# Keep AI Answers Current When Products, Stock and Policies Change

A correct fact can become a wrong answer when it arrives late or stays in a cache. Set different rules for stable knowledge, effective policies and live state, then verify the whole path to the customer.

Source: https://dtcup.com/blogs/insights/keep-ai-support-answers-current-freshness/

DTCUP / FIELD NOTES / OPERATE

# Keep AI Answers Current When Products, Stock and Policies Change

A correct fact can become a wrong answer when it arrives late or stays in a cache. Set different rules for stable knowledge, effective policies and live state, then verify the whole path to the customer.

Published Oct 2, 202624 min read

 ![Jack Hu](https://dtcup.com/wp-content/plugins/dtcup-knowledge/assets/jack-hu-small.jpg)**Jack Hu**7 years of experience building and operating brand websites, specializing in website development, paid advertising and SEO. Built a brand website from launch to US$10 million in annual sales.



  On this page  ## The Short Answer

Classify the claim before choosing a refresh interval. Preserve source, observation and effective-time meanings; scope cached answers to their dependencies; verify withdrawals through the customer path; and define a useful fallback when current evidence is unavailable.

## Key takeaways

- Stable guides, effective policies, live state and promises need different controls.
- A recent synchronization can contain an old observation.
- Verify the final answer, including caches and ongoing conversations.
- Test source withdrawal as deliberately as new content ingestion.
- Technical rollback must not restore a withdrawn business promise.
- Safe fallback and completed customer resolution are different outcomes.

![Dark green and pale sage cards rotate on a brass display beside an hourglass and a tray of ivory cards.](https://dtcup.com/wp-content/uploads/2026/10/dtcup-c14-operating-guides-20261002-1024x576.webp)A customer asks whether an item can arrive before Friday. The assistant finds a clear product page, reads a shipping guide and delivers a confident answer. Every sentence comes from something the business wrote. The answer can still be wrong because the stock disappeared this morning, the shipping cutoff changed yesterday, or the guide describes a service that is temporarily unavailable.

This is a different failure from inventing a fact. The system is repeating a fact that used to be suitable. Adding more instructions to be accurate will not repair a delayed inventory feed or remove an obsolete policy from a search index. Freshness has to be designed into the path between the business event and the customer answer.

My starting question is not how frequently to refresh the whole knowledge base. It is which statements become harmful when they are out of date, and what the assistant should do when it cannot establish their current validity. That leads to a smaller, more useful set of operating rules than trying to make every document real time.

This guide uses a fictional homewares company, Tern House, to build those rules. Its times, quantities and service targets are illustrative design choices, not universal standards or measured DTCUP results. The aim is to show how a support leader and an implementation team can agree on what current means, verify changes and keep serving customers honestly during an outage.

## Classify the fact before choosing an update schedule

Some knowledge changes slowly. A care guide may remain valid for a particular product revision for years. Its main risk is applying it to the wrong revision, not being a few minutes old. A brand description can be approved content with a named owner and a scheduled review.

Other facts are rules with effective dates. A return window, shipping threshold or promotion condition can change at a precise boundary. A version can be newly published but not yet effective. An older version may still govern a purchase made before the change. A single latest document does not answer all of those questions.

A third group describes operational state. Sellable stock, a parcel event, an account balance or an order cancellation status can change while the customer is talking. Copying these values into a general article creates an extra place that must remain synchronized with the operating system. For many tasks, a bounded read from the authorized source is the more appropriate design.

A fourth group describes promises that combine several facts. Delivery by Friday depends on more than stock. It can require destination eligibility, order cutoff, handling time, carrier service and current disruptions. Even perfectly fresh inputs do not justify a promise unless the application has a valid method for combining them.

At Tern House, the support team therefore creates four labels: stable reference, effective policy, live state and derived promise. These are operational categories, not labels the customer must learn. They determine who owns the source, how the system obtains it and which response is allowed when evidence is missing.

The important consequence is that there is no single knowledge refresh frequency. A daily review might be excessive for an unchanged care instruction and inadequate for a flash-sale inventory claim. A source-specific freshness rule is easier to explain and test than a marketing statement that the assistant is always up to date.

