emails or pages in tools such as Power BI or Looker Studio. We connect through exports and the interfaces these systems offer, and the report layout is your own template. The names are for orientation only and do not imply any partnership.
Law and risk
- Accuracy and responsibility: a report that goes to a board, a client or a regulator is the responsibility of the person who approves it. The pipeline makes the check faster, it does not replace it.
- Data protection: reports often contain personal data. You need a data processing agreement with every provider (Art. 28 GDPR), processing in the EU or UK, and access limited to the people who need each report.
- Labelling: AI‑generated text published to inform the public on matters of public interest must be disclosed under Art. 50 of the AI Act, unless it has been reviewed by a person who holds editorial responsibility. Internal reports are outside that rule. More under labelling obligations.
- Out of scope: reports that rate or rank individual employees or applicants are high-risk under Annex III of the AI Act. We do not build them, see risk categories.
- Staff: Art. 4 of the AI Act asks for sufficient AI literacy. This overview is general information, not legal advice.
How to measure success
Compare a few reports from before with the same reports from the pipeline. Useful figures are the time from period end to approved report, the number of corrections the reviewer makes, the data problems caught before reporting, the share of statements with a source and the share of reports delivered on time. We agree the measures with you at the start, because results depend on your data and your reports, and we do not promise figures in advance.
Where this applies
- Architects and engineers: site reports and snagging lists, see AI for architects.
- Construction: progress reports and site diaries, see construction AI.
- Retail and ecommerce: weekly store and range reports, see AI for retail and AI for ecommerce.
- Financial services: management and regulatory figures with sources, see AI for banks and financial advisers.
- Meetings: notes and action lists as short reports, see AI meeting notes.
Sources
- Regulation (EU) 2016/679 (GDPR), Art. 28: EUR-Lex
- Regulation (EU) 2024/1689 (AI Act), Art. 4, 50 and Annex III: EUR-Lex
- OWASP, Top 10 for LLM Applications