Answer finance questions on agreed definitions.
Plain-language questions answered on agreed definitions, each with an as-of date and a status, and “unavailable” when the data isn’t there.
What did returns cost on creator-led TikTok Shop orders in October?
Unavailable. Returns are recorded by order, not by creator, so this can’t be answered yet. Returns on Canvas field jacket orders on TikTok Shop cost $4,400 in October.
Returns · definition v3 · as of Nov 23 · status: unavailable
A question the data can’t answer yet. The answer says so, and gives the part that is known.
Sound familiar?
- Two people ask the same question and get two numbers.
- A change to the model waits in a queue for weeks.
- The chatbot answers with confidence, and nobody can tell whether it’s right.
A confident wrong answer travels faster than a late right one.
What changes.
- Definitions
- Buried in each analyst’s queryAgreed once, versioned and reused
- Answers
- A number with no contextA number with its definition, as-of date and status
- Missing data
- A confident zero“Unavailable,” and what would be needed
- Questions
- A queue for the data teamSelf-serve on supported questions, a queue for the rest
How it works. 5 steps, 4 stages.
- AI2
- Rules and automation1
- Your team2
- Approval
1Define
Your teamApproval
Agree the definitions
Finance and data approve each measure, its version and its owner.
2Ask
AI
Plan the question
Turns a plain-language question into a query on approved measures, and asks when it’s ambiguous.
- Power BI
- Metabase
3Answer
Rules and automation
Calculate on governed measures
Queries run on the agreed definitions. Missing data returns “unavailable”, never zero.
- Snowflake
- dbt
AI
Draft the explanation
Says what the numbers show and what they don’t, without inventing a cause.
4Review
Your team
Review the misses
Wrong, partial and unsupported answers are reviewed, then become new tests or new measures.
Every miss becomes a test
What you get. Real work, every week.
TikTok Shop contribution was 12% in October.
- Definition
- v3 · after fulfillment, fees, returns and commission
- As of
- November 23
- Status
- Closed, October
- Evidence
- October true-ups
- Owner
- Finance lead
Every answer carries its record: the definition, the date, the status and who owns it.
- Ambiguous period: asks which month
- Restated actuals: uses the closed figure
- Missing attribute: returns “unavailable”
- Plausible but wrong joinNew test added
The tests every answer passes first, including the questions it should refuse.
How you’ll know it’s working.
Accepted answers
Answers that reviewers agree with, out of the questions asked.
Questions answered as unavailable
A healthy count shows the system says so when it doesn’t know.
Reviewer corrections
How often a person changes an answer, and why.
Baselines agreed before we build.
Works with what you have.
Planning & BI
- Power BI
- Snowflake
- Metabase
- dbt
Commerce
- NetSuite
- We configure first
- Your BI tool’s own copilot and your existing metrics. If your measures are already defined there, we test the questions you really ask against them.
- We build when
- Definitions live in people’s queries, and questions that need finance’s judgment wait for days.
What a first build includes.
Built in your accounts, around your team. Scope, fee and baseline agreed before we start.
- 1Design
- 2Build
- 3Run it, or hand it over
We build
- A set of agreed measures, with versions and owners
- A test set of real questions, including unanswerable ones
- Answers that carry definition, as-of date and status
- A review loop for misses
We start from
- The questions your team really asks
- Existing metric definitions and reports
- Read access to the governed data
From your team
- Owner
- Finance lead, with the data lead
- Involves
- Finance
- Data
- Business owners
Questions teams ask.
Why not let the model do the math?
Models are unreliable at arithmetic across many records. In a 2026 research benchmark (FinBalance), the best model got 46.5% of balances exactly right. So queries and code do the calculating, and AI plans the question and explains the result.
Does governing the measures really help?
It did in one test. In an experiment run by dbt, answers were 98.2% accurate on governed measures against 90.0% without them, on a small set of questions. It’s a vendor’s own test, so we run yours on your data.
Who owns it once it’s running?
You do. It’s built in your accounts, documented, with training for the people who run it. Ownership usually sits with your finance lead, with the data lead. We can keep running it with you, or step back.
How do we start?
Usually with a finance review: two to four weeks, a fixed fee credited to the build, and a plan that’s yours. If the job is already clear, we can scope the build directly.
Part of your finance engine. One of 11 use cases, in 5 parts.
See the whole finance engineIn reporting & decisions
Get answers you can trace. Start with a finance review.
We reply within one business day.