From first call to report, in your test environment.
Every test follows the same six steps and one of four ways of connecting to your AI. Whichever you choose, three things never change: tests run in a sandbox or test environment, every case is synthetic, and our answer key never leaves our systems. This page is also written for the person who has to approve the engagement.

Six steps
- Scoping call — 30 minutes
We learn what your AI does, which decisions it makes on its own, and where a wrong answer would cost you most. You leave knowing whether a test is worth doing; if it is not, we say so.
- Test plan — a few days
We propose the domain, topics, number of cases, difficulty mix and answer format. Example: "500 GST cases on input tax credit, blocked credits, reverse charge, place of supply and time limits, in SAP document format." Fixed price, quoted; fixed deliverables; agreed before work starts.
- Test-environment connection — about a week
You give us a way to send questions to your AI in a test or sandbox environment, using one of the four options below. We also map your tax codes and account names to ours, so marking is fair to your set-up.
- Test run — about a week
Your AI answers every case. Answers come back as fixed fields — eligible yes or no, amount, tax code, account, legal basis — so each field can be marked exactly.
- Marking and review — one to two weeks
Each answer is marked against a certified answer key built by qualified professionals: at least two independent solvers, a reviewer, and a senior adjudicator where they disagreed, with every answer tied to the law. Genuinely ambiguous cases are reported separately and never counted as simple failures.
- Report and walkthrough
You receive the report and a walkthrough call: overall score, score by topic, high-severity errors, recurring failure patterns, and every failed case with the correct answer and the rule it turned on.
What your team provides
| Item | Effort |
|---|---|
| Scoping call attendance | 30 minutes |
| Test-environment access for the test period | One-time set-up |
| Tax code and account mapping | About an hour with a finance user |
| A named contact to confirm answers | Light, during the test run |
Four ways to connect. Pick what your security team prefers.
- 01Option A — We run the testYour AI agent connects to our testing service.
- 02Option B — You run it, we mark it (most common)We send you the questions only — never the answers.
- 03Option C — A sealed runner in your environmentA small, locked program in your own account.
- 04Option D — Full test pack licenceA copy of the pack, licensed under strict terms.
Option A — We run the test
Your AI agent connects to our testing service. We send each question, your agent answers, and we mark the answers. You receive the report.
- Your effort: a secure connection to your agent in a test environment.
- Best for: agents that can be reached over an API or an agent-to-agent connection, including SAP Joule agents where your landscape allows it. Testing Joule agents.
Option B — You run it, we mark it (most common)
We send you the questions only, never the answers. You run them through your AI inside your own systems and upload its answers to us securely. We mark them and return the report.
- Your effort: run a batch of questions and upload a file.
- Best for: organisations that do not allow outside systems to call their AI.
Option C — A sealed runner in your environment
For the strictest environments, we install a small, locked program in your own account. It holds the questions only, runs them through your AI locally, and sends just your AI's answers to us for marking.
- Your effort: a one-time installation, approved by your security team.
- Best for: regulated organisations with no outbound access to AI systems.
- For SAP customers: planned as an SAP BTP application from Q4 2027. The SAP BTP app.
Option D — Full test pack licence
Where a customer needs the full test pack in-house, for example to run regression tests in its own evaluation framework, we license a copy under strict terms: named users, no sharing, no use for AI training, deletion at the end of the licence, and a copy marked uniquely to you.
- Best for: software vendors running their own regression testing on every release.
- Priced higher, quoted, because it carries the most risk to the test's future fairness. Licences · Benchmark licence.
Side by side
| A. We run it | B. You run, we mark | C. Sealed runner | D. Licensed pack | |
|---|---|---|---|---|
| Connection to your AI | Yes, from our service | None | Local only | None |
| What you receive | Report | Questions and report | Report | Questions, answers, report |
| Answer key leaves us | No | No | No | Yes, under licence |
| Set-up effort | Low | Low | Medium | Low |
| Status | Live | Live (most common) | Live; SAP BTP application planned from Q4 2027 | Live |
Results in every option are delivered as a written report and structured files. There is no self-service results dashboard today.
