Modern Gravity is generally availableBlog ↗
Seen, but not chosenNational moving and storage companyAsk an AI a business question and you get a clean number. "Headcount at the end of September was 412." It reads the same whether it is right or wrong. That is the problem we set out to measure.
In our internal blind test on synthetic business data, accuracy rose from 43.8% for Claude alone to 98.1% for Claude with Gravity, and wrong figures stated as fact fell from 46 answers to none, out of 105 scored answers each. That is evidence from this test, not a prediction for your company's data.
The problem: a wrong number that looks right
Real company data is messy. A customer record has a blank name. A month is only half over. An exit is entered three weeks late. Each one can change the answer without changing how sure the answer sounds.
A refusal leaves you with a question to resolve. A wrong figure stated as fact can make its way into a report before anyone checks it.
The test, in plain words
We built a synthetic company with 19 kinds of real-world data issues at known places. We wrote the correct answer to every question before any AI saw one.
Then we asked 43 business questions, three times each. The same model answered twice: once alone with read-only access to the data, and once with Gravity installed the way a customer installs it.
Every answer was stripped of anything that showed which setup wrote it. A program and a grader model from a different family graded every answer blind. A third grader, from another model family, then audited every answer. We opened the key that says which setup wrote which answer only after grading was done.
What we found
The rates below cover 105 scored answers to 35 of the 43 questions. The other 8 questions (3 that ask "why did it change?" and 5 multi-step problems) have no single right figure, so the figure grader does not score them.
A scored answer counts as right if it gives the right figure, or if it declines to give a figure where the answer key says no figure should be given. A refusal where a figure was expected does not count as right.
| Claude alone | Claude with Gravity | |
|---|---|---|
| Accuracy (right answer or honest refusal)* | 43.8% (46 of 105) | 98.1% (103 of 105) |
| Wrong figures stated as fact | 46 of 105 answers | 0 of 105 answers |
With Gravity, none of the 105 scored answers stated a wrong figure as fact.
We ran the same test on an open model, DeepSeek. Accuracy went from 39.0% alone (41 of 105) to 94.3% with Gravity (99 of 105). This result covers the two models and setups we tested.
An honest refusal tells you to investigate. A confident wrong answer gives you no such warning.
The fine print
We want you to read this part too.
- Accuracy counts an honest refusal as a right outcome when the answer key expected no figure, for example a measure nobody has defined or a month that is not over. Counting only right figures confirmed by both the program and the audit, it was 60.0% with Gravity (63 of 105) against 41.0% alone (43 of 105). So Gravity's 103 right outcomes are 63 confirmed right figures, 35 honest refusals, and 5 answers the audit graded right without saying which kind.
- 20 of Gravity's 35 honest refusals followed a small-number rule that we have since changed. The rates above describe the setup as tested, before that change.
- Gravity answers took longer (median 19.9 seconds against 16.3) and cost more to run.
- This was an internal test on synthetic data. It measures answer errors. It does not measure savings or how Gravity will do on your data.
- The paper also shows where Gravity was wrong.
Right answers are the minimum
At a minimum, an AI on your company's data has to answer most questions correctly, and say so when it cannot. That is the part this test measured.
The value comes next. Once your numbers are right and checked, they become the base for the rest of the company's work: planning, analysis, forecasts and decisions. Your plan, your board pack and your AI tools all start from the same agreed numbers, instead of each one working out its own version. That is how Gravity is meant to be used. This test measured the first part, answer errors, not those later uses.
What it means for a finance or HR leader
The number you take into the board meeting must be one you can defend. If your team uses AI for headcount, hiring, revenue or churn, the question to ask is not "is it fast?" It is "how would I know if it was wrong?"
Check one AI answer against your records
You do not need our test to see if your AI has the same problem. In your company-approved AI, pick data you are allowed to use and a question whose answer you can check. Paste this before your next question:
Before you answer my next question from our data, tell me: which date range you are using; whether any month in that range is incomplete; how you treat blank or missing values; whether any rows are dated in the future; which exact definition of the measure you are using, and who approved it; and how many records the number rests on. Then answer, and show me the records and the calculation so I can check them.
Treat the explanation as one more claim to check. Verify the source records, the calculation and the approved definition yourself. A believable explanation is not proof.
Check the answer against your records and the agreed definition, including any reason it gives for not giving a number. A refusal can be the right outcome when the data cannot support a number.
Read the method and the paper
*Right answer or honest refusal. Method and limits: methodology page.
Internal synthetic-data test, October 2026, 43 questions asked three times each. Grades finalized after a model-based audit of all 516 answers, which moved the headline in Gravity's favor; no person hand-graded a sample; no outside audit. Accuracy and wrong-figure rates use 105 scored answers to 35 of the 43 questions.