[![Four kinds of support facts require different controls: stable reference needs version scope, policy needs effective dates, live state needs an authorized recent observation, and a derived promise needs all required inputs plus a valid decision rule.](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/C14-facts.png)](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/C14-facts.png)Refresh frequency follows the consequence of a stale statement, not the format of the source file.## Keep four clocks separate

The first clock is when the business fact became true. A warehouse released a receipt at nine in the morning. A policy became effective at midnight in the store’s stated timezone. Call this the effective or event time, with the appropriate meaning for that source.

The second clock is when the source recorded the event. A warehouse operator might enter a release later, or an external service may deliver a delayed event. The record can be newly received and already describe an older state. That difference matters when a system says it synchronized a moment ago.

The third clock is when the assistant’s data path processed the information. A connector fetched the record, an index accepted a revision, or a cache received a response. This is evidence about the pipeline. It is not automatically evidence that the underlying business data is current.

The fourth clock is when the customer answer was produced. The system should compare the evidence available at that moment with the rule for that type of fact. A previous turn’s answer might have met the rule then and fail it now.

Imagine that Tern’s connector runs at ten o’clock and imports inventory observed at eight. Reporting last sync at ten can reassure an operator while hiding two hours of observation age. A better internal display shows both source observation time and processing time, plus whether the source supplies enough information to interpret them.

If the source does not provide a trustworthy observation time, record that limitation. Do not invent a timestamp by copying the current server time into a field named stock checked. The current time can truthfully describe when the request was made, while the state age remains unknown.

Timezone belongs in the contract as well. A promotion ending at midnight must name which midnight. Store-local time, customer-local time and UTC can describe different boundaries. Keep an unambiguous stored instant where possible and explain the customer-facing local time when the policy requires it.

A date printed on a page can refer to editorial review, legal effectiveness or technical deployment. Those are not interchangeable. Ask the content owner to label the meaning explicitly so that neither a developer nor a language model has to guess which clock controls eligibility.

## Write a freshness contract that changes behavior

A useful contract has a source, scope, maximum acceptable age where relevant, invalidation events and an allowed fallback. It also has an owner. Without the fallback, an age limit merely creates an error condition that the assistant may try to talk around.

For Tern’s inventory status, the proposed contract says to read the authorized availability service for the exact item, selling location and channel. The service must report the status and the observation information required by the application. If that evidence does not meet the approved condition, the assistant cannot describe the item as currently available.

For a return policy, the contract says to select the approved revision applicable to the relevant market and order date. The central issue is not a short time-to-live. It is whether the revision is valid for this customer situation and whether a newer correction or withdrawal has changed its availability for use.

For a care guide, the contract says to use an approved product-revision match and stop using the entry if its owner withdraws it. A review deadline can trigger an internal review queue. It does not need to tell every customer that a stable instruction has become false simply because a calendar reminder is due.

For a delivery promise, the contract requires the supported promise service or an approved calculation with all necessary inputs. If only a standard transit range is known, the assistant can explain that range and its conditions. It cannot convert the range into a guaranteed arrival date.

Choose numerical age limits through a business discussion about change rate, consequence and source capability. A five-minute target is not automatically safe, and a daily target is not automatically careless. An unavailable item can sell out within seconds; a historical order policy may remain valid indefinitely for that order cohort.

Document whether the rule concerns observation age, pipeline delay, review age or effective-date validity. Those dimensions may need separate checks. Saying the knowledge must be fresh within an hour is too vague when one person means connector activity and another means the time of the underlying warehouse event.

Make the contract executable enough that two testers reach the same verdict. Given the same record, time and customer scope, they should agree whether the system can answer, must recheck, can provide a bounded explanation or must hand off. A contract that depends on the tester’s intuition has not yet resolved the design question.

## Measure the whole path from change to answer

Tern changes a product’s available status at nine o’clock. The change event reaches its connector at two minutes past, the retrieval layer is updated at four minutes past, and an old answer cache remains usable until eleven minutes past. A customer asking at six minutes past can still receive the old statement.