The rule that protects everyone
Tests run only on a sandbox or test system, never production. A finance agent under test may try to post entries; in a sandbox, that harms nothing. This is written into every contract.
We don't guard your data carefully. We don't have it.
We write our own test cases. Every one is original and synthetic, never taken from a real taxpayer, company, employee or customer record, yours or anyone else's. We send cases to your AI and assess what it returns. A standard test needs none of your data at all.
What we need
- A way to send questions to your AI in a test or sandbox environment: one of the four options above.
- A mapping of your tax codes and account names to ours. Names only, no balances or transactions.
What we never access
- Your production systems
- Your customers', employees' or suppliers' records
- Your ledgers, invoices or filings
- Your AI's training data or model weights
How access works
Access is scoped, time-bound and revocable. It is granted for the engagement and ends at its close. Named individuals only: we tell you who, with their credential and role, and we tell you if the list changes.
If you issue us an access key, it is held in a managed secret store, used only for the agreed endpoint, never logged, and deleted at the close. You can revoke it at any time.
Who the reviewers are
For a service delivered by people, the real security question is the person looking at your system. Every reviewer is a qualified professional checked against their profession's own register, ICAI, ICSI or the relevant body, with credential number and expiry recorded. Every reviewer signs confidentiality and conflict-of-interest agreements, and sees only the cases assigned to them. How we qualify professionals.
Audit trail
Every case sent, every answer received, every judgment made and every reviewer involved is recorded and available to you on request. Each run is also logged by benchmark version and date, so the record can be read later by an auditor or buyer. Evidence.
Where our data lives
Our systems run in an Indian cloud region, with encrypted storage and backups in a second Indian region. Answer keys are held in a separate, isolated store that only our marking service can read. Data rights and confidential packs · Data processing.
When real data is involved
Some services, such as the expert review queue for live decisions, involve your real case data. These need a data processing agreement under India's DPDP Act, reviewers limited by name, and agreed retention. We tell you at scoping whether your request falls inside the standard, synthetic-only shape.
Separation from Zynoviq's Enterprise AI division
Zynoviq also runs an Enterprise AI division that builds AI products. The people who run and mark your tests are a separate delivery team from the division's staff. We never test, grade or produce data for the division's products, our reports never recommend them, and if your system includes one we tell you in writing before we test. How we stay separate · Conflict-of-interest policy.
Certifications — the honest position
Zynoviq's professional data and evaluation business does not currently hold SOC 2 or ISO 27001 certification. We say so plainly rather than implying otherwise. We are working towards ISO 27001 by December 2027, and we complete an external penetration test of our own systems before our first enterprise contract, and every year after that. If your policy requires a certification from every supplier regardless of the work, tell us on the first call. Trust.
How tests run — FAQ
How long does a test take from start to finish?+
About four to six weeks from the scoping call to the report for a capability assessment; about four weeks for an AI Quality Diagnostic; about two weeks for the free sample test.
How much effort does our team need to put in?+
A 30-minute scoping call, a one-time test-environment set-up, about an hour mapping tax codes with a finance user, and a named contact during the test run.
What answer format must our AI return?+
Fixed fields where it can — eligible yes or no, amount, tax code, account and legal basis — plus any explanation. If your AI only answers in plain language, qualified professionals mark it instead, at a different price, quoted.
What if our AI can only be used through a screen, not an API?+
Your team runs the questions through it and uploads the answers to us for marking: option B below.
Does any option need access to our production system?+
No. Every option runs in a sandbox or test environment, and this is written into every contract.
Can you complete our security questionnaire?+
Yes. Ask for our security pack; we answer standard questionnaires honestly, including where we do not yet hold a certification. Trust.
Zynoviq