Generative Reporting: The Next Software Category

The category before this one

Reporting automation has existed for 30 years. Crystal Reports in the 1990s. SSRS in the 2000s. Looker and Tableau in the 2010s. Databox, DashThis, and Whatagraph pushed templated monthly reports into the SaaS era. Each generation expanded who could produce a report without being an analyst.

The ceiling was always the same. Tools could arrange charts. They could not write the narrative. The written summary, the context, the interpretation - those required a human.

What changed

Language models crossed the threshold where they can reliably draft structured business writing from numeric inputs. Not creative writing. Not opinion. Structured, fact-grounded business narrative. Given a table of numbers and a prompt, a modern LLM produces a paragraph that reads like an analyst wrote it, provided the generation is disciplined.

That last clause is the entire game.

The hallucination containment stack

Generative reporting only works if numeric claims are never invented. Four layers are required:

1. Source-of-truth pinning

Every metric that appears in the report has an explicit source, stored as metadata with the claim. "MRR grew 11.2% MoM" is not a sentence; it is a sentence plus a citation pointing at the Stripe query and its result.

2. Constrained generation

The LLM is given the numeric facts as structured input and told to use them exclusively. Free-form numeric generation (where the model makes up a number) is blocked at the pipeline level.

3. Post-generation verification

After the draft is produced, a separate pass extracts every numeric claim and verifies it against the source. Any claim that does not match is flagged or regenerated.

4. Human approval gate

Optional but recommended. A named reviewer approves before send. For recurring reports with stable structure, the approval burden is light after the first few cycles. For high-stakes artifacts (investor updates, board reports), approval is always worth the friction.

What separates generative reporting from AI dashboards

• Output medium. A dashboard is a tool. A report is an artifact. You can email a report. You cannot email a dashboard, only a link to one.
• Interaction model. A dashboard is designed to be explored. A report is designed to be read once.
• Audience fit. Exec stakeholders usually want the artifact. Analysts want the tool. Generative reporting serves the first audience.
• Lifecycle. A dashboard is evergreen. A report is point-in-time. Both are useful; they are not substitutes.

Where the category is going

Three pulls on the product surface over the next 24 months: agent integration (Donum as a tool in Claude, Cursor, and enterprise agent stacks via MCP), brand layer sophistication (reports that are indistinguishable from a human-written artifact for the target audience), and verification quality (hallucination near-zero in steady state).

The category will likely converge with adjacent tools: investor CRM tooling will ship built-in update generation; agency reporting tools will bolt on narrative layers. Donum competes by being the best at the whole loop rather than any single piece.