All work
142025SCAD

Executive Dashboard Narrator

Bilingual executive briefs on the morning of the review — grounded in the actual numbers, with zero hallucinated figures.

Before

Leadership wanted prose commentary on top of Power BI dashboards — month-over-month narrative, not just charts — but writing it lagged the data by two days.

After

A scheduled job reads dashboard datasets, runs significance tests, and writes a four-paragraph bilingual executive brief grounded in the real numbers. Briefs now land the morning of the monthly review, it's the default lead-in to the executive deck, and there have been zero hallucinated numbers in 9 months.

Challenge

Leadership wanted prose summaries on top of Power BI dashboards — not just charts — with month-over-month commentary.

Approach

Scheduled job reads dashboard datasets, runs significance tests, and writes a 4-paragraph bilingual executive brief grounded in the actual numbers.

How it was built

  1. 1

    Read the data, test significance

    Phase 1

    Started from the datasets, not the charts — running significance tests so the narrative highlights changes that actually matter.

  2. 2

    Templated numbers, generated prose

    Phase 2

    Templated every figure and let the model write only the connective prose — the design choice that guarantees no invented numbers.

  3. 3

    Bilingual, on schedule

    Phase 3

    Produced Arabic + English briefs on a schedule so they're ready the morning of the review instead of two days later.

Key architecture decisions

Template the numbers, generate only the words

Why · LLMs invent figures. Injecting templated, verified numbers and letting the model write only prose is what delivered zero hallucinated numbers in 9 months.

Significance testing before narration

Why · Without it, the brief would narrate noise. Testing first means the commentary focuses on genuinely meaningful movements.

Impact

  • Bilingual briefs delivered the morning of the monthly review (was 2-day lag)
  • Adopted as the default lead-in for the executive committee deck
  • Zero hallucinated numbers in 9 months of running (numbers are templated, not generated)
-2 days
lag
0
hallucinations
AR + EN
languages

What I'd tell someone building this

  • 01 · For number-heavy generation, template the numbers and let the model handle only language.
  • 02 · Run significance tests first so you narrate signal, not noise.
  • 03 · 'Zero hallucinated numbers' is an architecture choice, not a prompt.

Tech stack

Power BIGPT-4Azure FunctionsLogic Apps

Ask anything about Executive Dashboard Narrator

AI scoped to this project · GPT-OSS 120B