RESEARCHThe State of AI Search, Q3 2026: 110 B2B SaaS brands measured across four AI models ↗The State of AI Search, Q3 2026 ↗
Run my read
Modern Gravity™ · Proof

Need more proof?

Your AI may do better on your data. Seven outside tests show why it is worth checking: researchers at universities and AI companies keep finding that AI answers on company data are often wrong, and they look right.

01What other tests found

Seven outside tests. The same pattern.

Each test asks AI to do real business work on business data. Results are as each source reports them. The tests use different tasks and scoring, so read them as a pattern, not a ranking.

Real company databasesBEAVER, MIT with Harvard, TU Munich, Intel and AWS AI

The best AI systems wrote a correct query for about 11% of questions. With hints for every step, the best was 30.1%.

9,128 real questions from query logs, 812 tables, 19 business areas. Source

Business database workflowsSpider 2.0, ICLR 2025

The best system solved 21.3%, against 91.2% on the earlier, cleaner academic test.

632 real text-to-SQL workflow problems. Source

Company document searchEnterpriseRAG, August 2026

Only 26.8% of answers met every requirement at once. When the needed facts were missing, the best model declined correctly 42.7% of the time.

983 expert-checked samples across six business areas, 13 leading AI models. Source

Sales and service workCRMArena-Pro, Salesforce AI Research

Leading AI agents succeeded about 58% of the time in one turn, and about 35% over a longer conversation.

Sales, service and quoting tasks in a CRM setting. Source

Complex business tasksAgentArch

The best setup reached 35.3% on the complex task and 70.8% on the simpler one.

18 agent setups on business workflows. Source

Dense financial documentsOfficeQA, Databricks

The best agent reached 54.4% on the harder question set.

Questions that need figures found and combined across long financial records. Source

Financial analyst tasksFinance Agent v2.1, Vals AI, October 2026

The top-ranked AI agent scored 67.85%.

449 private questions on public company filings. Source

02Where GravityBench fits

GravityBench found the same problem.

On synthetic company data, Claude alone had 43.8% accuracy and stated a wrong figure as fact in 46 of 105 scored answers. With Gravity, accuracy was 98.1% and no scored answer stated a wrong figure as fact.

Accuracy means a right answer or an honest refusal. Internal double-blind test by Modern AI Labs, October 2026, 43 questions, synthetic data; grading audited by a third model; no outside audit. It does not describe accuracy on your business.

See our results and the fine print →See exactly how it was tested →Read the full paper →

03Why it is hard to see

"Our AI seems fine" is the usual first answer.

It looks rightA wrong number reads like a right one.

Tone, format and confidence do not show whether a figure is right. Only a check against a verified answer does.

Hard questionsFamiliar figures are not enough.

A useful check includes questions where definitions and reporting periods change the answer, alongside the familiar figures.

Fluent, not completeAn answer can meet most needs and still be wrong.

In EnterpriseRAG, models met up to 83.8% of individual requirements, but only 26.8% of answers met all of them.

No warningMissing facts often got an answer anyway.

When the facts were missing, the best model in EnterpriseRAG declined correctly less than half the time.

04What the research says helps

Test the whole system, not just the model.

The research points to what surrounds the model: agreed business meaning, good reading of the source, and tight control of access.

MeaningHints helped, but a big gap remained.

In BEAVER, hints for every step raised the best result from about 11% to 30.1%. Business context helped, and most questions were still wrong.

ReadingBetter reading of the source helps.

Databricks reported that better document preparation gave a 16% average gain across every agent setup it tested.

ControlRead-only access limits what an agent can change.

Anthropic advises limiting tool permissions, and notes that an agent with read-only database access can be used far more widely than one that can write. It does not make answers accurate or keep sensitive data private by itself.

Gravity is built for the business-meaning part. Your business certifies what each number means, answers are computed from your own data, and an answer the data cannot support is withheld, not guessed. Check both the right answers and how many questions get answered on your own questions before you rely on it.

05Test it yourself

The best proof is your own numbers.

Have your finance team choose 10 questions it answers every month, and write down the verified answers, definitions and periods first. Ask your AI the way your leaders would. Score fully right, incomplete and wrong answers and refusals separately, and flag each wrong figure stated as fact. Ask for our short buyer test kit: six live tests you can run on any AI tool, including ours.

Ask for the buyer test kitBring us 10 questions
06Sources

Read them yourself.

Every figure on this page was checked against its source in October 2026. Modern AI is not affiliated with these researchers, and they do not endorse Modern Gravity.

Three products. Structured Services. One Truth Layer. Discovery (live). Gravity (GA), with Agentic BI built on it. Friction (GA). Modern Foundation (Structured Services).