top of page

What an AI Workflow Should Do When a Source Document Is Stale

Writer: Branden Bell
Branden Bell
Nov 11
4 min read

Updated: Nov 12

Check the version. Two cream policy pages carry an olive tab and a terracotta bookmark.

A support lead reads a tidy AI draft promising the replacement will ship within two business days. The draft cites a real policy file. Then someone notices the file is last quarter’s version, and the current policy has a different rule for this kind of order.


When an AI workflow encounters a stale source document, stop the action that depends on it. Verify the current source with its owner, replace the stale input, and regenerate the affected draft. Keep a note of which version was checked and what the reviewer released.


Three reasons to stop and check


  • The document was superseded. The maintained record points to a later version. Use the current version that applies to the case.

  • The document has no reliable status. It has a filename or upload date, but no owner, version or effective date. Ask the owner to confirm whether it applies.

  • Two sources disagree. One document is newer, but the older one may still govern orders placed before a change. Ask which rule applies to the specific order and effective date.


A recently downloaded copy can contain yesterday’s rule. A policy published months ago can still be current. Check its status in the maintained record and confirm the scope before deciding. The source-of-truth guide shows how to name the record that owns each fact.


Worked example: a response that can wait


Imagine a fictional parts supplier. Its assistant drafts support replies from a shared collection of policy PDFs. A customer asks when a replacement can ship. The draft quotes policy v3, which says qualifying replacements ship within two business days.


  • What the reviewer sees. The policy register lists v4 as current. It says orders fulfilled by a partner require partner confirmation before the team can quote a ship date. This order is partner-fulfilled and was placed after v4 took effect.

  • What stops. The reviewer holds the draft with the two-day promise. They log v3 and v4 on the ticket and remove v3 from this order’s drafting inputs. They keep older policies labeled with their effective dates for audits and cases where those versions still apply.

  • What moves forward. The owner verifies v4 applies, then asks the partner for confirmation through the approved process. The team sends a reviewed acknowledgment that it is checking availability. The ship date remains open.

  • What closes the ticket. The partner confirms a ship date. The reviewer verifies that date against the confirmation and releases the updated response. The ticket retains the current policy version and confirmation reference.


The pause applies to the promise that depends on the policy. The team can still acknowledge the request and work on the missing information. If the corrected message is approved and then changes again, use the exact-message approval guide.


Put the version check where someone will see it


Make a small source record beside the action in your workflow. Here is a copyable version.


  • Source and owner. Replacement policy register, maintained by the support lead.

  • Version and scope. v4, with its effective date and order types.

  • Checked at. The time a reviewer compared the draft’s source with the maintained record.

  • Hold condition. The source is superseded, missing, inaccessible or ambiguous for this case.

  • Resume decision. Named reviewer confirms the applicable source and checks the revised output.


If your team drafts in Claude Projects, its documentation explains that files in the project knowledge base provide context across its chats. Treat that as a drafting input. Assign someone to update the project’s reference files when the maintained policy changes. Claude’s Projects guide


NIST’s generative AI profile recommends reviewing generated content against predefined guidelines and evaluating the feedback loop between content provenance and human reviewers. A short source record is one way to put that principle into practice for a small workflow. NIST Generative AI Profile



Frequently asked questions


How often should a source be checked?


Choose a check interval based on how often the rule changes and the cost of getting it wrong. For a price, promise or approval that someone is about to send, compare the relevant fact with the current record at release.


Is the latest modified date enough?


Use it as a clue. A file can be moved, downloaded or reformatted after the actual policy was superseded. Confirm the version, effective date and scope in the maintained record.


What can the workflow do while a source is on hold?


It can capture the request, flag which fact needs review and route it to the owner. A reviewer can approve an acknowledgment that avoids the disputed commitment.


Should we delete the older document?


Follow your retention process. Move superseded versions out of active drafting inputs and label them so a reviewer can distinguish them from current guidance.


Test one rule your team uses today


Pick a policy that changed this year. Trace which copy an assistant receives, and try one fictional request with an old version attached. Check whether the workflow catches it and names the person who can resolve it. Start with the practical Claude guide if you are mapping your current setup. If you want help redesigning one workflow around its real sources and approvals, share your workflow with Bell Consulting Solutions.

 
 
 

Comments


bottom of page