Your AI should know which version is current
When AI cannot tell current from superseded, it answers with confidence from the wrong document. Source-of-truth is an operating control, not a folder habit.
Confident answers from superseded documents are not a model problem. They are a source-of-truth problem. A manager asks for the current refund policy. The assistant returns a clean summary. The team acts on it. Two days later someone notices the summary came from a draft that was replaced last quarter. The numbers were wrong. The tone was still confident.
Nothing in the chat looked broken. The model did what models do: it used the material it could reach. The company had not marked which document was current, who owned that call, or how superseded versions should behave when intelligence is in the path. Access existed. Authority over currency did not.
This failure is quiet and expensive. It shows up as policy drift, customer promises that cannot be kept, and post-mortems that blame "the AI" when the real gap was company memory without version discipline. Teams then add more documents, more drives, and more prompts. The pile grows. Currency still is not governed. Reframe: currency is authority, not storage A folder called "latest" is not a control. A filename with "FINAL_v7" is not a control. Currency is a decision someone must own and something the operating path must enforce.
Reframe the problem as authority over truth. For each work class that AI may read, name the source that counts as current. Name who can promote a candidate into that role. Name what happens to the previous version when promotion occurs. Without those three answers, every retrieval is a popularity contest between documents that look equally available.
Company context is not "everything we ever filed." It is the bounded set of material the path is allowed to treat as live. Superseded material may still need to exist for audit, training, or dispute. It should not sit in the same retrieval lane as the live rule unless the path is explicitly historical. Mixing those lanes is how confident nonsense gets into customer replies.
Governed action depends on that split. If an output can become a commitment, the acceptance step must know which version underpinned it. Otherwise accountability floats. The person who approved the message cannot defend the basis later because the basis was never named. Business consequence when versions are ambiguous Three costs arrive quickly when AI can see current and superseded material without a currency rule.
First, decisions fork. Two teams ask the same question and receive two different "current" answers because retrieval ranks by similarity, not by authority. Both answers look polished. Operations then reconcile by email, which is another form of unmanaged company memory. Continuity breaks even when people are trying to do the right thing.
Second, risk concentrates on the acceptance step. A human may sign off a draft without noticing the underlying policy was retired. Approval boundaries look intact on paper. In practice, the boundary was crossed the moment the wrong version entered the context package. Governance that ignores currency is theatre.
Third, trust erodes inside the company. People learn that AI summaries need a second search "just to be sure." That second search is unpaid verification labour. It also trains the organisation to treat every AI path as provisional. Scale then becomes politically hard, because leaders correctly sense that the system cannot tell live from legacy.
Enterprise AI does not fail only when models hallucinate from thin air. It fails when company memory contains multiple claimants to the same truth and no governed rule decides which one wins. What a source-of-truth control must hold You do not need a perfect knowledge graph on day one. You need a small set of rules operators can apply under pressure.
Start with object classes. Policies, product facts, pricing rules, customer commitments, and internal process guides do not share the same consequence. Attach a currency owner and an approval boundary to each class. Low-consequence internal notes can tolerate lighter controls. External commitments need named authority before a version becomes live.
Define the promotion path next. A draft becomes a candidate. A candidate becomes current only when a named role accepts it. Acceptance should write three facts into company memory: what is now current, what it replaced, and who approved the change. If those facts live only in chat, continuity will not survive the next tool change.
Separate retrieval lanes. Live paths should prefer current sources. Historical paths should declare that they are historical. Mixing them without a flag is how superseded clauses re-enter customer work. Also define what "I cannot find a current source" should do. Guessing from an old file is not helpfulness. It is an authority failure.
Finally, make version conflict visible. When two documents claim currency for the same class, the path should stop or escalate rather than silently pick a winner. Silence is the expensive option. Walter is the governed company layer that helps intelligence become coordinated work. Models will keep changing. Document stores will keep expanding. What should remain stable is how the company remembers which version is current, who may promote it, and how that fact travels into governed action.
Without that layer, every new assistant inherits the same ambiguity. With it, company context can include currency as a first-class property, not a naming convention. Governance can attach approval boundaries to promotion. Continuity can survive model swaps because the truth rule does not live inside a prompt that will be rewritten next quarter.
Walter's stance is deliberately narrow. The claim is not that one vendor will clean every drive, or that people should stop drafting. It is that enterprises need memory, governance and continuity while models change - so intelligence can act on the version the business has actually authorised. Pick one high-consequence document class this week - a policy, a pricing rule, or a customer-facing process. Write the currency map on one page.
What is the single current source? Who owns promotion into that role? Where do superseded versions go, and can AI still read them on live paths? What must an acceptance step record when someone acts on that source? What should the path do when currency is unclear?
If any answer is missing, stop scaling summaries for that class. Scaling retrieval into an ungoverned pile multiplies confident errors.
A source-of-truth setup sheet is a practical tool we may publish later to help teams record owners, promotion rules and retrieval lanes without another workshop cycle. Until then, the one-page map above is enough to expose whether your AI can tell live from legacy.
If your assistants can reach three "final" versions of the same rule, the constraint is company memory design - not another model. Learn more or join the waitlist at hirewalter.ai.
