Benner House
Menu

Twelve questions to see whether your data is ready to teach.

Answer twelve questions about the records your business keeps and how you would use them. You get a score for each of the four stages of AI data work, the gaps that matter most, and where to start. Nothing you enter leaves your browser.

A spirit level, bubble off center, rests on an uneven stack of archive boxes as a hand slides a wooden shim underneath.

Twelve questions, four stages

Answer for the AI use case you care about most. If you don't know an answer, choose the first option: not knowing is a finding too. It takes about five minutes, runs entirely in your browser, and sends nothing anywhere.

Prepare
1. Do you know where the records for your top AI use case actually live?

Systems, shared drives, inboxes, and vendor portals, each with a person who can answer questions about it.

2. How much of that data could a machine read as it is today?

Scanned PDFs, faxes, handwriting, tables saved as images, and call audio all need conversion before a model can use them.

3. How many of your records are copies, templates, or junk?

Forwarded email chains, form letters, drafts saved five times, and auto-generated notices teach a model to repeat boilerplate.

Protect
4. Do you know which of your sources contain PHI or other personal data?

Names, account numbers, birth dates, and addresses hide in free-text notes and attachments, not just in labeled fields.

5. If you redacted a dataset today, could you say what share of identifiers it missed?

Redaction recall is measured against a hand-labeled sample. Without that number, "redacted" is a hope.

6. Do you have a documented right to train models on these records?

Vendor contracts, customer terms, licensed content, and consent language can all restrict training use.

Teach
7. Do your records say what happened next?

A claim note that records the payout, a ticket that records the fix, a discharge summary linked to a readmission. Recorded outcomes are labels you don't have to pay for.

8. Could you hand a model 300 expert-verified examples of the task tomorrow?

Real inputs, each paired with the answer a senior person in your business would sign off on.

9. If two of your experts graded the same answer, would they agree?

Without a written rubric, experienced people often disagree, and a model trained on their labels learns the inconsistency.

Prove & Run
10. Do you have a test set that would show whether a model got better?

A fixed set of real cases with agreed answers, kept out of training and run before and after every change.

11. Could you list every source and change behind a dataset you've used for AI?

Sources, licenses, transformations, and reviewers, in a form you could hand to a customer, an auditor, or your own counsel.

12. After launch, how will production failures become new training data?

Someone has to sample real traffic, label what went wrong, and feed it back on a schedule.

How scoring works

Each question has four answers worth 0 to 3 points. The three questions in each stage add up to a stage score out of 9, shown as a percentage on the chart. All twelve together give an overall score out of 36, also shown as a percentage, which places you in one of three bands:

  • 0–39%: Start with an inventory. The data may be valuable, but nobody can tell yet.
  • 40–69%: Ready for a focused pilot. Enough foundation to test one use case once its weakest stage is fixed.
  • 70–100%: Ready to build and measure. The foundations are in place; the risk shifts to measurement and upkeep.

Any stage that scores below 60% gets its own advice and the services that address it, even when the overall score is high. The four stages are weighted equally on purpose. A project with excellent labels and no test set fails as surely as one with a perfect test set and no labels.

Where the questions come from

The twelve questions are a short form of the nine dimensions we score in a paid AI Data Readiness Assessment, organized by the four stages of our approach. The difference is evidence. The scorecard records what you believe about your data; the assessment reads samples of the actual records to check.

Your answers stay in your browser

The scorecard runs entirely on your device. Nothing you select is sent to Benner House or anyone else, and nothing is stored once you close the page. If you choose to share your result, the share link carries your answers in the part of the address after the # sign, which browsers do not send to web servers. You can send that link to a colleague, or to us if you'd like an assessment to start from your answers. More detail is in our privacy policy.

What a score can't tell you

A self-assessment measures confidence, and confident teams are sometimes wrong about their own data. Treat a high score as a hypothesis to test on a sample of real records, and a low score as a list of questions worth answering before anyone trains anything.

Keep reading

Related services

From Insights

Bring us one messy data source and one question.

Tell us what the data is and what you want a model to do with it. A data engineer, not a salesperson, will reply within one business day with questions and a proposed first step.

Start a project Take the readiness scorecard