Your manager approved the quote on Friday. By Monday, the price, recipient or customer record has changed. The AI must not treat that old yes as permission for a different action.
A yes belongs to a proposal, not an entire conversation
Consider an illustrative UK maintenance business. Its AI assistant prepares a £1,200 equipment order and asks the operations manager to approve it. The manager reads the supplier, delivery date and amount, then gets pulled into another job. During the delay, the supplier revises its price to £1,450 and the customer cancels the installation. An approval arriving later should not release the revised order simply because the conversation still contains the word yes.
This is not primarily a problem of the AI misunderstanding language. It is a problem of permission becoming detached from the facts that justified it. The workflow may remember the approval perfectly while forgetting that the underlying situation is different. Human review has happened, yet the action no longer represents the reviewed decision.
Recent technical guidance recognises this gap. PrompTessor's 6 October guide explicitly describes revalidating current state after approval and gives a price-change example. Its point is practical: a suspended workflow needs a resume policy, not merely an approve button. This is practitioner guidance, not evidence that a particular percentage of businesses has suffered the failure.
Separate three moments in your own process: when the proposal was assembled, when the person reviewed it and when the external action will happen. They can be seconds apart or separated by a weekend. Any material change between those moments can make the original decision unsuitable, even when the same employee retains authority to approve a replacement.
What this means in practice is straightforward. Ask your supplier to demonstrate a quote changing while approval is pending. The acceptable outcome is not a confident explanation after sending. It is a blocked send, an explanation of what changed and a new decision on the revised proposal. If the product cannot show that distinction, its approval feature is incomplete for this workflow.
Human review can happen correctly and still authorise the wrong action if the proposal changes afterwards.
Define exactly what the reviewer is approving
An approval record needs enough detail to distinguish the proposed action from its neighbours. For a quote, that may mean the customer account, recipient address, line items, total including VAT, payment terms and attached document version. For a refund, it may mean the original transaction, amount, currency and destination. For an access change, it means the named account, permission and system, rather than a general instruction to help the employee.
Those details should come from the authoritative business systems and the actual outgoing payload. Do not ask the AI to summarise a proposed action and then quietly approve whatever different action it produces later. A polished review card is only useful if it describes the operation that the execution layer will actually submit.
QueryNow's guide, reviewed on 2 November, identifies Microsoft Agent Framework approval controls and LangGraph interrupts as named implementation examples. It also argues that evidence needs the exact request, reviewer identity and timestamp. Treat those product features as building blocks, not an automatic guarantee that your own integration binds the eventual action correctly.
A useful review screen shows both the commitment and its basis: send this version of this quote to this address, for this customer, under this policy. Where an action covers several records, show the included set. An approval for five listed customers should not expand to whichever customers happen to match a live filter when the job eventually runs.
There is room for bounded flexibility. You might authorise a purchase up to an agreed delivered price from a named supplier before a stated date. In that case the approved object is a clearly defined envelope, not a frozen amount. Record those tolerances explicitly and enforce them outside the model. Never infer a wider tolerance from the reviewer sounding relaxed in chat. A person approving one proposal and a person granting a bounded operating mandate are making different decisions.
Approve the outgoing commitment or an explicit envelope, never a vague intention to continue.
Use both an expiry time and change-sensitive checks
Time limits are helpful, but a timer alone does not solve stale approval. A quote can become obsolete one minute after review if the recipient changes. Another proposal may remain acceptable after several hours because nothing relevant has changed. You need a maximum execution window plus checks for the changes that matter to the specific business action.
An October HackerNoon discussion of authorisation leases illustrates a permission record tied to resource version v317, a ten-minute validity period and one permitted use. Those are design examples, not statutory limits or universal defaults. The useful idea is that permission has scope, duration and an invalidation route rather than lasting indefinitely in a message history.
Choose the duration according to the commitment. A rapidly changing booking slot may need a short reservation and an immediate availability check. A routine supplier order might allow a longer window while still checking price and cancellation status. A public statement needs its approved text and attachments protected even if the publication date is already agreed. Record which changes are material instead of forcing every harmless internal note to trigger a complete review.
Immediately before execution, re-read the relevant state and compare it with the approved basis. For systems supporting version-aware writes, require the expected version so a competing update cannot slip through between the final check and the write. When several external systems are involved, a single database check cannot make the whole operation atomic. Use supported reservations or conditional operations, and send unresolved conflicts to a person rather than claiming certainty.
