How we use AI

Most of this isn't AI. Here's the part that is.

Every median, every completion rate, every severity, and the whole regulated report are arithmetic and rules. They give the same answer every time, and they run whether or not any model is installed. A language model writes prose in five places, and it is never allowed to state a number, a time, or a name of its own.

what decides what
Computed
Recognized
Generated

Roughly where the work happens. The narrow band is the only place a language model writes anything.

01 / The distinction

Three different things get called AI.

The useful question is not how modern the technology is. It is how much the software is deciding, and whether you can check the result. That gives three tiers, and they behave differently enough that a policy should treat them differently.

Tier 1

Computed

Arithmetic and rules we wrote. No model of any kind. Run it twice, get the same answer.

  • Time on task, medians, completion rates
  • Severity and outcome scoring, read from the vocabulary your study defines
  • Whether a duration is sound enough to be counted at all
  • The regulated human-factors report, end to end
  • The checks that tell you what the report is still missing
  • Search
Tier 2

Recognized

On-device recognition that turns one medium into another. It reads and transcribes. It does not conclude.

  • Spoken words in the recording become timed, searchable text
  • Text on the participant's screen becomes a search index, and nothing else
  • Runs entirely on the recording machine
  • Scores nothing, ranks nothing, decides nothing
Tier 3

Generated

A language model, running locally, writing prose or proposing a draft. Wording varies between runs.

  • A plain-language summary of a session
  • A stakeholder findings report, every claim cited
  • Answers to questions you ask about your own study
  • A first-draft study plan for you to edit
  • A review of report prose a person wrote

02 / The boundary

What the model is not allowed to do.

These are not instructions asking a model to behave. They are properties of how the software is built, and most of them exist because an earlier version got it wrong and we found out the hard way.

It cannot state a numberEvery figure is computed first and handed to the model as fixed text. It reports numbers; it never derives them.
It cannot state a timeEvery quote it cites is checked against what was actually recorded. A quote that cannot be matched is dropped rather than softened, and a matched quote is stamped with the real moment from the recording.
It cannot invent vocabularyWhere a finding is classified, the model picks from a fixed list of codes, and the software renders the words. It cannot coin a category.
It cannot change your studyWhat the model writes is stored beside your study, not inside it. It never creates an observation, a task timing, or a recorded error. The single exception is the draft study plan, which becomes yours only when a person reviews it and applies it deliberately.
It cannot reach the networkInference happens on your machine. An automated check runs on every build and fails it if unexpected outbound network code appears anywhere in the analysis path.

03 / If your policy says no

You can switch it off and keep almost everything.

This is not a workaround. The capabilities are licensed separately, so a configuration without any generative AI is something you can simply buy, and the software is built to be complete without it.

NO GENERATIVE AI

The language model is a separate installation. Leave it out and the software never asks for it, never mentions it, and never checks for it.

  • All recording, moderating and observing
  • Task timing, outcomes and every metric
  • Surveys and their scoring
  • The full regulated report and its completeness checks
  • Review, search, export

NO MACHINE LEARNING AT ALL

For a policy that covers recognition as well, the searchable layer is licensed separately again and can simply be left off.

  • Recording, logging, task timing and outcomes
  • Surveys and reporting
  • What you give up is the searchable layer: spoken words and on-screen text

Being straight about tier 2: recognizing speech and reading text on screen are machine learning, and a policy that says no AI may well cover them. We would rather say so than describe them as something else. What we will say is that they only ever transcribe, they never draw a conclusion, and nothing they produce leaves your machine.

Bring us your policy

We will map it to the feature list, line by line.

If your organization regulates or restricts AI, send us the policy. We will show you exactly which capabilities it touches, which it does not, and what a compliant configuration looks like. A detailed technical paper, down to the level your security reviewers will want, is available under NDA.