· Zenous Team · 3 min read
Auditing an AI-Native Program: What to Check When Agents Do Half the Work
Classic delivery audits assume humans did the work. On programs where agents draft, reconcile, and report, five audit questions change. Here is the checklist extension we use in 2026.
A delivery audit answers one question: can the evidence support what the status report claims? On an AI-native program, where agents draft documents, reconcile data, write code, and produce the status report itself, the classic audit still applies. But five new checks decide whether the audit means anything.
1. Provenance: who actually produced this artifact?
The first check is boring and decisive. Sample ten artifacts (plans, reconciliations, commit sets, status narratives) and establish for each whether a human, an agent, or a pair produced it, and who approved it. On healthy programs this takes minutes because the evidence trail exists. On unhealthy ones the answer is a shrug, and that shrug is the finding. If provenance is unknowable, every downstream claim inherits the doubt.
2. Review debt: is approval real or ceremonial?
Agents produce faster than humans review. The gap accumulates somewhere, usually as rubber-stamping. The audit test: take the approval log, pick approvals at random, and ask the approver to explain the artifact. A reviewer who approved forty agent outputs in an afternoon did not review forty outputs. Measure the ratio of agent output volume to genuine review capacity. When it exceeds roughly five to one, the program is shipping unreviewed work with a signature on it.
3. Silent scope: what did agents change that nobody decided?
Scope change used to arrive through a change request. Agents introduce a quieter channel: a data model “tidied” during reconciliation, requirements text “clarified” in redrafting, code “improved” beyond the ticket. None of it malicious, all of it undecided. The check is a diff audit on a sample of agent-touched artifacts against their inputs, looking for changes with no corresponding decision record. Three findings here usually predict thirty.
4. Skill atrophy: can the team still do the work?
This one is new to 2026 audits and teams hate it, which is how you know it matters. If agents have drafted every estimate for six months, ask a planner to walk through the last estimate’s logic. If agents reconcile all data, ask who would detect a subtly wrong reconciliation. The program’s resilience is the humans’ remaining ability to check the machine. Where that ability has atrophied, the audit flags a single point of failure that happens to be a language model.
5. Vendor agent opacity: what is inside the black box?
Programs increasingly consume vendor products with embedded agents. The contract questions from the vendor management playbook still apply, plus one: can the vendor produce the same provenance and evidence trail you demand of your own agents? A vendor who cannot say what their agent touched in your data is a vendor whose SLA does not cover your actual risk. Get the answer in writing before the audit forces the question.
What stays the same
Everything else. Decision rights, plan integrity, RAID freshness, benefits traceability: the classic 40-60 criteria carry over unchanged, because agents did not repeal the reasons programs fail. They added a new way for a program to look healthy while drifting: fluent artifacts, confident reports, and nobody quite sure who decided what.
That is the audit’s job now. Not to slow the agents down, but to confirm the program can still explain itself.
Our 7-14 day delivery audit runs the classic criteria plus the five checks above. The free project audit checklist covers the self-assessment version. If your program has agents in the loop and no recent independent read, book a consultation.
