How it works
Checking AI’s work takes someone who could have done it.
AI agents build your models and reports. Scripts check the numbers and claims, reviewer agents read the work, and the analytics director signs what you act on: a person with 15+ years in analytics who has built two analytics teams. It starts with a free call and a fixed price.
Step 1
A free call, then a fixed proposal
The call takes up to thirty minutes and covers one question, the decision waiting on it and the data you have. The proposal fits on one page and fixes the price before work starts; changes are priced separately. Scoping deeper than the call is paid work, priced like any other.
A piece of work starts from €12k, and a retainer for repeat work from €8k a month (see pricing).
Step 2
Agents build the work, and three stages check it
AI agents
Build the work
Agents write the queries and code, build the models and draft the report from your data. The analytics director sets the question and runs them, and the agents fix what a later stage sends back.
Can AI replace a data science team?Is AI-made analysis cheaper?
Number and claim checks
Deterministic
Scripts, not AI, check every deliverable: that totals add up, that a figure printed twice says the same thing, and that each quote is in the source it cites. They find the same fault every time, and miss a wrong definition used consistently.
AI review panel
Three reviewer agents and a judge
Every deliverable gets a review by an agent. On what you act on, such as a model going live or a decision memo, the review is a panel: three reviewer agents working apart, each on a different AI model, and a judge who rules on what they found.
Analytics director
A person, not an agent
The director signs what you act on, such as a model going live or a decision memo, and can send it back like any other stage.
Stages 2, 3 and 4 can each send the work back to the agents. If two fix rounds at a stage fail on the same kind of problem, the work stops and the director decides. A fixed piece passes the earlier stages again before it returns to the stage that sent it back.
Step 3
What you receive
A model, an analysis or a decision memo. A recurring report arrives as a document, a notebook or a simple web page, not as a Looker or Power BI build. The work runs on FinDaS’s side, so nothing is installed on your systems.
Each deliverable comes with a short method note listing the checks that ran on it, as every example on this site does. What you act on also carries the analytics director’s sign-off. A routine refresh, the same signed-off report on new data, keeps its method note and needs no new signature.
AI on its own
Why not just use ChatGPT?
The wrong answer looks exactly like the right one. A wrong churn model still scores every customer and a wrong forecast still draws a clean line, with no error to tell you to check.
On DSAgentBench, a 2026 benchmark of 275 end-to-end data science tasks (a preprint, not yet peer-reviewed), the strongest AI agent tested, Claude Sonnet 4.6, had a task success rate of 56.7%. Three human testers in the same study reached 85.1%.
A 2024 meta-analysis of studies pairing people with AI found the pair did better than the AI alone where the person was better at the task, and worse where the AI was. A company without a data team has nobody in that seat.
Known questions are different. In a test by dbt Labs, an AI first modelled the data for 11 set questions; Claude Sonnet 4.6, the model that led DSAgentBench, then answered them through dbt’s semantic layer with 98.2% accuracy. So a routine refresh runs on the agents’ checks, and what you act on carries the analytics director’s sign-off.
Start here