The content team sees a completed update and believes the change is done. The customer path tells a different story. In this fictional sequence, the effective exposure interval runs from nine o’clock until eleven minutes past, because the last relevant answer surface did not stop serving the old claim earlier.

Do not simply add every component’s configured delay to produce an alleged measurement. Some work runs in parallel, some begins only after an upstream event and some fails. Measure the actual timestamps along representative paths. Use a planned worst-case budget only when its assumptions and boundaries are explicit.

The acceptance test is a customer question, not just a successful connector job. Ask a known question before the change, apply the approved revision in a test environment, and ask again through the route customers use. Check the selected evidence, final wording, citations and any saved response that can be reused. Record the first verified compliant answer separately from the time the technical job finished; those timestamps may differ substantially.

Repeat the check for a continuing conversation. Some systems retain a previous answer or retrieved passage in conversation context. Updating the search index does not necessarily stop the next turn from relying on that old context. The application needs a way to revalidate time-sensitive claims rather than treating conversation history as current authority.

Also inspect suggested questions, agent drafts, summaries and exported conversation notes where those outputs influence subsequent service. They may have different reuse rules. You do not need to rewrite an accurate historical transcript, but you do need to prevent an old transcript statement from silently becoming a present-tense instruction.

Assign an owner to the end-to-end result. Separate teams can own the source, connector and assistant, but somebody must be accountable for verifying that an important withdrawal actually changes the customer answer. Otherwise each component can pass its own check while the combined service remains wrong.

[![A fictional change becomes effective at 09:00, reaches the connector at 09:02 and retrieval at 09:04, but an answer cache remains stale until 09:11, leaving an eleven-minute customer exposure interval.](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/C14-timeline.png)](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/C14-timeline.png)A successful import does not establish that every answer path has stopped using the old fact.## Combine stable knowledge and live data without blurring them

A helpful answer can use both kinds of evidence. The stable guide explains how an item is shipped; the live service reports whether the exact variant is available through the relevant channel. These inputs should remain distinguishable inside the application even when the final response reads naturally.

Suppose the guide says that in-stock orders normally enter processing on the next business day. The inventory call reports unavailable. The assistant should not retrieve the guide and assume its in-stock condition has been satisfied. A conditional instruction is not proof that the condition currently holds.

The reverse matters too. A positive stock observation does not establish that expedited delivery is available to the customer’s address. The answer can state the verified availability and then request the information needed for a supported shipping check. The assistant should not fill the gap with the fastest service described in a general document.

Design the output of each tool around what the customer task needs. A stock tool can return an approved availability category and evidence time rather than the entire warehouse ledger. Private order tools require the identity and permission checks described in [the order-access guide](https://dtcup.com/blogs/insights/ai-order-support-identity-permission-boundaries/). Recent information is not automatically information that the requester may receive.

For a price, include the market, currency and applicable customer context required by the source. A current number for the wrong market is not a fresh answer to the right question. Promotions may require a separate eligibility check; do not treat a product’s base price as the final payable amount.

Where the system cannot combine the inputs reliably, narrow the answer. It can explain what the guide says and identify which current condition remains unverified. That is more useful than either pretending to know the final outcome or refusing every part of an otherwise answerable question.

Be equally careful in the interface. A green current badge beside an entire answer can imply that every claim was checked live. If only one field was checked, the display should not lend that status to the surrounding general explanation. Internal evidence records can be granular even when the visible answer stays concise.

## Cache facts within their boundaries, not just similar questions

Caching can reduce latency and repeated source calls. The risk is reusing a valid answer outside the conditions that made it valid. Two customers asking the same words may belong to different markets, product revisions or permission contexts.

Begin with the unit being cached. A public care-guide passage is different from a customer-specific order result, and both are different from a complete generated answer that combines several sources. The wider the combined answer, the more dependencies must be considered before it is reused.

For public policy knowledge, useful cache boundaries can include source revision, market, language and the policy condition needed by the task. For private data, authorization and account context require explicit treatment. Do not put a customer-specific response in a shared question-and-answer cache because its wording resembles a common support request.

