An employee asked your AI assistant a compliance question. It answered clearly, cited the policy, and applied a rule that was replaced in March. Six months later an auditor asks which version it used, and nobody can say.
A supplier invites one of your account managers to dinner. Estimated cost: $120 a head. Before accepting, the employee does the right thing and asks the company AI assistant whether it's allowed.
"Yes. Under §3.2, gifts and hospitality up to $150 can be accepted without approval."
On 1 March 2026 your Gifts & Hospitality Policy was updated to version 5, and the limit dropped to $100. The old version 4 is still sitting on the intranet, in last year's training deck and in a shared folder. The AI found one of those. The employee followed the advice in good faith, and an approval that should have been required never happened.
The dinner is minor. The question that follows isn't: how many other answers came from superseded versions, and can you show which version each one used?
AI assistants answer from the passages that best match the question. An old policy that happens to be worded like the question will beat the current one, and nothing in the process checks effective dates, superseded status or approval state. A superseded rule and a current rule look exactly alike once they've been turned into an answer.
That's a governance problem as much as a technical one. Regulators increasingly expect an AI decision record to show exactly which policy and which version was relied on, not just "the gifts policy" (CX Today, Velt). If the assistant can't tell versions apart, nothing downstream can record the difference.
"Can I accept a $120 dinner from a supplier?"
The superseded policy is picked because its wording fits the question, not because it's current.
Clear, cited, applied in good faith. Against a rule that no longer exists.
Your AI applies the version it read earlier, after the policy was updated.
Your AI is told when a rule it read has changed since, so it re-checks before answering.
Your AI treats two versions of a policy as identical and answers from whichever it found first.
The versions are paired and you see exactly what differs, such as the limit dropping from $150 to $100.
There's no change log for half your policies, so your AI can only guess what was edited and when.
ARR shows what changed between documents, even with no history or backup to compare against.
Your AI states the requirement and leaves out the exemption written on the next line, or the reverse.
When a requirement and its exception sit together, your AI is handed both, so neither arrives alone.
The section numbers in your evidence pack drift as policies are revised, and nobody can prove what the AI actually saw.
Every reference carries a fingerprint of the exact text, so a reviewer can confirm what was relied on (in shared setups).
A well-run policy management system keeps one approved, current version. That's the right foundation. But copies escape it: exported files, attachments in old emails, last year's training slides, a team's own summary on the wiki. An AI assistant that can read across your organisation will find those too, and it has no reason to prefer the official one.
ARR doesn't replace your policy system or your judgment. It changes what the AI is handed: the current rule with its exceptions, a clear signal when two versions disagree, and a warning when something it relied on has changed.
ARR helps your AI tell versions apart and notice when they change. It isn't an audit-trail system by itself; recording which version each answer used is still the job of your logging and governance tools. We've measured these abilities in our own benchmark on code and text, not yet on policy documents, so treat this as how ARR is built to work here. That's what early access is for.
If it applied the old rule, or can't say which version it used, that's the gap ARR was built to close.
We're opening ARR to a small group first. Tell us how your policies are managed and we'll be in touch when your place is ready. No payment required.
More in ARUKAS Field Notes:
Your AI is answering from an old document
When legal AI gets the clause wrong
Chunking breaks meaning
Why AI coding agents break things three files away
Running parallel coding agents without them overwriting each other
AI citation errors start before the AI writes anything
Duplicate files, duplicate totals: where AI reconciliation goes wrong
What AI prior-art search doesn't tell you it missed
How ARR works →