A pulp factory where 50,000 people waited two or three days to "look up a number."
Suzano is the world's largest pulp manufacturer, 50,000 employees. Daily, huge numbers of people need production, inventory, and logistics data. The old way was to ask IT to write an SQL report, and one query averaged two or three days of waiting.
IT drowned in ad-hoc requests while the business side sat idle. Classic case of data within reach but people unable to touch it. Worse, by the time IT's queue cleared, the business decision point had long passed. A buyer who wanted to compare prices waited two days for the table, and the supplier's quote window had already closed.
How they fixed it
Suzano worked with Google Cloud and Sauter to build a Gemini Pro agent: employees ask in plain language, it translates to SQL, and queries SAP material data on BigQuery.
The key design is read-only. The agent can query but not modify. That kills data-tampering risk and lets the compliance team sleep. You ask "how much wood-chip inventory did the Brazil plant have last month" and get an answer, no SQL, no waiting on a person. This read-only stance is not laziness, it is a deliberate guardrail: better to do less than let AI touch production data.
The numbers, translated
Query time for the 50,000 employees dropped 95%. What took two or three days now lands in minutes. The analytics team was freed from 200-plus weekly ad-hoc requests. Self-service query hit 78% adoption within one quarter.
The feel: a line supervisor who wants to see a batch of raw material inventory no longer opens a ticket and waits on IT. They type a question and get an answer. Across 50,000 people, this "ask anytime" removes a mountain of waiting and handoffs. Telecom giants like TELUS saw similar gains, 57,000 members using AI regularly, 40 minutes saved per interaction.
A concrete scenario
Suppose you are Suzano's procurement, comparing two suppliers' on-time delivery last quarter. Old way: request to IT, get the table two days later. Now you ask in the chat box, get the comparison in three minutes. The decision moves from "next week" to "right now," and a backlog of sourcing optimizations gets done in one burst.
The pitfalls and edges, a cold shower
First, it is a single-task specialist, not a general BI assistant. Suzano did not try to build an all-knowing business-intelligence copilot. They locked onto one high-volume, well-defined job, natural-language-to-SQL. Narrow scope is exactly why it was measurable and why they can claim "95%" with confidence.
Second, it depends on clean SAP schemas and Google Cloud. If your enterprise data is messy and table relationships unclear, the agent's SQL will still be wrong. It amplifies the data order you already have, not rebuilds it. Clean the data before the agent.
Third, 95% is a drop in time, not a guarantee of accuracy. Queries got faster, but whether the answer is right still rests on underlying table quality and human spot-checks. Do not skip review just because it is fast.
Three lessons you can lift directly
- Start from the highest-frequency, most standardized lookup. Do not build an all-powerful analysis god on day one.
- Read-only, no writes. Keep data-mutation risk out the door so compliance can sleep.
- Govern the base tables first. The agent is an amplifier, not a rebuilder.
Why this matters to you
If your company's data is scattered and looking up a number means begging IT, Suzano's pattern is worth copying: do not chase an "all-purpose data-analyst AI" on day one. Find your highest-frequency, most standardized lookup, wire a large model to your database, and ship it as a read-only natural-language query. Small companies have less data and can land this faster than Suzano. Three rules: read-only to stay safe, narrow before wide, clean the base tables first. What you save is not one person's effort, but the whole organization's drag of waiting on data. IT also gets freed from "report monkey" work to do things that actually need judgment.