A time-to-live is one control. An invalidation event is another. When a policy owner withdraws an entry, waiting for an ordinary expiration timer may be unacceptable. The design should identify which derived passages, cached answers and active retrieval collections depend on that entry and make them ineligible for new answers.

Be precise when discussing web cache settings with developers. [MDN’s Cache-Control reference](https://developer.mozilla.org/en-US/docs/Web/HTTP/Reference/Headers/Cache-Control) distinguishes reuse that requires validation from responses that should not be stored. Those HTTP mechanisms are useful at the transport layer; they do not automatically enforce the business validity of every fact inside a generated answer.

For example, a response can be technically revalidated against an unchanged endpoint while the upstream supplier feed is delayed. Conversely, an unchanged historical policy can remain the correct policy for an old order. Business freshness needs its own evidence and scope checks rather than relying on a cache header as a complete solution.

When a source changes, consider whether a versioned dependency key is more reliable than trying to find every paraphrase of an old answer. A cached response that records the revisions it relied on can be rejected when any required revision is withdrawn. This is an implementation pattern to evaluate, not a promise that a particular platform already provides it.

Test cache isolation with realistic pairs of requests. Use the same question for different markets, different product variants and different authorized customers. A fast response is a failure if it belongs to the previous requester. Performance and correctness have to be assessed together.

## Treat removal as a first-class update

Teams often test whether new material appears and forget to test whether old material disappears. A withdrawn guide can remain in an index after its original page is deleted. A cached answer can remain after the indexed passage is removed. A conversation summary can retain the same instruction again.

Give every knowledge item a stable identity and every approved revision a clear relationship to the previous one. Deleting a file should not erase the ability to identify which indexed passages came from it. Renaming a source should not accidentally create a second permanent copy that remains eligible after the first is withdrawn.

Platform behavior must be checked rather than assumed. For example, [Microsoft’s Azure AI Search guidance](https://learn.microsoft.com/en-us/azure/search/search-how-to-index-azure-blob-changed-deleted) describes specific change and deletion-detection requirements for storage indexers. The useful operational lesson is to verify the deletion mechanism in the actual data path; a new-file import test does not establish that removed content will stop being retrieved.

At Tern House, withdrawal creates a record that the item must no longer support new answers, while preserving the audit evidence needed to understand earlier use. The implementation then removes or filters the relevant retrieval entries and invalidates dependent answer material according to its retention and access rules.

The test asks a question that previously selected the withdrawn source. Passing means the old claim is no longer offered as current guidance and the assistant follows the approved alternative. A test that merely confirms the file is absent from an administration folder is incomplete.

Keep deletion, expiration and supersession distinct. A superseded policy may still be needed to explain an older purchase. A withdrawn inaccurate instruction may need to be blocked for every new answer. An expired promotion may remain useful as historical context without being presented as an available offer.

These distinctions prevent a common overcorrection: deleting every old version and then leaving the assistant unable to explain the terms that applied at the time of an order. The objective is controlled eligibility, not maximum destruction of history.

If deletion propagation fails, surface it as a service issue with an owner and containment option. A visible failed removal is actionable. A green sync status that ignores removals gives the team confidence it has not earned.

## Design what happens when the source is late or unavailable

There are several different failure states: the source cannot be reached, the response is too old, the scope does not match, the response is incomplete or the source reports that the value is unknown. Returning a generic error for all of them makes both customer support and debugging harder.

For Tern’s stock question, a failed current check leads to a narrow statement that availability could not be verified, plus a working route to check or contact support. It does not lead to the last positive stock claim, unless the business has explicitly approved a bounded use of that observation and the answer communicates its limitation without implying a current promise.

For a stable care question, a temporary outage of an unrelated order service need not stop the answer. The guide may still be approved and applicable. Degraded mode should remove the unsupported capability rather than disabling every useful response.

For a policy question with conflicting validity information, the assistant can explain the parts that are established and escalate the disputed eligibility. It should not select the more generous policy just to keep the customer happy or the less generous policy just to reduce business exposure. The resolution rule belongs to the policy owner.

