How we cut content time-to-market from 7 days to 1 hour with a data-driven approach (Как мы сократили Т2М контента с 7 дней до 1 часа, применив подход на основе данных и вывернув известный мем наизнанку)
B · Useful, limited Mikhail Miroshnikov of Sber's B2C data and recommendations department describes replacing a chain of human approvers with three GigaChat checks, with process detail and volume figures, but accuracy and conversion claims are self-reported without method.
Summary based on the full text·Original: about 12 minutes to read
In short
A Habr post on Sber's company blog (3 November 2026) by Mikhail Miroshnikov of the B2C data and recommendations department. A push notification once took about two weeks and up to nine approval steps; by late 2023 the cycle was 5-6 working days. Sber grouped the approvers (editor, proofreader, reputation risk, legal, part of compliance) into three GigaChat checks inside its own content tool, Creator. For content those checks can cover, brief to CRM upload now takes about an hour. A person intervenes only when a check flags a high risk.
Key numbers
- about 1 hourbrief to CRM upload for content that passes the AI checks, against a 5-6 working-day cycle at the end of 2023
- 50-60editing rules derived by clustering years of editor and proofreader comments
- 5 days to about 3 minuteslegal review time, SLA vs the AI lawyer check
- about 11,000 to over 22,000content units produced in 2023 vs last year; the author's team went from 10 to 5 content managers
- 30%share of SMS and push messages currently running through the full end-to-end AI process
As reported by the source.
Russian·Web page·opens habr.com
What it reports
- Generated share: 0% of texts at the end of 2023; 30% of SMS and pushes on GigaChat and DeepSeek by the end of 2024; 70% of short formats generated by GigaChat alone by the end of 2025. Approval stayed the bottleneck.
- AI-Boost (editing and proofreading): rules came from all approved content of the last couple of years, whose comments clustered into about 50-60 rules for SMS and pushes, which make up 75% of communications.
- Smart ToV (reputation risk): a traffic-light list of toxic topics and wording, plus a restricted mode for regions hit by emergencies. Human approval only on a "red" flag.
- AI lawyer: risk maps for five key financial products listing risky wording, possible liability and mitigations. Legal review that took five days under SLA now takes about three minutes.
- Checks were piloted through a mail bot with false positives before release into Creator.
- Volume: about 11,000 content units in 2023, over 22,000 last year. The author's team doubled productivity while going from 10 to 5 content managers; the others changed roles or teams.
- Claims: AI lawyer and Smart ToV accuracy above 90%; checks over-flag more often than they miss; CTR and conversion of AI texts at the level of in-house copywriters.
- The full end-to-end process launched only last quarter and covers 30% of SMS and pushes; the author expects twice that by year end. A new "AI validator" role checks generated drafts instead of writing.
How they got it
A first-person account of the process redesign. Rule sets were derived from historical approver comments. No description of how accuracy or conversion was measured.
Limits
- Accuracy, conversion and time figures are self-reported; no sample or test design given.
- "1 hour" applies to the roughly 30% of SMS and pushes covered end to end; emails, landing pages and other channels still go to human editors.
- Comparisons with Amazon, Klarna and Persado are relayed from public claims, not checked.
Why it matters for companies
The method is reusable in any regulated content operation: mine the history of approver comments, turn the recurring ones into explicit rules, and run those rules as checks that can be changed quickly when law or circumstances change. Keeping checks separate from the generator lets compliance adjust one lever instead of retesting the whole pipeline. The team-size change (10 to 5 content managers with double output) is the workforce effect to plan for.
Key passages in translation
Russian original
Править фильтрующие проверки значительно легче с точки зрения релизов и операционного ресурса на разработку.
Changing the filtering checks is far easier in terms of releases and development effort.
Russian original
То, что раньше по SLA занимало пять дней, теперь занимает минуты три.
What used to take five days under the SLA now takes about three minutes.
Russian original
Причём проверки чаще докапываются до несуществующих рисков, чем пропускают реальные.
And the checks more often dig up risks that do not exist than miss real ones.
Summary based on the full text. Read the original for the full argument and data.Read the original on habr.com