Waiting on an analyst
Simple questions sit in a queue for days because only two people can write SQL.
Home / Packages / AI assistant on your data
Package 03 · Fixed price
AI assistant on your data
A chat assistant connected to your warehouse that turns a plain-English question into SQL, runs it read-only, and shows you the query and the rows behind every answer. Built on your metric definitions, with guardrails you set, in your own cloud account.
from $1,200Fixed price, agreed up front
7 daysDelivery
Your cloudRuns in your account
Who it is for
You will recognise one of these
A warehouse nobody queries
The data is there. The bottleneck is that asking it anything needs an engineer.
Trying AI without exposing data
You want to see what AI on your data is worth, without rows leaving your account or a vendor holding your keys.
WHAT YOU GET
Everything in the package
Delivered in 7 days, in your own cloud account, with the code and documentation handed over.
- An assistant in Slack, Microsoft Teams or a web chat, connected to your warehouse through a read-only database role
- A semantic layer: your table descriptions and metric definitions, so 'revenue' means what finance says it means
- Guardrails: read-only, allow-listed tables, masked personal-data columns, row limits and a monthly cost cap
- Every answer shows the SQL it ran and the rows it used, so anyone can check it
- An evaluation set of 30 real questions from your team, with the pass rate measured and reported before launch
- Deployed in your AWS or Google Cloud account, using the model provider you choose on your own API key
- A one-hour training, the runbook, and the code in your repository
How it works
4 steps, 7 days
STEP 01
Kick-off
Day 1. We collect the questions people actually ask, agree the tables in scope and the model provider.
STEP 02
Definitions and guardrails
Days 2 to 3. The semantic layer, the read-only role, masking and limits.
STEP 03
Build and evaluate
Days 4 to 5. The assistant is built and scored against the 30-question set until the pass rate is acceptable.
STEP 04
Pilot and handover
Days 6 to 7. Your team uses it for real, we fix what they find, then training and handover.
Scope and the small print
- Works with BigQuery, Redshift, Snowflake, PostgreSQL and MySQL.
- Model usage is billed to your own API key, typically a few dollars a day at normal use. A model hosted inside your cloud is possible when nothing may leave it.
- This complements dashboards rather than replacing them. Teams usually pair it with the Metabase package.
- The price covers one warehouse and up to about twenty tables in the semantic layer. Larger scopes are quoted up front.
Questions
Asked before buying
Will it make numbers up?
It does not invent numbers: it writes SQL, runs it, and reports what came back, with the query shown. It can misread an ambiguous question, which is why the metric definitions and the 30-question evaluation exist, and why every answer is auditable.
Does our data go to OpenAI or Anthropic?
The question and the table structure go to the model provider so it can write the query. Rows do not, unless you choose to let the model summarise results. If nothing may leave your account, we run a model inside your cloud instead.
What about personal data?
Columns you mark as sensitive are masked at the database-role level, so the assistant cannot read them no matter what it is asked.
Which chat tools are supported?
Slack, Microsoft Teams and a web chat page. Others on request.
Let's talk
Not sure it fits? Ask.
A 30-minute call, no deck. We will tell you whether this package is the right one, what we would do first, and whether you even need us.