答案就在您的文件里，只是没人有时间去找。
一个用自然语言查询数据的环境——「三年内各地区营业额」「下滑的产品」「这两个文件之间的差异」——并按需生成表格、图表与报告。
- 无需 SQL 与数据透视表
- 计算过程可见
- 涵盖数据库与文件
The bottleneck is not the data. It is the person who can work it.
The data exists, and it is often clean. What is missing is somebody's time.
Only one person can do it
Analyses go through whoever masters pivot tables or SQL. Their queue is the real delay, and a two-minute question waits three days.
Files live in parallel
Three exports of the same month, two different perimeters, no reference version. Time goes into reconciling files rather than deciding.
The BI tool does not cover everything
Certified indicators live in the dashboards. Everything else — the file that arrived yesterday, the one-off export, the question asked only once — is analysed by nobody.
From the question to the table
The system does not guess a result: it writes a query, runs it, and shows you which one.
Mapping the sources
Databases, spreadsheets, exports from business tools: the perimeter is declared, columns are described, possible joins are identified. This is done once, with your teams, and it determines how reliable everything else will be.
Interpreting the question
“Revenue by region over three years” is translated into perimeter, dimensions, measure and period. When the question is ambiguous — which fiscal year, which definition of revenue — the system asks rather than choosing for you.
Generating the query
A query is written and run against the data, across millions of rows if necessary. Nothing is estimated from memory: the calculation is performed by the data engine, not by the language model.
Verifiable output
The result comes back as a table or chart, together with the query used and the number of rows involved. You can redo the calculation, or have your team redo it.
Formatting and reporting
The result becomes an exportable table, a chart or a written report — with the figures, their perimeter and their limits. The conversation continues: “and excluding the Spanish subsidiary?”
What you can ask for
In plain language, as a continuing conversation — each question starts from the previous result.
Multi-dimensional analysis
Revenue by region, by product, by period, by rep, crossed however you like, across several fiscal years.
Trend detection
What is growing, what is declining, what is off its usual seasonality — with the rows that explain the gap.
File comparison
Two exports, two perimeters, two versions: differences identified line by line, including rows present on one side and missing on the other.
Consistency checks
Duplicates, outliers, missing fields, totals that do not add up: flagged before a decision is made on them.
Visualisation
The chart that fits the question — trend, breakdown, comparison — generated on demand and exportable.
Written report
A written summary of performance, with the figures, their source, and what the data does not allow you to conclude.
Three departments, three uses
CRM and ERP exports are cross-referenced by hand every month to prepare the sales review.
Copilot wired into both sources, with perimeter definitions validated once, generating the usual set of tables.
Preparing the review goes from a day to an hour, and questions raised in the meeting get answered in the meeting.
Reconciling budget against actuals is done on files nobody is certain are the right version.
Automated file comparison, consistency checks, and flagging of discrepancies before analysis.
Discrepancies are identified and documented before the meeting, with the detail of the rows behind them.
Production data is exported daily but only looked at when something goes wrong.
Plain-language querying of the history, with outlier detection and drift against the norm.
Drift is seen before the incident, by the people who know the shop floor rather than by the data team.
What it replaces, and what it does not
Our position, frankly: we do not replace a decision-grade BI tool on its certified indicators. We make accessible everything that today is analysed by nobody.
| Data copilot | Classic BI tool | |
|---|---|---|
| Questions covered | Any one-off question, on any declared source | The indicators planned when the dashboard was built |
| Time to answer | Minutes, by the person who has the question | However long it takes to change the report, via the data team |
| Skill required | Being able to phrase the question | SQL, or mastery of the tool |
| One-off sources | Yesterday's file is usable immediately | It must first be integrated into the data model |
| Certified indicators | Not its job — the dashboard remains the reference | Exactly its job, and it does it better |
| Verifiability | Query and rows used exposed with every answer | Fixed calculation, validated once at design time |
What we measure at pilot
On numerical data, the measurement is about correctness first: we validate the perimeter on your real data before any deployment.
- delay between the question asked and the figure obtained, on the declared perimeter
- -80 %
delay between the question asked and the figure obtained, on the declared perimeter
- of answers exposing the query run and the rows used
- 100 %
of answers exposing the query run and the rows used
- of scoping to declare the sources and validate the definitions
- 2-3 weeks
of scoping to declare the sources and validate the definitions
这些量级来自我们在可比范围内于试点阶段的实测。在贵司语料上，它们会在规模化之前实测，而不是事先承诺。
常见问题
Relational databases, Excel and CSV files, exports from business tools. Sources are declared at scoping: that step determines reliability, not the model used.
Every answer exposes the query run and the number of rows involved. Your team can redo the calculation. And we validate the perimeter on your real data before any deployment.
Several hundred thousand rows without difficulty: the calculation is done by the data engine, not the language model. That is precisely what carries the volume.
No, and we do not offer it as such. Certified indicators stay in the dashboard. The copilot covers what today nobody analyses.
Rights are those of your systems. A sales rep does not get an answer built on HR data or on another territory's perimeter.
Not necessarily. Deployment on your infrastructure is possible, and it is the standard case when the data is sensitive.
本方案不做什么
We do not promise one hundred per cent reliability on numerical data, and we are wary of anyone who does. An AI system on figures must be audited, not taken on trust: that is why every answer exposes its calculation and its rows. Nor do we replace a decision-grade BI tool on its certified indicators — if what you need is a regulatory dashboard, this is not the right solution.
Which question does your data never answer fast enough?
Bring a representative file and two real questions. We will work out together whether the perimeter is usable as is.