AI in manufacturing
Natural-Language Reporting: Why “Just Ask the Data” Is Hard
Ambiguous questions, messy schemas and wrong joins. What it takes for plain-language reports to be right.
Trikaan Research & Engineering4 min read
KEY TAKEAWAYS
- Plain-language questions hide ambiguity that must be resolved.
- A semantic layer of business definitions is essential.
- Show the generated query logic and source records.
- Test against a library of real questions.
‘Show me sales by plant last month’ sounds simple. Is that invoiced or dispatched? Which month — calendar or financial? Including returns? Converting plain questions into correct reports requires more than a capable language model.
Here is what makes it hard, and what makes it work.
Sources of error
| Problem | Example |
|---|---|
| Ambiguous terms | ‘Sales’ — orders, dispatches or invoices? |
| Time ambiguity | ‘Last month’, ‘this quarter’ |
| Schema complexity | Many similar tables and columns |
| Joins | Double-counting across one-to-many relations |
| Units | Tonnes vs kilograms vs pieces |
The semantic layer
Define business entities and metrics once — what ‘sales’, ‘yield’ and ‘plant’ mean, which tables and filters apply, which joins are valid. The AI generates queries against these definitions rather than raw tables. The same definitions power human reports, so numbers match.
Clarify, then answer
When a question is ambiguous, the assistant should ask or state its assumption clearly: ‘Showing invoiced sales, calendar October, excluding returns.’
Transparency
- Show filters and definitions used.
- Link totals to underlying records.
- Show data freshness.
- Respect user permissions in every query.
Testing
Build a library of real user questions with verified answers. Run it whenever definitions, models or data change. Track accuracy and fix the semantic layer, not just the prompts.
Frequently asked questions
What is text-to-SQL?
Converting a natural language question into a database query.
Why do AI reports sometimes double-count?
Incorrect joins across one-to-many relationships; a semantic layer with tested relationships prevents this.
Can users trust AI reports?
When definitions are governed, sources are linked and accuracy is tested regularly.
Does it replace BI tools?
It complements or replaces many routine reports; complex analysis may still use BI tools.
- #natural language query manufacturing
- #text to SQL
- #conversational analytics
- #AI reporting
- #ask your data
Trikaan Research & Engineering
Trikaan
Written by the researchers and engineers who build Anvaya, ASTR and Seed Track, with input from Trikaan's delivery and engineering teams.
