ZynoviqHow AI fails at professional work
AI rarely fails professional work by producing nonsense. It fails by producing something that looks right. This is the library of failure patterns we design cases to catch, each with a short synthetic example. It is a map of what to test for, not a measure of how often each failure occurs.
The patterns
| Pattern | What it looks like | How we test for it |
|---|---|---|
| Fluent but wrong | A clear, well-structured payroll answer that computes a deduction on the wrong wage base. | Every answer is marked against a certified answer key, not judged on how convincing it reads |
| Missed proviso or exception | Asked whether a factory can claim GST credit on a canteen it is legally required to run, the AI says food is a blocked credit — and misses that the law allows the credit when the employer is required to provide the service. | Cases built around the exceptions practitioners know, not just the general rule |
| Outdated rule | Answers a payroll question using rules the labour codes have since replaced. | Every certified answer carries a "valid as of" date and is re-reviewed when the rule changes |
| Wrong rule for the facts | Treats an inter-state supply as intra-state and splits the tax into CGST and SGST instead of charging IGST. | Cases where small changes in the facts change which rule applies |
| Invented authority or citation | Supports its answer with a circular number or case name that does not exist. | Professionals check every section, rule and circular the AI cites |
| Overconfident instead of escalating | Gives a definite answer on a classification question where qualified professionals genuinely disagree, instead of flagging the doubt and referring it to a person. | Genuinely ambiguous cases where the correct behaviour is to hand the case to a person |
| Inconsistent answers | Asked the same case twice, says the credit is allowed the first time and blocked the second. | The same case is run more than once, and differences are reported |
| Wrong tool or record | An agent reads the supplier's details from the wrong master record, then reasons correctly from the wrong facts. | Tool calls and steps are marked where the integration exposes them, not just the final answer |
| Premature action | An agent books an input tax credit before checking that the conditions for it are met — for example, that the goods have been received. | Agent scenarios where the required checks come before the action |
| Quiet regression after a change | After a prompt or model change, the overall score holds steady while reverse-charge cases that used to pass now fail. | The same regression set is re-run after every change, and new and recovered failures are compared |
When the AI acts, more can go wrong
Agents add failure modes that an answer-only assistant does not have. We test for these where the mission includes actions:
| Failure mode | Synthetic example |
|---|---|
| Loops | An agent repeats the same lookup or step without reaching a decision. |
| Incorrect parameters | The right tool is called with the wrong period, amount or tax code. |
| Changing records it should not | An agent edits a master record or posted entry outside its remit. |
| Unsafe execution | An agent completes an action that should have needed a person's approval. |
Why patterns matter more than single mistakes
One wrong answer is an anecdote. A pattern tells your team what to fix: a missing exception in the knowledge base, a retrieval step that pulls the wrong record, a prompt that discourages escalation. Every failure we report comes with the correct answer, the rule behind it and its severity — and becomes a regression case so the same pattern is caught if it returns.
Professionals also design cases to provoke these patterns on purpose. That is our professional red team.
Related research
AI failure patterns — FAQ
Are these failures from real customer systems?+
No. Every example on this page is synthetic, written to illustrate a pattern. We never publish or describe results from a customer's system.
How common is each pattern?+
We do not publish prevalence figures here. How often a pattern appears depends on the AI system, the domain and the workflow — which is why we test each system on its own cases. Our published findings on public models are in the India GST AI Report.
Do you test for all of these in every engagement?+
Each mission specification sets which patterns matter for your AI. An answer-only assistant does not need agent-action tests; an agent that posts entries does.
Is this security testing?+
No. These are professional failures — wrong judgments, wrong rules, wrong actions. Data leakage, prompt-injection and infrastructure testing need a qualified security firm. AI red team.