ZynoviqKnow where your AI is right, wrong and confidently wrong.
We run your AI on held-out professional cases in your test environment and mark every answer against certified ground truth. You see each failure with the correct answer and the rule behind it, then we re-test whenever your system or the law changes.
Your platform already tests. It cannot tell you what correct is.
AI platforms and enterprise software increasingly provide native testing, evaluation and governance. They run cases, track versions and record metrics. A test is only as good as its expected answer, and for tax, payroll or company-law work that answer is a matter of professional judgment. Someone has to know what the AI should have done, and why.
| Your platform | Zynoviq |
|---|---|
| Runs test cases | Supplies the cases and the professionally correct answers |
| Technical pass or fail | Professional correctness, with the rule behind every answer |
| Tied to one vendor | Vendor-neutral: works with any model, agent platform or evaluation tool |
| Your own test cases | Held-out cases your AI has never seen |
| Static unless you update them | Refreshed when the law changes |
The two are complementary. Your platform tells you whether the system passed the test. We supply the cases and the professionally correct answers, so the test represents professional reality.
Evaluate once, re-test after change, try to break it.
What we mark
An AI can get the conclusion right and the reasoning wrong, or say the right thing and then take the wrong action. We mark each dimension separately so you can see which kind of failure you have.
| Dimension | The question we ask |
|---|---|
| Knowledge | Did the AI use correct, current information? |
| Reasoning | Did it follow the logic a professional would follow? |
| Decision | Did it reach the right conclusion? |
| Workflow | Did it follow the correct process, in the correct order? |
| Tool use | Did it choose the right tool, record or source? |
| Action | Did it take the correct action, or correctly decline to act? |
| Policy | Did it respect professional and organisational constraints? |
How answers are marked
| What | How | Status |
|---|---|---|
| Structured fields | Rule-based: exact match for codes and yes/no answers, ₹1 tolerance for amounts, the key legal section required | In use today |
| Free-text reasoning | Marked by qualified professionals | In use today |
| AI-assisted marking of free text | Calibrated against human markers; people review every disagreement and a sample of agreements | Planned — opt-in per customer |
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. The AI boundary.
What you receive
- Overall score and scores by capability, so you can see where the AI is strong and where it is weak
- Critical failures, each with its severity and failure category
- Expected versus actual for every case, side by side
- Expert explanation of why the expected answer is correct and where your AI went wrong
- Rule and authority: the section, standard or written policy behind each answer, with its valid-as-of date
- Regression cases: the failures to re-run after every change
- Next steps: what to fix first and what to re-test
Delivered as a written report plus a spreadsheet or JSON file and a regression set. There is no self-service portal. See a sample report.
How tests run
Every option runs against your AI in your test environment, never production, and every case is synthetic.
| Option | What happens | Who runs it |
|---|---|---|
| API or agent endpoint | Your AI connects to our testing service; we send each case and mark the answers | We run it |
| Batch file run | We send the questions only, never the answers; your team runs them and uploads the outputs | You run it, we mark it |
| Your own framework | You run a licensed pack in your own evaluation tools under strict licence terms | You run it |
| Runner inside your environment | A small, locked program in your own account holds the questions only and returns your AI's answers | Planned for SAP BTP |
How tests run in your environment, including security and data handling.
Ways to start
Every engagement begins with one workflow and one domain. A free 25-case sample test shows you the quality of our work. When you need more, the AI Quality Diagnostic tests about 100 cases on the workflow you choose, for one fixed price, quoted. From there: a capability assessment across a domain, continuous evaluation, a red-team engagement and the Regulatory Regression Service. Packages and how to start.
Testing agents, not just answers
When your AI acts, by proposing a tax treatment, posting an entry or routing a decision, we mark each step it takes against the certified answer: the judgment, the amount, the tax code, the posting and the legal basis.
We also check two things that matter once an agent works on its own: whether it hands genuinely hard cases to a person instead of answering confidently, and whether it gives the same answer when the same case is asked again. Testing Joule agents · Custom AI agents.
Who we test for
What we will not do
We will not build your AI, run it, fix it for you, or tell you it is certified. We test synthetic cases in test environments only, and we never test another country's law until qualified professionals are in place. We certify our answers, not your system. What we do not do.
Benchmarks & Evaluation — FAQ
Does Zynoviq replace our platform's testing?+
No. Your platform runs the test and tracks the metrics. We supply what correct means: held-out professional cases, certified answers and the rule behind each one, and we can run the evaluation for you where that helps.
Which domains can you evaluate today?+
GST and Indian indirect tax are live. Payroll under the labour codes and company law are active on request. TDS and income tax are next. Every domain and its status.
Do you need our data or production access?+
No. Every case is synthetic and every test runs in your test environment. We never touch production, your ledgers or your customers' records. How tests run.
Is the report a certification of our AI?+
No. It is a set of findings against certified ground truth, with a versioned record of what was tested, against which benchmark version, when and with what result. It is evidence of testing, not a certification or assurance. Evidence.
Who marks the AI's answers?+
Structured fields are marked by fixed rules. Free-text reasoning is marked by qualified professionals, checked against their profession's register and calibrated before any paid work. No answer is written or certified by AI.
How do we start?+
With a free 25-case sample test, or a fixed-price AI Quality Diagnostic on one workflow. Both show you the quality of the work on your own AI before anything larger. Packages and how to start.