Answer
What should an AI system refuse to infer?
Causation from correlation, a rule from one case, a missing value from its neighbours, and conclusions about named parties.
Causation from co-occurrence, a general rule from one observation, a missing value estimated from its neighbours, and any conclusion about a named person or business that the data does not directly support.
These four are not arbitrary. Each produces an output that is more useful than the truth, which is what makes them attractive and what makes a general instruction about care insufficient. A system that declines to infer causation produces a less satisfying answer than one that does, and the less satisfying answer is the correct one.
Causation is the first and the most common. Two things move together, or one precedes the other, and the natural sentence says one caused the other. The available evidence supports co-occurrence and the claim asserts a mechanism, and the gap between them is where most bad business decisions with data attached come from. The permitted form states what was observed and, where relevant, that a causal reading would require something the data does not contain.
Generalisation from a single case is the second. One customer behaved this way, one run failed like this, one region shows this pattern: each of these is an observation with a sample of one, and the sentence that turns it into a rule is unsupported regardless of how plausible the rule is. This matters most where the sample is small and invisible, which is the normal condition for anything about a specific business.
Filling a missing value is the third and the most insidious because the result is a number. Where a field is empty, a plausible value can be derived from the surrounding ones, and once derived it is indistinguishable from a measured one. Any system that does this must mark the value as estimated, and the safer default is to leave it absent and report the absence, because a gap is information and a filled gap is not.
Conclusions about identified parties are the fourth and carry consequences beyond accuracy. Inferring that a named person is dissatisfied, that a named business is in difficulty, that a customer is likely to leave: these move from analysis to allegation the moment they attach to someone identifiable, and they are frequently wrong in ways the subject cannot correct because they never see them. The restriction is not only prudence about error; it is about what a business should be willing to assert about people.
The way to enforce these is structural rather than instructional. A schema that requires an evidence field cannot express causation without something to point at. A validation that rejects estimated values in a measured field prevents the third. A rule that any statement about a named party must quote a record makes the fourth checkable. Instructions to be careful do none of this, because the inference feels like reasoning at the moment it is made.
The inferences worth prohibiting are the ones that produce the most useful-sounding sentences, which is exactly why they need a rule rather than a preference.
Siddharth Sharma, Context Theory
Related questions
Is a system that refuses to infer anything useful?
No, and that is not what these four prohibit. Interpretation is the point of most analysis. What is prohibited is a specific set of moves that produce claims stronger than the evidence supports, and everything else — stating what a pattern implies operationally, noting that two findings are consistent, identifying what would need to be true — remains available and is where the value sits.
What about an inference the user explicitly asks for?
Then supply it labelled as what it is. Asked what caused this, the useful answer names the candidates, says what evidence would distinguish them, and states which the data supports. That is more useful than either a refusal or a confident single cause, and it is the shape that survives being wrong.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
2026 speed-to-lead benchmark · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Each of the four prohibited inferences produces a more useful-sounding output than the supported alternative, which is why a general instruction about care is insufficient and a structural prohibition is required. | Comparing the supported and unsupported versions of the same finding for which reads as more useful. |
| Software | A value derived to fill a gap is indistinguishable from a measured one once written, so any estimated value must be marked as such and the safer default is to leave the field absent and report the absence. | Checking whether a dataset distinguishes measured values from derived ones at the field level. |
| Constraint | An inference attached to a named person or business becomes an allegation the subject cannot correct because they never see it, which makes the restriction a question of what a business will assert rather than only of accuracy. | Identifying whether the subject of any inferred conclusion has any route to see or dispute it. |
| Response | Structural enforcement works where instruction does not, because a schema requiring an evidence pointer cannot express an unsupported causal claim while an instruction competes with the appearance of reasoning at the moment of inference. | Attempting to record a causal claim in a schema that requires a supporting observation for each assertion. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one