For an operation such as cancellation, a timeout is especially important. The application may not know whether the action completed. It should represent a pending or unknown result and check the authoritative operation state through a designed recovery path. Retrying blindly can create duplicate effects in systems that do not protect against them.

Provide useful customer wording for each supported failure class. A sentence such as I cannot confirm the current availability of this variant is more informative than something went wrong. Follow it with a genuine next step, not a made-up callback or an unmonitored email route.

Do not use a timestamp as permission to make an unsafe claim. Saying available as of yesterday may still mislead a customer who is asking whether to purchase now. The fallback must fit the decision, not merely disclose that the evidence is old in small print.

Define a recovery condition. The service should return to normal only when the required source checks succeed and affected stale material is no longer eligible. A connector becoming reachable again does not prove that it has caught up on all missed updates.

[![A freshness decision matrix separates valid evidence, stale or missing live data, a withdrawn source, and an uncertain action result, with different answer or recovery behavior for each.](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/C14-fallback.png)](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/C14-fallback.png)Degraded service should preserve useful, supported answers while refusing to invent current facts.## Plan a policy cutover without erasing legitimate history

Tern plans a new shipping threshold effective on the first day of a month at the stated store timezone. The content owner approves the new version in advance. Publishing it into the repository should not make it applicable to customers before the effective boundary.

The team tests a question just before the boundary and a question just after it. It also tests an older order whose conditions are governed by the previous policy. These are three different cases, not one latest-policy check repeated with different wording.

The knowledge record needs the applicability facts that make the distinction possible. Those might include market, sales channel, order date and the type of request. The assistant should ask for missing information only when it is necessary and appropriate for the task, using authorized order context when available.

A scheduled cutover can fail because of indexing delay, a timezone mistake or a cached response. Prepare the approved fallback before the boundary. If the new rule cannot be confirmed after it becomes effective, an old current-policy answer should not continue merely because the previous release is technically available.

Rollback has to respect business truth. Rolling software back to yesterday’s build can be reasonable. Reinstating yesterday’s withdrawn promise may not be. Keep application version and policy applicability separate so that a technical recovery does not quietly reverse a business decision.

Where a prior policy still applies to an old order, preserve that scoped use. Where an old statement was incorrect or explicitly withdrawn, block it according to the approved withdrawal decision. The team needs to know which kind of old version it is restoring before calling a rollback safe.

Record the cutover result in terms a support manager can understand: the new rule answered correctly for the new cohort, the old cohort retained its applicable explanation, and unsupported cases followed the approved route. This is stronger evidence than a deployment timestamp alone.

## Build tests around changes, not only static questions

A static test set can tell you that an assistant knows today’s policy. It cannot establish that the assistant stops using that policy tomorrow. Freshness tests need a sequence: known starting state, controlled change, expected transition and customer-facing verification.

Start with a small set of critical claims. Choose a live stock status, a scheduled policy change, a withdrawn instruction, a private order state and a conditional delivery explanation. For each, state which system owns the fact and what should happen if the update arrives late.

Test the same question before and after the change. Then test a paraphrase, a continuing conversation and a question that combines the changed fact with a stable guide. The goal is to reveal alternative paths that still retrieve or repeat the old claim.

Include an out-of-order event. If a newer status has already been accepted, an older delayed event should not overwrite it merely because it arrived last. The correct handling depends on source identifiers and sequencing guarantees, so test the actual integration contract instead of relying on arrival time alone.

Include a missed event followed by a reconciliation run. Event-driven updates can be fast when they arrive, but the team still needs a way to detect gaps. A periodic comparison with the authoritative source can be part of the design, with its cost and coverage clearly understood.

Include withdrawal while an answer is cached and withdrawal during a conversation. Also test a newly approved revision that is not effective yet. These cases catch two opposite problems: old content remaining active too long and future content becoming active too early.

Run the tests across scopes that matter to the business. A market-specific change should not contaminate another market. A product revision update should not overwrite instructions for a different model. A permission change should affect retrieval eligibility even if the document’s prose has not changed.

Capture evidence sufficient to reproduce a failure: question, scoped request, source revision, observation time, processing state and actual answer. Keep sensitive values out of general review material. A screenshot of the final answer alone rarely explains which freshness boundary failed.

