Service 03 — audit the data, retrieval, authority, and evidence.
AI audits
A model cannot repair a stale export, reconcile two definitions, or decide which shadow table deserves authority. It only reads the pile fluently. We audit the data it sees and the actions it can take as one system — because reliable answers and governed behavior share the same boundaries.
The work
We treat readiness and governance as one chain. The source, the retrieval path, the prompt context, the capability, and the resulting evidence must each point back to an owner and a contract.
01
Map authoritative sources, duplicate corpora, exports, vector indexes, freshness, lineage, ownership, sensitive fields, and contradictory definitions.
02
Evaluate whether representative questions retrieve the right evidence, exclude stale or unauthorized copies, cite their sources, and abstain when the data cannot support an answer.
03
Identify context that consumes tokens without improving decisions, along with data that expands the prompt-injection, privacy, and exfiltration surface.
04
Trace permissions, tool access, pre-action checks, approval boundaries, side effects, monitoring, human escalation, and evidence that can be verified outside the model transcript.
What you leave with
Tangible, not theatrical
A plain assessment of sources, retrieval, evaluation, governance, and operations — including where the honest status is not ready.
The duplicate, stale, low-value, or oversized context being embedded, retrieved, and paid for without improving the answer.
Sensitive data, capabilities, attack paths, approval gates, evidence, monitoring, and accountable owners connected in one view.
A sequenced cleanup and control plan, plus a bounded pilot recommendation only when the data and operating model can support it.
The xerocopy difference
AI audits are often model report cards. That is too late in the chain. xerocopy begins with the copies the model is asked to trust. If two indexes disagree, if lineage ends at an export, or if a source has no owner, another evaluation harness will not create truth.
Fewer copies reduce retrieval noise, token spend, sensitive-data exposure, and prompts available to poison. Contracts at the boundary make the remaining context explainable. The same discipline then governs actions: the model proposes, policy commits, and the evidence survives the transcript.
Garbage in, garbage out — except with AI, the garbage is embedded, retrieved, and billed per token.
Related work
Follow the problem
Start here
Send a paragraph. We will tell you whether this service fits, what a small first engagement could answer, and when we are not the right shop.
hello@xerocopy.sh →