ZynoviqTell us what your AI needs to do.
You don't need to know which dataset, benchmark or evaluation framework you want. Describe the work your AI does and where a mistake would hurt. We map it to the professional capabilities involved and propose a mission.
What helps us most
A few plain sentences are enough. The more of these you can answer, the better our first reply will be:
- What your AI does — answers questions, drafts, classifies, decides, takes actions.
- Who uses it — your staff, your customers, other systems.
- Which professional workflow it touches — GST returns, invoice processing, payroll, purchase approvals, case escalation.
- What decisions it makes — and which of them it makes without a human.
- What systems it interacts with — SAP, Oracle, Salesforce, Tally, your own APIs and knowledge bases.
- What failure would be costly — a wrong tax position, an unauthorised payment, a missed escalation.
How we turn that into a mission
| You tell us | We work out |
|---|---|
| "Our agent checks supplier invoices before posting in SAP." | GST capabilities involved (input tax credit, reverse charge, e-invoicing), the actions to test (post, hold, escalate) and the professionals needed. |
| "Our copilot answers payroll questions for HR." | Labour-code and payroll capabilities, the policy documents that set the answer, and the questions where it must refer to a human. |
| "We are switching models next quarter." | A benchmark of your current behaviour and a regression set to run before and after the change. |
Missions can cover one or more layers of the work: expert ground truth, evaluation, regression and red team. Each Evaluation Mission runs one or more Data Missions underneath — the production step in which professionals create and certify the cases your AI is measured against.
Selling AI software? Ask about a benchmark licence covering all your customers. Software vendors and partners · License a benchmark.
What happens after you submit
- We reply within 2 working days
with questions, or a first view of what the mission could cover.
- Scoping call
we go through the workflow, the systems, the sources that define a correct answer and what you need to decide.
- Written mission specification
scenarios, capabilities, professionals, deliverables, environment, timeline and quote. You approve it before any work starts.
- Work starts
professionals are matched and calibrated, ground truth is created and adjudicated, and your AI is evaluated as agreed. How it works.
What a mission specification covers
Before any work starts you approve a written specification. It settles, in advance, every question that could otherwise turn into a dispute about what was delivered.
| Item | What it settles |
|---|---|
| Customer and AI system | Who the mission is for, and which model, agent or assistant is in scope |
| Domain and task type | The professional area and what the AI does in it — answer, classify, decide, act |
| Ground truth | What counts as correct: the law, standard or written policy that sets the expected outcome |
| Grading | How correctness is marked — rule-based for structured fields, professionals for free-text reasoning |
| Expert level | The capability and seniority required of the professionals |
| Review depth (QA) | How many independent solvers and reviewers, and how disagreements are adjudicated |
| Difficulty mix | The balance of routine, hard and edge cases |
| Evidence requirements | The rule, section or policy clause each answer must cite |
| Review criteria | What reviewers check before a case is accepted |
| Prohibited content | What cases must never contain — real personal or client data, for example |
| Acceptance criteria | What makes the delivery complete and acceptable to you |
| Deliverable | Cases, scores, failure report, regression set, formats |
| IP and ownership | Who owns the cases and results, and on what licence |
| Confidentiality and reuse | Whether anything may be reused; your information stays separate from reusable intelligence unless you explicitly authorise otherwise |
| Timeline | Milestones and delivery dates |
| Commercial model | Fixed, recurring, licence or usage-based — quoted against this specification |
Good to know
- Synthetic cases, test environments. We never need your data, and we never touch production.
- Human-certified answers. No expected answer is written or certified by AI.
- Findings, not certificates. Our reports are not a certification or assurance of your AI. See a sample report.
- Need certified training data instead? Request a dataset.
Evaluation Missions — FAQ
What is an Evaluation Mission?+
A defined piece of work in which qualified professionals set the ground truth for what your AI does, and that ground truth is used to evaluate, regression-test or red-team it. Each mission is agreed in a written specification before any work starts.
Do we need to know which benchmark or framework we want?+
No. Tell us what the AI does and what failure would be costly. Mapping that to scenarios, capabilities and professionals is our job.
How is a mission priced?+
Quoted per mission, against the approved specification — based on the number, type and difficulty of scenarios and the professionals required. Never by the hour or by headcount.
What if our domain is not live yet?+
Tell us anyway. GST and Indian indirect tax are live; payroll and company law are available on request; other domains depend on whether we can qualify the right professionals. We will tell you plainly what is possible and when. Domains and status.
Do we have to share data or system access?+
No data. Every case is synthetic. If the mission includes running your AI, it runs in your sandbox or test environment, never in production. How tests run.
Can we try something smaller first?+
Yes. If your AI handles Indian GST, the free 25-case sample test is the quickest way to see where it goes wrong.