Also check whether the permission was revoked and whether the approver still has the required authority. A customer withdrawal or a policy change can invalidate a decision without changing the quote total. The practical standard is not that every permission expires after ten minutes. It is that your workflow can explain why this permission still covers this action at the moment the business becomes committed.
A fresh timestamp does not rescue approval for a changed recipient, amount or record.
Build on existing workflow controls, then test the missing boundary
You probably do not need a new bespoke approval platform. Start with the workflow engine, CRM or transaction system already responsible for the action. Microsoft Agent Framework documentation describes request-and-response handling through RequestPort in its .NET examples and request_info with response handlers in Python. These mechanisms let a workflow pause and receive an external decision. They do not, by themselves, define your price tolerance, customer cancellation rule or approval expiry policy.
Ask whoever maintains the integration where those business checks run. The answer should name the execution boundary and the authoritative records consulted, rather than saying that the model has instructions to be careful. The workflow should preserve the original proposal, the approval decision and the result of its final validation. That gives the next operator a usable explanation when the action cannot proceed.
Test five deliberately awkward cases in a safe environment. Change the price while waiting. Change the recipient after approval. Cancel the customer request before execution. Deliver an approval after its allowed window. Finally, modify the target record between the last read and the attempted write. For each case, define whether the workflow must block, refresh the review package or proceed within a previously approved tolerance.
Include a control case where an irrelevant internal comment changes but the approved commitment remains identical. Otherwise a technically cautious system can produce unnecessary reapprovals until staff start ignoring them. The point is to catch material divergence, not reward the software for generating the most friction.
What this means in practice is that acceptance testing should inspect the receiving system, not just the agent's final message. Confirm that the email was not sent, the order was not placed or the access change was not applied in each blocked case. Keep your own test records independent of any dashboard, including the approved proposal, changed fields and observed outcome. Use existing version checks and workflow features before commissioning custom machinery.
A workflow pause is a mechanism. Keeping the resumed action within permission is your business rule.
Reduce review noise without letting old permission travel
The obvious objection is that repeated approval slows everything down. That objection is valid when systems ask staff to reconsider unchanged, low-risk work. It is not a reason to let a different commitment borrow an old decision. The better answer is to reduce unnecessary requests while being precise about the few boundaries that matter.
Let the assistant gather evidence, compare supplier options and prepare documents without repeatedly interrupting someone. Apply approval to consequential actions or policy exceptions. Where the business genuinely wants standing authority, describe a bounded mandate with permitted actions, amounts, recipients and expiry. Revalidate that mandate at execution rather than manufacturing a fresh interruption for every harmless preparatory step.
When new review is necessary, show the difference. A message saying the quote increased from £1,200 to £1,450 and the delivery date moved is more useful than another identical approve button. Make rejection, revision and cancellation normal outcomes. Identify who owns requests that expire over weekends so abandoned work does not silently accumulate or resume unexpectedly.
There is a UK data protection connection where changed records involve people. The ICO's accuracy guidance says organisations should take reasonable steps to keep personal data accurate and make its source and status clear. The page is under review following the Data (Use and Access) Act. It does not prescribe a ten-minute approval expiry, and a stale approval is not automatically a reportable breach. Nevertheless, acting on an outdated address or misleading account status creates a practical reason to connect approval checks with record accuracy.
Start with one workflow this week. List the fields that would change the decision, define a reasonable execution window and rehearse one pending approval across a material update. Measure expired requests, material-change blocks and reviewer effort separately. Success is not the greatest number of approvals collected. It is fewer unintended commitments, with a clear explanation of why each executed action still matched the authority the business actually granted.
Reduce unnecessary approval requests, not the connection between permission and the action it covers.
Key Takeaways
- Bind approval to the exact outgoing action or a written operating envelope, including recipient, amount and relevant record version.
- Combine an execution deadline with checks for material changes, revocation and reviewer authority.
- Use existing workflow pauses and conditional writes, but test the business-specific validation before resuming.
- Show reviewers what changed instead of asking them to reread an unchanged approval card.
- Verify blocked actions in the receiving system and keep evidence of both the approval and final validation.
