For most of working life, polish was a rough signal of effort, and effort was a rough signal of care. A report with consistent tables, a clear structure and a well-drawn diagram usually meant someone had spent time on it, and someone who spent time on it had usually checked it.
That signal has broken. An AI assistant produces a polished document as cheaply as a rough one. Formatting, structure, confident headings and tidy diagrams now cost nothing, so they tell the reader nothing about whether anyone checked the content.
CloudNala's view is that presentation quality and evidence quality are two different axes, and reviewers have to learn to score them separately. The habit of trusting a well-made page is deeply ingrained, and it is precisely the habit a fluent model exploits without meaning to.
Silverleaf Brokers is fictional, and its systems and figures are illustrative. In Part 1 its advisers began producing client review packs with an AI assistant and reviewing them as HTML. Three weeks in, the senior adviser caught an error that worried her more than any typo: a renewal comparison table, beautifully laid out, that quoted an excess from the previous year's schedule. Nothing about the page looked wrong. That was the problem.
Two axes, four quadrants
Plot any artifact on two questions. How well is it presented? And can each material claim be traced to a source and checked?
The rough, unsupported quadrant is not dangerous, because nobody relies on it. The rough, supported quadrant is a different problem: sound work that is hard to read and gets ignored. The polished, supported quadrant is where decisions should happen.
The polished, unsupported quadrant is where risk accumulates. Before generative AI it was rare, because producing polish took enough effort that people tended to check their facts on the way. Now generation moves almost every artifact up the vertical axis for free, and only deliberate evidence work moves anything to the right. Without that work, more and more of an organisation's documents drift into the top-left box.
How fluent text hides weak support
It helps to be specific about the mechanisms, because "the AI might be wrong" is too vague to act on.
Fluency carries confidence it has not earned. A language model writes "the client's current excess is R5,000" in the same register whether it read that figure in this year's schedule, last year's schedule or inferred it from a similar client. The sentence does not change shape when the evidence behind it weakens.
Stale sources look current. Retrieval brings back whatever matches the question. If last year's schedule is still in the document store and matches well, it can be quoted with this year's formatting. Our RAG series covers why freshness and permissions belong in retrieval, not in the prompt.
Tables and diagrams imply completeness. A table with eight rows suggests there are eight things. A diagram with five boxes suggests the system has five parts. Readers rarely ask what was left out of a well-drawn structure.
Consistency is mistaken for correctness. An AI-generated document is usually internally consistent. A wrong figure is repeated faithfully in the summary, the table and the recommendation, which makes it look checked.
Tool failures produce plausible gaps. When a lookup fails quietly, the agent may write around the gap instead of reporting it. The Agent Evidence Layer series covers this in When the tool fails but the answer looks complete.
None of these require the model to "hallucinate" in the dramatic sense. Each is an ordinary failure of evidence, made harder to see by good presentation.
Label the claim, not just the document
The most effective control we have seen is also one of the simplest: label every material statement by what kind of claim it is.
Verified fact. The statement has a source and a date it was checked. "Current excess R2,500 (2026 schedule, page 3, checked 14 October)." The reader may rely on it.
Assumption. The statement is believed but not confirmed, and it has an owner and a way to test it. "Fleet size assumed at 12 vehicles; adviser to confirm with client." The reader must not treat it as fact.
Recommendation. The statement is a judgement, with reasoning and the alternatives considered. The reader is being asked to decide.
Open question. Something material is unknown, with a named person who must answer it before approval. The reader may not approve while it is open.
Silverleaf added these labels to the assistant's instructions and to the review template. The assistant now has to attach a source to anything it presents as a verified fact, and anything it cannot source becomes an assumption or an open question by default. The renewal table that caused the scare would have shown "Excess: R5,000 (source: 2025 schedule)" and the date alone would have stopped it.
There is a cost. Labelled documents are slightly longer and slightly less smooth to read. That friction is the point. It slows the reader at exactly the places where slowing down is warranted.
What reviewers should look at first
A reviewer with limited time should not read top to bottom admiring the layout. They should look for the places where evidence is most likely to be thin:
- Numbers. Every figure that drives a decision: amounts, dates, counts, percentages. Is there a source, and is it the current one?
- Superlatives and absolutes. "Always", "never", "fully compliant", "no risk". These are rarely supportable.
- Comparisons. Tables comparing options are where stale or mismatched sources hide.
- The recommendation. Does it follow from the labelled facts, or from the general shape of similar documents?
- What is missing. Ask what a knowledgeable person would expect to see that is not there.
This is the same discipline good reviewers always applied to junior colleagues' work. The difference is that a junior colleague's uncertainty usually showed in their writing. An assistant's does not.
Polish still matters, after evidence
None of this is an argument against well-presented work. A clear, readable artifact gets better review than a messy one, and the rough, supported quadrant has real costs: sound advice that nobody reads changes nothing.
The order matters. Evidence first, then presentation. An organisation that measures its AI programme by how impressive the outputs look will get more impressive outputs. One that measures by how many material claims carry a checkable source will get outputs it can defend.
Next: Part 3, Comments are not approval. All parts are on the series page. For a related argument about accuracy claims, see Responsible AI accuracy: stop promising a single percentage.
How CloudNala can help
We help teams build evidence discipline into AI-assisted documents: claim labelling in agent instructions and templates, source requirements for the figures that drive decisions, and reviewer checklists that direct attention to where support is usually thin. The result is a short, practical standard your people can apply the same week.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book an AI Assurance Review or write to us at consult@cloudnala.co.za