# A third of the deviations came from the document system itself. On this pharmaceutical site, close to **30% of deviations** were attributed to a documentation or training cause. The internal reading was the one you hear everywhere: a problem of rigour in the teams. It was a problem of document design. The rate came down to 3%. 30% → 3% Share of deviations whose retained cause was documentation or training, before and after the programme. Total deviation volume fell without a single control being added. Share of deviations of documentary origin: 30% before, 3% after.30 %Before the programme3 %Afterno control added Pharmaceutical production site, figures anonymised at the client's request. The interesting fact is not the drop, it is what the drop reveals: **when a third of your deviations point at your own documents, your documents are the least controlled process on the site.** No awareness campaign fixes that, because the cause is not in the execution. "Those deviations are irreducible. It's the human factor." That is the conclusion mechanically produced by an investigation that stops at the workstation. The three causes actually identified on this site were all documentary: **procedures that contradicted each other**, **obsolete versions still circulating at the workstation**, and **training validated on texts that had since changed**. None of them was about operator behaviour. All of them were measurable from a document management export. ## Three levers, in this order. ### 1\. The scan A full export of the document management system, analysed: documents in force, overdue for review, never consulted, duplicates, contradictions detected by content comparison. Two days of work, and a map that points out the areas to address by itself. No rewriting started before this step. ### 2\. The restructuring Merging redundant documents, separating the stable principle from the variable parameter, removing with justification what no longer served any requirement. Every withdrawal went through change control with its rationale — a non-negotiable condition for the operation to be defensible. ### 3\. Connecting training The point that produced most of the gain. As long as the procedure lives in the document system and training lives in another tool, the two drift apart. Connecting them means a change propagates at once to the material _and_ to the qualification — and the "trained on an obsolete version" deviation disappears structurally. ## One more document has never prevented an error. This site had responded to every deviation by adding a clarification, a reminder, a verification. Ten years of that policy produced a document set nobody could keep coherent — and **it was that incoherence which generated a third of the following deviations**. The loop opened the day the standard answer to a deviation stopped being "add something" and became "understand why the document did not hold". ## Beyond the rate. Workload Fewer deviations opened means mechanically fewer investigations, fewer CAPAs and less review. The quality team got time back for the deviations that deserved it. Onboarding With the volume to absorb for qualification shrinking, the time before a new joiner became autonomous followed. It is the most visible benefit for production. Reporting An unexpected effect: operators reported more weak signals, once it became clear that the conclusion would not automatically land back on them. So the figure to remember is not 3%. It is **the share of your deviations that points today at your own documents** — data you already hold, and that almost nobody calculates. ## Go further. [ ### Make your DMS talk The scan, in practice. Read →](/en/blog/data-analysis-dms-document-repository/)[ ### Remove without losing the trail The condition for defensibility. Read →](/en/blog/documentation-simplification-auditable-traceability/)[ ### Writing a procedure The method, and what you leave out. Read →](/en/blog/write-quality-procedure-method/) ## The case, plainly. How do I calculate this rate for my site? + By taking the last twelve months of deviations and classifying the retained cause. Watch for the bias: many files labelled "human error" or "procedure not followed" are in fact about a document that could not be applied or was contradicted. Re-reading a sample of about thirty files is enough to estimate how much reclassification is due. How long did the programme take? + The scan, two days. The restructuring, several months, because it goes through change control and there is no reason to short-circuit it. The effect on the deviation rate is not immediate: it reads over the following quarters, as the cleaned-up versions reach the workstations. Is this result transferable? + The mechanism is; the size of the effect depends on your starting point. A site where 5% of deviations are documentary does not have this seam — and that is good news, it means you should look elsewhere. Which is exactly why you measure before committing to anything. Why is the client not named? + Because publishing a site's name alongside its deviation rate requires written consent, and we do not ask for it for a web page. The figures themselves are those of the file. We can present the case in detail during a conversation, under the same reservations. ## What share of your deviations points at your own documents? You already hold the data. Two days is enough to read it, and to know whether the seam exists on your site.