The problem
Navito's account managers ran quarterly business reviews with every customer. Each one needed a deck of usage trends, anomaly callouts, and forward-looking recommendations. Building a deck took six hours of data pulls, charts, and copywriting per AM per week - and most of it was the same shape, every quarter.
The previous attempt - a generic BI dashboard - moved the work but didn't reduce it. AMs still rebuilt the narrative from scratch.
What we built
A typed semantic model that names the metrics AMs actually use, paired with an LLM drafting step that turns those metrics into a first-pass narrative deck the AM only needs to edit.
- Typed semantic layer over the warehouse - measures, dimensions, definitions versioned in code.
- Templated draft generation with grounded prompts: every claim cites the metric query that produced it.
- AM editing surface - accept, reject, or rewrite each section with one click.
- Cost ceiling per deck enforced in code, with model fall-back when a draft can be served from cached aggregates.
How we shipped it
Architecture week one: the semantic model came first. We refused to let the LLM see anything that wasn't a defined, typed metric - that's why the system is trustworthy. Build phase ran four sprints with weekly demos to two pilot AMs, then a controlled roll-out across the team.
What we measured
- Hours per AM per week spent on QBR prep
- QBR meeting acceptance rate (customer says yes to the meeting)
- Edit rate per generated section (is the draft good enough?)
- Cost per generated deck
The result
AMs reclaimed 5.5 hours per week. QBR acceptance lifted 19% - better narrative, sent earlier in the quarter. Cost per generated deck holds at < $0.05, well under the ceiling. Most importantly: AMs reported they trust the output, because every claim traces back to a typed query.
Operator's note
"The handoff doc alone was worth the engagement. Our team picked up the codebase on day one and hasn't looked back." - Aditi S., Head of Engineering, Navito