ZynoviqThe AI boundary
Our product is human judgment, checked and certified. If AI wrote or certified our answers, our tests would inherit the same blind spots as the systems we test. So the rule is simple: AI never sets, changes or certifies professional ground truth. People write, solve, review and certify every answer.
Where software helps — and where it never does
AI never sets, changes or certifies professional ground truth. People write, solve, review and certify every answer. Our software routes work and runs fixed, rule-based checks (formats, completeness, duplicates). Marking of your AI is rule-based and human today; AI-assisted marking of free text is planned, opt-in and always checked by people.
| Task | Who or what does it | Status |
|---|---|---|
| Writing scenarios and expected answers | Qualified professionals | Always people |
| Solving, reviewing, adjudicating and certifying answers | Qualified professionals | Always people |
| Setting, changing or certifying ground truth | Never AI, never software | Always people |
| Routing work to solvers, reviewers and adjudicators | Workflow software | Fixed rules, no AI |
| Format, completeness and duplicate checks | Rule-based checks | Fixed rules, no AI |
| Scheduling tasks and runs | Workflow software | Fixed rules, no AI |
| Marking structured fields of your AI's output | Rule-based marking engine | In use today |
| Marking free-text reasoning | Qualified professionals | In use today |
| AI-assisted marking of free text | AI, always checked by people | Planned — opt-in per customer |
What our evaluation work runs on
| We run | What it does |
|---|---|
| A web workbench | Where professionals receive tasks and write cases, answers and reviews |
| Rule-based checks | Formats, completeness, duplicates, signs of real data — fixed rules, no AI |
| A workflow engine | Routes cases to second solvers, reviewers and adjudicators |
| A secure vault | Stores certified answers and keys, encrypted and access-logged |
| A marking engine | Compares structured answers with the certified key, field by field |
| Reports and delivery | Report cards, data files, licences |
What this business does not run
- No GPUs
- No AI model training or fine-tuning
- No hosting of AI models
- No AI agents deployed for customers
- No AI grading of answers today — see the planned, opt-in option below
Where AI appears at all
Only your AI, on your systems, when we test it. For preference data, the candidate answers may come from a customer's own AI model, run through the customer's own account and budget, never a model we run. Preference data. The one planned exception is opt-in AI-assisted marking, described below.
Why this is a promise about this business
Zynoviq also has a separate Enterprise AI division that builds and deploys AI products. This business has separate delivery teams, no one works on both, and we never test, grade or produce data for the division's products. How we stay separate · Conflict-of-interest policy.
Planned: AI-assisted marking — opt-in and human-checked
Today, marking is rule-based for structured fields (exact match for yes/no answers and codes, a ₹1 tolerance for amounts, the key legal section required) and qualified professionals mark free-text reasoning.
We plan to offer AI-assisted marking of free-text reasoning as an option. It is not available yet. When it is:
- Opt-in per customer. It is used only if you choose it; otherwise marking stays rule-based and human.
- Calibrated against human markers. Professionals review every case where the AI marker and a human disagree, plus a sample of the cases where they agree.
- Agreement reported. The agreement rate between AI-assisted and human marking is stated in your report.
- Never the source of truth. It only compares your AI's output with an answer people have already certified. It never writes, changes or certifies an answer.
The core promise does not change: AI never sets, changes or certifies ground truth, and this business owns no GPUs. Any other use of AI, for example clerical help such as checking a case's formatting, would be disclosed in the provenance of the records it touched. How we keep contributors from pasting AI answers: human-certified means a human did it.
The AI boundary: questions people ask
Does Zynoviq own GPUs or AI models for this work?+
No. This business owns no GPUs and trains, hosts or deploys no AI.
Doesn't Zynoviq build AI products?+
Zynoviq's separate Enterprise AI division does. This business has separate delivery teams, no one works on both, and we never test, grade or produce data for that division's products. How we stay separate.
Will you ever use AI in your process?+
AI never sets, changes or certifies professional ground truth. Marking of your AI is rule-based and human today; AI-assisted marking of free text is planned, opt-in and always checked by people. Any other use, such as clerical help with formatting, would be disclosed.