Trust is a product feature
A QBR deck generator for CSMs at Lumapps: why every design choice was a trust choice, and why the model never writes a number.
Lumapps customers get a Quarterly Business Review, every quarter. Before each one, a customer success manager (CSM) builds the deck. Most of that prep is not about the story. It goes into opening the CRM, BI dashboards and product analytics, finding the right figure for the right quarter, and pasting it into the right slide.
The obvious fix is to hand the whole deck to an AI. We didn’t. A QBR goes in front of the customer, and a wrong number there costs more than the time it saved. So we built a QBR deck generator around one requirement: a CSM has to trust every figure on every slide.
Where the hours go
Gathering the data alone takes about 30 minutes per deck. That sounds small until you multiply it across customers and quarters. The work is not hard. It is slow because the data lives in several places and each CSM collects it alone, deck by deck.
That half hour is time the CSM does not spend on what the customer came for: what changed this quarter, what is working, what to do next. We wanted CSMs back on that story.
The deck is the company’s reputation
A dashboard can be slightly off and nobody notices. A QBR deck can’t. The CSM presents it to the customer, often to an executive, and every number on it speaks for Lumapps. If one adoption figure is wrong, the meeting becomes a discussion about that figure.
That set the bar for the tool. The CSM must be able to present the deck without checking every number by hand.
Automate the copying, not the judgment
An LLM can write a whole deck, numbers included. It can also write a plausible number that is wrong, and nobody catches it until the customer does. So we split the work in two.
The figures come straight from the actual data sources through their APIs. No model reads, rewrites or rounds them. Lumapps Agent Hub, a governed AI engine, drafts the narrative around them, inside guardrails that keep it from adding figures of its own. The CSM then edits the story. Nobody needs to edit the numbers, because no person and no model typed them.
It is boring AI on purpose. There is no chat window and no generated insight. The CSM gets the right numbers on the right slides and a first draft of the words.
Four trust choices
Every design call came back to one question: will the CSM trust what they see?
- Facts come from the source. Contract and adoption figures are read live from the CRM when the deck is built. No model touches them.
- Gaps stay visible. When the data source has no value, the deck shows ”?” instead of inventing one. The CSM fills it in.
- It lives where the work is. The generator opens from a link on the customer record in the CRM. There is no new app to learn.
- It picks the right deck. The CRM knows whether a customer runs on Lumapps or Beekeeper, so the generator chooses the template.
How it works
The figures and the narrative take separate paths. They only meet in the template.
What I’d tell anyone building AI for customer-facing work
Trust is designed, not earned later. If people put your output in front of their customers, they will not use it until they trust it, and they will stop the first time it burns them. Decide early what the model is allowed to touch. Keep facts in the source system, show gaps instead of filling them, and put the tool where people already work.
The generator has just launched. Next we want to pull in the other sources CSMs still copy from, under the same rule: the model writes words, never numbers.