DATUMKEEP
AI Technology

AI runs through every part of it. So does the leash.

Most quality tools have an AI feature. Here it is the connective tissue: it reads your history into knowledge, guides the investigation, and asks the statistical engines the questions your engineers would ask. What makes that safe to buy is not the model — it is what the model is structurally prevented from doing at each step. This page is both halves.

Where it acts

Three surfaces, three different jobs — and three hard limits.

AI is not one feature on one screen here. It runs through the knowledge base, the investigation and the statistics — doing something different in each, and barred from something different in each.

Knowledge

It reads what nobody has time to read

  • Turns years of unstructured reports and closed cases into structured problem, cause and solution
  • Reconciles the different names your sites use for the same failure into one concept
  • Finds the prior cases that match a new incident, in context, before anyone searches
Where it stops

It cannot promote anything into trusted knowledge on its own. A person approves what becomes golden.

Investigation

It runs the method, you run the judgment

  • Drafts the causal chain step by step — 5-Why, Ishikawa, 8D — never from a blank page
  • Proposes hypotheses, and labels each one with where it came from: reasoning, a historical match, or your engineer
  • Flags when the same failure has happened before, and what was done about it
Where it stops

It cannot close an analysis. A quality gate checks the chain of evidence, and a person signs off.

Statistics

It asks the engines, it does not answer for them

  • Turns a plain-language question about your process into calls to the statistical engines
  • Sequences the right analyses and explains what came back, in words your team can use
  • Carries every honesty state through untouched — "not enough data" reaches you as "not enough data"
Where it stops

It cannot compute a number. Every figure comes from a deterministic engine the model has no way to author.

What it runs on

Whose models, and for what.

Not one model doing everything. Three jobs with different demands, matched to what each one actually needs — and a fourth that is deliberately not a model at all.

Reasoning

Anthropic Claude — frontier tier

The judgement calls: proposing and testing hypotheses, building the cause-and-effect structure, drafting the analysis, and reviewing the finished work against the evidence that supports it.

Fast routing

Anthropic Claude — smaller tier

The high-volume work where speed matters and the answer is bounded: pulling evidence out of what you wrote, drafting the routine grids, turning a statistical result into a sentence, and normalising imported data.

Retrieval

OpenAI embeddings

Finding what your plant already knows. Every incident, resolved case and document is indexed so a new failure can be matched against them by meaning rather than by keyword.

Statistics

No model — deterministic engines

Every number on the site comes from a statistical engine, not a model. The AI can decide which test to run and read the result back to you; it has never authored a figure and cannot.

Which exact tier serves which role is configuration, never written into application code — a build gate fails if a model identifier is hardcoded. That means we can move a role to a better model as one lands, without a release, and it means we can tell your technical reviewers precisely what is in use today.

How we differ

Everyone says “human in the loop.” Ask what happens when nobody is looking.

Cited answers and a review step are table stakes. These five are engineering decisions that hold when the room is empty — and each one is something you can put to us in a demo.

01

Your documents cannot give the AI orders

Usually: whatever is in the document goes straight into the prompt

Case files, incident text and imported reports are handled as data to be read, never as instructions to be followed — fenced off and labelled as untrusted before the model ever sees them, with invisible and direction-manipulating characters stripped out. A supplier report cannot contain a sentence that changes how your analysis behaves.

02

The model does not get to grade its own work

Usually: the same model that wrote the answer also rates it

A separate check — deterministic, with no AI in it at all — decides whether the analysis genuinely engaged the prior knowledge that was retrieved for it, or merely had it nearby. It can hold an analysis back for review. It cannot be talked round, because there is nothing in it to talk to.

03

Every proposed cause says where it came from

Usually: one undifferentiated list of suggestions

Each candidate cause is tagged with who originated it — the AI's reasoning, a match against your own history, or an engineer — and separately with how it was produced. So your team can weigh a model's idea differently from a fix that already worked on line 3, instead of guessing which is which.

04

When something degrades, it is on the record

Usually: a quiet fallback that looks identical to a good result

If a model is unavailable, returns something unusable, or a step cannot be computed, that is recorded as an event on the session and the output says so. The failure mode a buyer should fear is not an outage — it is an outage that still produces a confident-looking report.

05

No model has ever authored a number here

Usually: the model writes the statistics into the sentence

Control limits, capability, measurement studies and every test are computed by deterministic engines. The AI chooses which to run and explains what came back — including when what came back was a refusal. Ask it something the data cannot support and you get the refusal, not a plausible figure.

The how is our edge — the behaviour is transparent

You will notice this page tells you where the intelligence acts, what runs it and what it is barred from doing — but not how any of it is built. That is deliberate: the methods, the scoring and which exact tier serves which role are proprietary, and keeping them ours is part of what keeps your advantage yours. Everything we do claim is behaviour you can put to the test in a demo — the FAQ answers the rest, including that one.

What we don't do
We don't replace your engineers — judgment and sign-off stay with your team.
We don't invent root causes — every conclusion traces back to cited evidence.
We don't train on your data across customers — your knowledge stays yours.
We don't hide uncertainty — low-confidence findings are shown as low-confidence.

Private by construction

Your data is strictly isolated per organization and never used to benefit anyone else. Security is a first-class part of the design, not an afterthought — see Security for the details.

Root cause analysis · CAPA · 5-Why, Ishikawa and 8D live today — Scheduled monitoring routines · The effectiveness learning loop shipped · off by default — FMEA roadmap

Put it to work on your own failures.

The fastest way to judge the intelligence is on a failure you already understand. Bring one to a 30-minute demo and see what surfaces — including where it refuses.