REQVERA
The Impossible Answer · Decision note 02

How to answer an “always accurate AI” question

Illustrative situation · Not a customer case
The buyer’s question

Can you ensure AI-generated outputs are always accurate?

Do not guarantee every generated output. Define the task, show how performance was evaluated and name the review required before consequential use. Source citations and human oversight are controls, not proof of infallibility.

A response to adapt

No. We do not guarantee that every AI-generated output is correct or complete. For [specific feature and intended task], outputs are used as [drafts, suggestions or other verified role], within [documented input and use boundaries]. Before [specified consequential action], [named qualified role] must perform [actual review process] using [authoritative source or evidence]. For [system/model version], we can provide [approved evaluation report] describing the test set, acceptance criteria, results, known failure modes and limitations. [Verified controls] address [identified risks]; they do not eliminate all errors. Uses outside [approved scope] are [actual restriction or escalation rule]. If requirement [ID] requires error-free outputs without that review, we cannot confirm compliance as written.

Replace every bracketed field with verified facts. Remove any optional sentence you cannot substantiate. Do not submit this wording unchanged.

A buyer-ready accuracy evidence card

Answer these questions before sharing a performance claim. Keep results descriptive of the evaluated use, not a promise about all future outputs.

The unit being checked
State whether you assess a field, a factual claim, an entire response or a completed task. Do not present claim-level accuracy as the probability that a long answer contains no error.
The reference answer
Identify who established the expected result and from which authoritative material. A model grading another model's answer alone is not a verified reference for a contractual fact.
The important failures
Separate invented facts, missing requirements, wrong attribution and unsupported commitments. Record critical errors independently rather than hiding them inside an average score.
The missing-information test
Include cases where the source contains no answer. Check whether the system leaves the point unresolved or fills it plausibly. A citation should support the claim, not merely link to a related document.
The release boundary
Record the reviewer and the action held until approval. Recheck when the model, retrieval sources, language, instructions or use case changes; do not carry an old approval into a different configuration.

What makes the wording defensible

Replace an indefensible accuracy guarantee with a scoped account of testing, known limits and the human decisions that cannot be delegated.

  • Agree what accuracy means for this task

    Correctly extracting a date, producing a complete answer and giving sound advice are different outcomes. Define the task and error categories before offering a score. An answer can be factually correct yet omit an exception, attach the wrong evidence or fail the buyer's actual requirement.

  • Bound a result to the tested system

    Identify the model or system version, configuration, source collection, language and representative inputs. Explain sample selection and how errors were judged. A successful demonstration or a vendor benchmark does not establish your application's performance on this buyer's documents. Do not invent a percentage where no evaluation exists.

  • Describe a review people can actually perform

    Name the reviewer, the original evidence they can access, the release condition and the escalation path. “Human in the loop” is incomplete if nobody has time, authority or source access to challenge an answer. For actions with material consequences, identify what remains blocked until the required review is complete.

The evidence to obtain

  • An evaluation the buyer can interpret

    Obtain a dated report with the system version, test population, scoring method, error breakdown and limitations. Separate product-owner evidence from an independent assessment; disclose which it is. Do not call risk-framework alignment a certification of accuracy.

  • Traceable source and review records

    For representative outputs, retain the supporting passages and documented reviewer decisions. Show a correction and an unresolved case as well as a good answer. Protect customer information when sharing examples.

  • An operating rule, not a disclaimer alone

    Confirm the actual prohibited uses, permissions, escalation process and response to a reported error. Only describe abstention, validation or blocking controls that the offered implementation genuinely has.

The decision to approve

Decision owners
The AI or product owner substantiates the evaluation; the domain expert determines whether review is adequate for the use. Legal approves any contractual assurance, with the Bid Manager preserving its scope.
Proceed
Proceed when a truthful scoped answer is permitted, the evidence supports its claims and the buyer accepts the actual residual-risk and review arrangement.
Do not proceed
Do not mark an unconditional accuracy guarantee as met. Do not promise a reviewer or blocking mechanism that is not staffed or implemented. A mandatory zero-error condition remains a gap unless formally changed.
Escalate
Ask which outputs and decisions the assurance covers, the required evidence of performance and whether a reviewed-output workflow is acceptable. Follow the buyer's authorized clarification process.

Keep the decision in the final files

An approved AI-assisted draft can still change before submission. Reconcile the final claims with the buyer's current requirements and approved evidence, including limitations that editing may remove. REQVERA provides a final-control context; this editorial page does not evaluate a customer's model or certify its outputs.

See the authentic final-control workflow

Scope & primary references

Illustrative buyer question, not a measured customer outcome. General response practice for generative-AI assurances. NIST guidance is a voluntary risk-management reference, not a legal determination or a guarantee; use-specific regulation and liability require qualified review.