[![A change-based test suite covers before and after a policy boundary, a cached withdrawal, a delayed older event, a missed update recovered by reconciliation, and a continuing conversation.](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/C14-tests.png)](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/C14-tests.png)A current answer is only one state in a sequence. Test the transition that makes the old answer ineligible.## Measure exposure and containment, not just sync success

A synchronization success rate is useful for operating a connector. It is not a complete customer-quality measure. A job may succeed while processing no relevant changes, or update the index while leaving a stale answer cache active.

Track the time from an important change to verified customer-path behavior. Separate detection, processing and answer-surface verification where practical. The breakdown tells the team whether to improve source delivery, ingestion, invalidation or the final response path.

Also measure the claims affected. Ten stale answers about a decorative product description and ten stale cancellation confirmations should not be treated as interchangeable incidents. Classify consequence and task, then use that information to prioritize containment and prevention.

Consider a fictional review of one hundred live-status attempts during a source disruption. Seventy meet the approved evidence conditions and receive supported answers. Twenty correctly decline to confirm current status and provide the approved next step. Ten incorrectly reuse an expired observation.

Calling ninety answers successful would hide the service degradation. The review should show seventy supported current answers, twenty contained but incomplete tasks and ten stale-answer failures. The twenty safe fallbacks matter, but they are not twenty resolved customer requests.

Now suppose all ten failures came from a saved-answer path while direct tool calls behaved correctly. Increasing the connector polling frequency would not address the observed cause. The intervention should target the reuse path, then rerun the change tests and inspect whether customer completion recovers.

Keep the denominator tied to eligible attempts, and distinguish attempts from unique customer issues. Repeated retries can inflate counts during an outage. A service recovery report should explain whether customers eventually got a verified answer or simply encountered the fallback several times.

Use a review window that includes normal operation and known changes. A quiet week with no important updates may look excellent without exercising the invalidation design. Scheduled change tests provide evidence that passive monitoring cannot supply on its own.

## Give freshness an owner and a practical operating rhythm

Content owners should know which entries support automated answers and how to withdraw them. An ordinary document edit should not be the only available emergency control. The process needs a way to stop a known bad claim before a full editorial rewrite is complete.

Integration owners should monitor source delays, failed updates, removal failures and scope mismatches. A useful alert identifies the affected source and capability. A generic pipeline failed message forces support staff to discover customer impact through complaints.

Support leads should review the customer wording and the actual recovery path. They can tell whether a fallback helps the customer proceed or merely sounds polite. They also need a clear route for reporting a stale answer with enough context for investigation.

Product owners should decide which capabilities are worth the operational complexity. If a live availability service cannot supply reliable evidence, launching a simpler assistant that explains product information may be the better first release. Broad promises create maintenance obligations long after the launch demonstration.

A weekly review can focus on a small set of items: overdue source reviews, important upcoming policy changes, propagation failures, stale-answer incidents and unresolved customer effects. Give each exception an owner and an action. Avoid turning the meeting into a tour of every connector chart.

Before a promotion or policy launch, run a specific readiness check. Confirm source ownership, effective boundary, test questions, invalidation route and fallback. After launch, verify the real customer path. This links the knowledge operation to business events instead of relying entirely on recurring calendar maintenance.

Keep a concise decision record for numerical targets. State why the age limit was selected, which tasks it protects and what evidence would lead the team to change it. Otherwise a temporary development setting can become an unexplained permanent service promise.

The long-term objective is not to eliminate every delay in every system. It is to know when the available evidence is sufficient for the claim being made, prevent unsupported current statements and restore useful service quickly when a dependency fails.

[![A freshness operating contract records fact type and scope, authoritative source, timestamp meanings, reuse and withdrawal rules, degraded response, responsible owner and customer-path verification.](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/C14-contract.png)](https://dtcup.com/wp-content/plugins/dtcup-operating-guides/images/C14-contract.png)The contract connects technical updates to the statement the customer is allowed to rely on.## A first-week implementation that is small enough to verify

On the first day, choose one claim that frequently causes customer confusion. Inventory availability is a useful candidate only if the team can access a trustworthy source. If it cannot, begin with an approved policy change whose effective boundary and scope are clear. Write the intended answer and the situations in which the assistant must not give it.

