Who checks the machine? Building a review layer for AI output
Most AI errors that reach a client were visible to someone first. The fix is a review layer designed like any other control: what gets checked, by whom, against what, and what the AI may do without asking.
A confident wrong answer looks exactly like a confident right one. That is the core problem with AI output in a business setting. A fabricated figure in a board pack, a policy clause that does not exist, a summary that drops the one exception that mattered: none of these announce themselves. They are caught by someone who knows to look, or they are not caught.
Most people who use AI at work already know this. In a January 2026 Connext Global survey of 1,000 U.S. workers who use AI, only 17 percent said it could run with minimal human involvement, while 70 percent said reliability comes from AI paired with some level of human review. Yet the checking is rarely designed. It is left to whoever happens to read the output last.
The cost of leaving it informal is visible. In a VentureBeat survey of 157 enterprises in June 2026, half had shipped an AI feature that passed internal evaluation and still caused a customer-facing failure. Only 5 percent said they fully trusted their automated evaluations.
Four failure patterns worth training for
Invented specifics. Numbers, dates, citations, names and clause references that sound plausible and appear in no source. The tell is precision without provenance. The rule: any specific fact that will leave the building must trace to a document someone can open.
Silent omission. The summary is accurate but leaves out the exception, the caveat or the dissenting view. This is harder to catch than an invented fact, because nothing on the page is wrong. Reviewers need the source alongside the output, not the output alone.
Borrowed confidence. The model answers a question it should have declined: a legal interpretation, a medical judgement, a forecast with no data behind it. Tone is no guide. A well-written paragraph is not evidence.
Skewed framing. Bias rarely shows up as an offensive sentence. It shows up in whose perspective is treated as the default, which options get a full paragraph and which get a clause, and which customers a scoring rule quietly disadvantages. Checking for it means asking who the output would look different for.
Precision without provenance is the most reliable warning sign we know.
Design the review layer like a control
Review is a process, so write it down like one. For each kind of AI-assisted output, decide four things in advance: what is checked (facts, figures, completeness, tone), who checks it, what they check it against, and what happens when it fails.
Match the depth of review to the cost of being wrong, not to how confident the output sounds. We sort outputs into three tiers: internal and reversible, internal and consequential, and external. Only the first gets light-touch review. External output always gets a named human sign-off.
Agents need permissions, not just review
Once AI moves from drafting to doing (sending, updating, filing, refunding), reviewing the output is too late. The control has to sit on what the agent is allowed to do. Give each agent its own identity rather than a borrowed login, grant the narrowest access the task needs, and list the actions that always require human approval: anything that moves money, contacts a customer, or changes a record of account.
Write those permissions on one page a manager can read. If the list cannot be explained to the person accountable for the process, the agent has more access than anyone has agreed to.
Make checking a habit
Reviewers tire, and AI output is long. Make checking easier: require the system to show its sources inline, flag every figure for confirmation, and keep a short log of errors caught. That log becomes your best training material, because it shows the team exactly how your tools go wrong on your work.
- Connext Global, Only 17% Say Workplace AI Is Reliable Without Human Oversight, via Business Wire, 18 February 2026. businesswire.com
- VentureBeat, Enterprise AI is entering an evaluation gap, June 2026 VB Pulse survey. venturebeat.com
- —Treat precision without provenance as the main warning sign. Every external fact must trace to a source someone can open.
- —Decide what is checked, by whom, against what, and what happens on failure, before the output exists.
- —Scale review to the cost of being wrong. External output always gets a named human sign-off.
- —Agents that take actions need their own identity, the narrowest access, and a written list of actions that require approval.