A customer asks whether installation is included in a quoted price. Your team's AI assistant answers “yes” and links to the product guide. The guide exists, but that section applies to a different package. Someone checks that the link opens, then sends the answer. The business has now made a promise it may not be able to keep.
This is what human verifiability should mean in day-to-day work: can a colleague quickly find the right document and see whether it supports the exact answer? It matters if you are considering a staff knowledge assistant, automated email drafts or AI search through policies and product documents.
Researchers comparing 11,000 queries across five search systems in 2026 found that the systems differed in the sources they cited and how they expressed uncertainty. Another study of Google AI Overviews found that AI-generated documents were cited more often than human-authored ones in its sample, even after accounting for retrieval rank. These studies looked at public search, not a business's private documents. They are a reason to test how your own assistant selects and uses sources; a citation on screen does not do that test for you.
Check three different things
Before someone uses an answer that matters, make three checks:
- Identity: Does the link open the original document, and is it the approved version for this customer, product, location or date?
- Support: Can the reviewer point to the exact passage that supports the specific claim, including conditions and exceptions?
- Action: Is the answer safe to use for its intended purpose? A staff research note and a customer commitment carry different consequences.
In the installation example, a link to a guide is only a start. The colleague needs the current guide for this package, a passage that covers this configuration and the authority to promise that scope.
Design the assistant to show the relevant passage, document version or effective date, and any conflicting source. If it cannot find support, it should say so and send the question to a person. A plausible answer is not a substitute for a missing rule.
Test the questions that cause real work
A smooth demo tells us little about the awkward questions staff actually face. Start with 20–30 questions from real enquiries and searches. Include:
- A current rule with an exception, such as a return window that varies by product.
- A changed rule where the old and new service plans are both in the files.
- Two similar product names with different installation terms.
- A missing answer, such as a delivery promise that requires live capacity information.
- A restricted record that this member of staff should not be able to see.
Mark right source, claim supported, conditions preserved and safe next step separately. A relevant source can still lead to a wrong conclusion. Count corrections and time spent checking, not just time saved generating answers. If checking takes longer than the search it replaced, the pilot has not yet shown its value.
Repeat some questions after a policy update. Does the assistant still quote the old rule? A named owner for each document collection and a clear version rule may do more for reliability than changing the model.
Put the check where the risk sits
An outgoing quote, refund decision or statement about someone's account needs the exact claim and evidence visible before the message goes out or a record changes. Keep the original source one click away so a busy colleague can check it.
We would start with one document collection and one task. The result to look for is faster checked answers and fewer incorrect commitments. If the source files contradict each other, fix those first. A more fluent assistant will not resolve a rule the business has not settled.
Test one workflow before scaling it.
Estimate the time it could release, then check the effort, risks and evidence needed for a pilot.