Question 1

Is DatumKeep an RCA tool, an SPC tool, or a knowledge base?

Accepted Answer

All three, on one core. Root cause analysis, CAPA and 8D, statistical process control, incidents and monitoring, and the knowledge base are modules over a single Problem → Cause → Solution store. That is the point: an SPC excursion can open an investigation that inherits its evidence, and the investigation’s conclusion becomes knowledge the next investigation retrieves. Buying them separately is what leaves you with five tools that don’t talk to each other.

Question 2

Which modules are live today, and which are roadmap?

Accepted Answer

Live: root cause analysis and CAPA, with 5-Why, Ishikawa and 8D supported inside them. Rolling out behind a per-tenant flag: the effectiveness learning loop, and scheduled SPC monitoring. Roadmap, on the same engine: FMEA, reliability and audit workflows. We would rather you knew which is which before the demo than after the contract.

Question 3

We already have a QMS and a CMMS. Does this replace them?

Accepted Answer

No. Your QMS holds the record of what was decided and your CMMS holds the work. DatumKeep is the layer that makes the reasoning reusable — it ingests what those systems already hold and turns closed cases into evidence the next analysis can retrieve. It sits alongside them, not instead of them.

Question 4

Who uses it inside the plant?

Accepted Answer

Quality and reliability engineers run the analyses; supervisors and reviewers sign them off; operators contribute observations without ever being named as a cause. Five roles ship — admin · supervisor · engineer · reviewer · operator.

Question 5

How is this different from asking a general AI assistant about our failures?

Accepted Answer

A general model will produce a confident answer from nothing. DatumKeep will not: every statistic comes from a deterministic engine the model cannot author, every retrieved case cites the record it came from, and a quality gate blocks the report when the evidence floor is not met. The difference isn’t the model — it’s everything around it that refuses.

Question 6

Does the AI invent root causes?

Accepted Answer

No. It drafts and proposes; your engineer confirms. Nothing enters the knowledge base without a person accepting it, and every proposed step cites the record it was drawn from. A quality gate scores the analysis across six dimensions and will not let a report through below the evidence floor.

Question 7

Can someone talk their way past the quality gate?

Accepted Answer

The evidence-floor block has no user-facing override — it is the one state with a single button and no way forward. Other gate outcomes offer a documented “proceed anyway” that is recorded on the session, so a reviewer can always see that it was used and why.

Question 8

What does it do when there is not enough data?

Accepted Answer

It says so, in the same shape as an answer. Capability is not reported below the sample the estimator needs, control limits are not drawn on four readings, and a retrieval below the relevance threshold is served flagged rather than padded with weak matches. There are eight documented states where the product declines to answer, and each one names its own threshold.

Question 9

Which AI models do you use?

Accepted Answer

Reasoning runs on Anthropic Claude, fast routing on a smaller Claude tier, and embeddings on OpenAI. The model identifiers are configuration, never written into application code — a continuous-integration gate fails the build if one is hardcoded, so changing tier is an environment change rather than a release.

Question 10

Will you tell us exactly how it works?

Accepted Answer

The specific models, data methods and scoring are proprietary — they are our edge, and keeping them ours is part of what protects your advantage too. What we are always transparent about is where the AI acts, where it is not allowed to, and every guarantee we make about it. Your security and compliance teams get the assurances they need, in writing.

Question 11

Does it assign blame to operators?

Accepted Answer

Never. No-blame is a design principle, and it is enforced twice: the analysis prompt forbids it, and a post-check rewrites any proposal that slips through. Analyses focus on systemic and process causes, not individuals. That’s what makes engineers trust it on the floor — and every conclusion remains fully traceable and export-ready for the audit room.

Question 12

How do you keep our knowledge separate from other customers?

Accepted Answer

Multi-tenant by construction, with Row-Level Security enforced at the data layer. Every connection sets a transaction-local tenant identifier before it can read anything, and the policies are enforced by the database rather than by application code that could forget one. Your company-private knowledge is never blended into another tenant’s, and never used to train anyone’s model. It is processed by our AI providers under a no-training commitment, and by no one else.

Question 13

Do the AI models train on our data?

Accepted Answer

No. DatumKeep runs on the commercial APIs of Anthropic and OpenAI, and both commit that data submitted through them is not used to train their models. You retain the rights to what you send and to what comes back. To be precise, because this is the question security reviews turn on: your knowledge is not used for training and is never blended into another customer’s knowledge base. It is transmitted to those providers to be processed — that is how any AI feature works — under a no-training commitment and a short retention window, after which the provider deletes it. Nothing is kept by them to build a model.

Question 14

Where is our data hosted?

