The Series Is Where It Breaks
Co-Founder & CEO of CLRT
Picture the review process most firms have built for AI-produced work. A draft arrives, someone or something checks it against the brief, the facts are tested, the tone is corrected, and it ships. Then the next one arrives, and the same careful process runs again. Each document is judged on its own merits, which is how documents have always been judged, and at low volume it works. The trouble begins when the volume rises, because a body of work has properties no single document has. It can repeat itself. It can contradict itself. It can lean on one example until the example becomes the message. None of those failures lives inside a document, so none of them can be found by reading one. They live in the gaps, and the gaps are what nobody was assigned to read.
CLRT recently produced a thirty-part long-read series from its own Insights library, made to run one piece a day, each running to about 2,400 to 3,000 words with every source linked inline. A team of writer agents drafted from one shared specification, and every piece passed through gates that would satisfy most quality teams: a scan for banned content, limits on the text of every phone card, a structural check, every source link opened and confirmed, and every rendered card viewed by a person. Every piece passed. Taken one at a time, each article was accurate, well sourced and on brief, which is the ordinary definition of quality for AI-produced work. By that definition the series was ready. The editor pass that followed, reading the thirty as one body of work, is the subject of this piece, because what it found was invisible to every check scoped to a single article.
The series-level pass read the thirty side by side and found three things. One airline chatbot case, Air Canada's, appeared in six of the thirty articles. Each use was legitimate; the case is well documented and fits the argument wherever it appeared. Read in sequence, it became the series' signature, and the editor cut it to four, moving the other days to different verified cases. Day 29's charts repeated a chart from Day 9 and another from Day 19, each correct where it stood, so the plates were replaced until no chart repeated another day's. And Days 5 and 15 cited different adoption figures for the same company, each traceable to a real source published at a different moment; the company's own later update settled it. Not one of these is an error inside an article. Each is a relation between articles, and a relation cannot be seen from either end.
The mechanism is worth stating precisely, because it decides what can catch it. A check, whether a careful reviewer or a model acting as judge, is scoped to the object it is handed. Give it one article and it can test everything that article claims. It cannot test what the article shares with twenty-nine others it has never seen. Nor are these defects random. Writers drawing on the same priors converge on the same famous example, the same clean chart, the most quoted figure, because availability is what a model's sense of a good example is made of. Parallel writers sampling one distribution will recycle by default, and each will do so while behaving correctly. The recycling is not a lapse by any writer. It is the predictable behaviour of many writers who cannot see each other, and it only becomes visible from the place nobody was asked to stand.
Scale makes this sharper rather than softer. The series ran to 165 phone cards and 502 source links, each checked at the piece level, and piece-level work grows in a straight line with volume. Relations do not. Thirty documents contain 435 pairs that could repeat or contradict each other, and a hundred contain nearly five thousand. The pattern is general. A legal team drafting client letters with AI can send two letters, both reviewed, that quote a client different figures. A marketing team can ship a quarter of campaigns, each approved, that all lean on one customer story. Vendor research from BetterUp with Stanford's Social Media Lab, read directionally, found that 40 percent of US desk workers believed they had received AI workslop in the previous month. Our reading is that some of what a reader calls slop is not a bad piece at all. It is a good piece the system has already produced.
The conclusion is uncomfortable for any firm that has scaled AI output by scaling per-item review. At volume, quality is a property of the portfolio, not of the piece, and it needs an owner whose job is to read across. That owner needs things a per-piece reviewer never does: a memory of which examples, figures and claims the body of work has already used, a decision about which source is canonical when two disagree, and the authority to change a piece that is correct on its own because it is wrong in context. None of this is a tool you install. It is a judgment about what in your body of work must agree, what must not repeat, and who answers when it does, made before the volume arrives, because afterwards the defects are already with readers who see the whole.
A check scoped to one document cannot see the defect that lives between two of them.
A deeper dive
There is a detail in the recycled case that makes the argument better than any of the six articles did. In Moffatt v. Air Canada, decided by British Columbia's Civil Resolution Tribunal on 14 February 2024, the airline's chatbot told a grieving customer that a bereavement fare could be claimed after travel. A separate page on the same website said the bereavement policy did not apply once travel was complete. Air Canada argued the customer could have found the correct information on another part of its site. The tribunal asked why customers should have to double check information found in one part of a website on another part of it, and ordered the airline to pay CA$812.02. The defect sat between two published pieces of the same body of work, each presumably owned by a different team. The tribunal treated the website as one body of work with one owner, which is how every reader treats your firm's output. Your client does not experience your letters one at a time. Your market does not read your campaigns in isolation. They hold the whole, and they notice the seam.
The second-order trap is the obvious fix. Give the checker more context, hand it the whole series and ask it for contradictions, and the problem looks solved at thirty pieces. It does not stay solved. Context runs out long before volume does, and a model asked to find disagreement will find it everywhere, flagging legitimate variation, a figure quoted at two dates, an example used twice on purpose, as error. The harder discipline is deciding in advance which kinds of sameness are defects and which differences are contradictions. When Days 5 and 15 disagreed, the right answer was neither the majority view nor the newer article; it was the company's own published update, because provenance outranks recency and both outrank repetition. That is a rule about your facts, not a setting in a tool, and it differs for a law firm, a bank and a newsroom. Stanford's 2026 AI Index reports 362 documented AI incidents in 2025, up from 233 the year before. As that record grows, the question a client or a regulator asks will increasingly be the tribunal's: who was responsible for all of it.
Work with CLRT
CLRT builds AI production systems for firms whose output is judged as a whole, by clients, by regulators and by a market that reads everything. The work is less about making each piece good, which the models increasingly manage, and more about deciding what across your body of work must agree, what must never repeat, which source wins when two disagree, and who owns the seam. That is judgment about your operation, and it has to exist before the volume does. If you are scaling AI-made documents and reviewing them one at a time, talk to CLRT about where your series will break, or start with a diagnostic at ascent.clrtstudio.com.

Vishal Sachar is the Co-Founder and CEO of CLRT in Dubai. He helps businesses put agentic AI to work, leads practical AI workshops for leadership teams and operators, and writes on agentic systems, AI governance and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.