Next, draw the actual path on one page. Include the source, connector, retrieval layer, response cache and conversation state used by that claim. Ask each owner what records a successful update and what records a failed withdrawal. Gaps in those answers are implementation work, not reasons to invent another dashboard score.

On the following day, create a controlled change in a test environment and record the four clocks. Ask the ordinary customer question through each supported entry point. If the source changes correctly but the answer does not, keep the failure intact long enough to identify its cause before clearing every cache and losing the evidence.

Then rehearse an unavailable source. The reviewer should see the approved narrow response, a working next step and enough internal evidence to explain why the assistant withheld the current claim. Restore the source and verify that recovery includes catching up on missed changes.

Finish the week by documenting the verified path, remaining limitations and the next claim to cover. Do not generalize one passing inventory check into a claim that all knowledge is current. Expand the same discipline to policies, order state and other consequential facts, keeping each scope and acceptance condition visible.

## Questions about keeping AI support knowledge current

### Should every support fact come from a live API?

No. Stable approved explanations can be served from a maintained knowledge base with the correct scope and revision. Live state needs an appropriate current source. The design should match the fact and customer decision rather than making every answer depend on unnecessary network calls.

### Does updating a document immediately update the assistant?

Not necessarily. The source change may need to pass through a connector, index, cache and conversation context. Verify the final customer path, including withdrawal and continuing conversations, rather than treating the source edit or successful import as complete evidence.

### Can an assistant answer from old data if it shows a timestamp?

Only if that bounded use is appropriate for the task and approved by the business. A timestamp does not make an old availability claim suitable for a current purchase decision. Explain uncertainty and offer a valid next step when the current condition cannot be verified.

### Why keep old policy versions at all?

Some older versions remain applicable to historical orders or events. Preserve them with explicit eligibility rules rather than making every old version searchable as current guidance. Withdrawn errors and legitimate historical policies need different treatment.

### What should be improved first when stale answers appear?

Trace one failure from the business event to the final response and identify where the old claim remained eligible. The cause might be delayed source data, an index, a cache or conversation reuse. Fixing the observed boundary is more effective than increasing every refresh frequency without diagnosis.

For the broader design, connect this operating contract to [knowledge ownership and maintenance](https://dtcup.com/blogs/insights/build-maintainable-ai-support-knowledge-base/), [policy conflict resolution](https://dtcup.com/blogs/insights/resolve-conflicting-ai-support-policies/) and [customer-service evaluation](https://dtcup.com/blogs/insights/build-ai-customer-service-test-set/). Freshness is the link between an approved fact and the moment a customer is asked to trust it.



 ## Sources

1. [developer.mozilla.org — Cache-Control](https://developer.mozilla.org/en-US/docs/Web/HTTP/Reference/Headers/Cache-Control)
2. [learn.microsoft.com — search-how-to-index-azure-blob-changed-deleted](https://learn.microsoft.com/en-us/azure/search/search-how-to-index-azure-blob-changed-deleted)

CONTINUE / THE TOPIC

## Keep the next question connected.

[NEXT BEST READ**Build an AI Customer Support Service Your Team Can Actually Run**](https://dtcup.com/blogs/insights/ai-customer-support-operating-model-implementation/)[RELATED GUIDE**Build an AI Support Knowledge Base That Can Keep Its Promises**](https://dtcup.com/blogs/insights/build-maintainable-ai-support-knowledge-base/)[RELATED GUIDE**When Two Policies Disagree: Give AI a Source-of-Truth Rule**](https://dtcup.com/blogs/insights/resolve-conflicting-ai-support-policies/)[RELATED GUIDE**Why an Order Number Is Not Permission to Reveal an Order**](https://dtcup.com/blogs/insights/ai-order-support-identity-permission-boundaries/)

Still have a question?[Ask the DTCUP Community ↗](https://dtcup.com/community/)