Accepted Answer

In Europe. The application, the PostgreSQL database and file storage for documents and reports all run in EU regions. AI processing is the exception, and we state it plainly: prompts are sent to Anthropic and OpenAI, which may process them outside the EU under the standard contractual clauses that govern those transfers. No customer data is stored outside the EU.

Question 15

Do you have SOC 2 or ISO 27001?

Accepted Answer

Not yet, and we would rather say so than imply otherwise. Formal certification is ahead of us, not behind us. What is true today and verifiable in a technical review: tenant isolation enforced by Row-Level Security at the data layer rather than by application code, access tokens held in memory and never written to browser storage, EU data residency, and a no-training commitment from our AI providers. We are glad to walk your security team through any of it, and to complete a vendor questionnaire.

Question 16

How do you handle logins and sessions?

Accepted Answer

Access tokens live in memory only — never in localStorage or sessionStorage, so a cross-site script cannot read them. They expire after an hour and are refreshed against an HTTP-only cookie; logging out invalidates that cookie on the server rather than only in the browser.

Question 17

Can we get our data out?

Accepted Answer

Yes. Every analysis exports as an audit-ready PDF carrying its full traceability chain, and your incidents, cases and measurements are exportable as structured data. Your knowledge is an asset you own — it should be as portable on the way out as on the way in.

Question 18

Are the control charts real SPC, or a charting library?

Accepted Answer

Real SPC. All eight Nelson rules are evaluated on every chart, control limits and specification limits are drawn differently and never confused with each other, and limits are not drawn at all when the baseline is not established. Not established is not the same as zero, and the chart says which one it is.

Question 19

Why show a confidence interval instead of just Cpk?

Accepted Answer

Because a capability index from a small sample is a guess with a wide interval around it, and the decision should not be made on the optimistic end of that guess. Every capability figure ships with its interval, and the number a decision uses is the conservative bound — never the point estimate.

Question 20

How is the effectiveness score calculated?

Accepted Answer

From real-world evidence: after a given fix was applied, did the same failure come back, or did it hold? The scoring is deliberately conservative under small samples — a solution proven once is never treated like one proven a hundred times, and a solution with no applications yet shows an empty track rather than a zero.

Question 21

Do you check whether the measurement system itself is trustworthy?

Accepted Answer

Yes. Gage R&R runs as a first-class study, and the workspace flags when a conclusion is leaning on measurements whose own variation has not been validated — including the case where the study design is inadequate, which is reported as not assessable rather than as a pass or a failure.

Question 22

What does the monitoring actually watch?

Accepted Answer

Five checks run together on a cadence you set: Nelson rules against a locked baseline, CUSUM and EWMA drift, trend projection, capability assessment, and data staleness — the one that notices a parameter stopped reporting at all. Every run records an outcome, including the quiet ones, because “nothing found” and “never ran” are different facts.

Question 23

Do we need years of history before it is useful?

Accepted Answer

No. During onboarding we ingest what your organization already holds — past reports, resolved cases, documented fixes — so your very first analysis builds on your own experience rather than a blank slate. A curated industry foundation can supplement it, but your organization’s proven experience always takes precedence — value from analysis one, compounding return.

Question 24

What can you connect to?

Accepted Answer

Guided import for incident history and master data from the systems you already run — the CSV and Excel exports your ERP, CMMS or MES already produce, plus your quality documents — mapped to your equipment and process vocabulary as they land. Your equipment can also post measurements straight to a batch API. Native connectors that log in to those systems for you are on the roadmap; we would rather name what exists than what is planned.

Question 25

How long does onboarding take?

Accepted Answer

It depends on how deeply you integrate. A pilot typically runs about a month: we ingest a slice of your history, configure your equipment and process master data, map the platform vocabulary to the terms your plants use, and run real analyses against failures you already understand — the only honest way to judge the output. Deeper integration with your existing systems extends that; a pilot on exported history does not need it.

Question 26

Does it speak our terminology?

Accepted Answer

Yes. You map the platform vocabulary to the terms your plants actually use, per tenant — so the same equipment, process step or failure mode reads the way your engineers say it across every site and team.

Question 27

What does it cost, and can we try it?

Accepted Answer

Pricing follows usage and integration depth, so we quote it rather than list it — what drives the number is how many analyses you run, how much history you bring in, and how deeply you integrate with your existing systems. Evaluation happens as a guided pilot of about a month, with a capped number of analyses, so you can judge the output on your own failures before committing — there is no self-serve sign-up. Tell us your scope on a 30-minute call and we will give you a real number.

Questions, answered.

The product

Trusting the AI

Data & security

Statistics

Rollout

Commercial

See it on your own failures.