Sber cut bank content approval from seven days to one hour
Sber replaced a lengthy approval chain with three AI checks: editorial, reputational and legal. The rules were compiled from specialists' past comments and risk maps, while humans were retained for disputed cases. According to the team, the journey from brief to CRM upload shrank from seven days to one hour, although the system currently covers only 30% of SMS messages and push notifications.
AI-processed from Habr AI; edited by Hamidun News
Text generation is usually not the main bottleneck. A draft push notification can be produced in seconds, then spend several days moving between an editor, lawyer, reputational risk team, compliance and product owner. Sber decided to automate this queue itself: standard specialists' comments were converted into machine checks, while humans were reserved for cases where the risk is genuinely high.
According to the project team, the journey from brief to uploading the finished communication into the CRM shrank from seven days to one hour. This is not a universal solution: the complete process currently works for about 30% of SMS messages and push notifications. Still, the case is useful because it demonstrates a practical automation architecture for a regulated company.
Three checks instead of eight approvers
The system divides control into three layers. The first, AI-Boost, checks spelling, editorial rules and product wording. The team examined accumulated comments from editors and proofreaders and distilled recurring requirements into roughly 50–60 rules. It supplemented them with dictionaries, product guidelines and message-length limits.
The second layer, Smart ToV, looks for reputational risks: inappropriate humor, toxic topics, questionable slang and the wrong tone during emergencies. It uses a traffic-light system. A green result passes automatically, while red sends the material to a person.
The third layer is an AI lawyer. For five financial products, the team created risk maps covering dangerous wording, potential liability and remediation options. The model does not merely flag content: it shows the risk it found and suggests a safer alternative. A check that previously had an SLA of up to five days now takes about three minutes, according to the project's authors.
A simple example: the bank prepares a push notification about a financial product. The model must detect a promise of non-guaranteed returns, verify the mandatory disclaimer and determine whether an upbeat tone suits the current news context. If the violation is unambiguous, the copy is returned for revision. If the risk is borderline, a specialist makes the decision.
Where the rules come from
The most valuable part of the approach is not the chosen LLM but the data used for checking. The team exported several years of approved content, analyzed comments from editors, lawyers and risk managers, and then grouped recurring remarks. Otherwise, the model would have to receive an abstract request to “check whether the text is good,” which cannot be measured and updated consistently.
The rules were split into hard and contextual categories. An error in a product name always blocks publication. Slang or emoji may be acceptable for a young audience but inappropriate in another campaign. This separation makes it possible to change an individual filter when a new law or crisis emerges without rebuilding the entire generation system.
According to Sber's internal assessments, the AI lawyer and Smart ToV exceed 90% accuracy. The checks produce false alarms more often than they miss genuine risks. For a bank, this is a deliberate bias, but the figure cannot be treated as an independent benchmark: the article does not disclose the testing methodology or the size of the control sample.
Practical takeaways
A product team should start not with an agent but with a log of previous revisions. First, identify recurring decisions, define critical violations and assemble a set of labeled examples. It is then useful to run the check in shadow mode, compare its decisions with those of experts and only afterward allow low-risk materials to pass automatically.
There are limitations. The case was described by Sber's own team, so the time-savings and accuracy figures require independent verification. The full process was launched recently, does not cover all content and currently works better with short formats. Moreover, automation does not eliminate editors and lawyers: it shifts their role from reading every text sequentially to configuring rules, handling exceptions and overseeing quality.
The practical conclusion is simple: in corporate content, it is more effective to place an LLM not only at the start of the pipeline, where it writes a draft, but also after it—as a set of narrow, measurable filters. The main gain comes from eliminating the approval queue, not from generation speed.
- Nov 12, 2026Claude in Chrome requires approval for every click on category 3 sites
- Nov 12, 2026Norwest raises record $4 billion fund, betting on agents, voice and physical AI
- Nov 12, 2026AWS shows how 100% retrieval precision can hide an AI agent's failure
- Nov 12, 2026Garage Eight dropped peer approvals for 69 MRs in an internal service using AI review
Key Questions Answered
What are the key details on Three checks instead of eight approvers?
The system divides control into three layers. The first, AI-Boost, checks spelling, editorial rules and product wording. The team examined accumulated comments from editors and proofreaders and distilled recurring requirements into roughly 50–60 rules. It supplemented them with dictionaries, product guidelines and message-length limits. The second layer, Smart ToV, looks for reputational risks: inappropriate humor, toxic topics, questionable slang and the wrong tone during emergencies.
Where the rules come from?
The most valuable part of the approach is not the chosen LLM but the data used for checking. The team exported several years of approved content, analyzed comments from editors, lawyers and risk managers, and then grouped recurring remarks. Otherwise, the model would have to receive an abstract request to “check whether the text is good,” which cannot be measured and updated consistently. The rules were split into hard and contextual categories. An error in a product name always blocks publication.
What are the key details on Practical takeaways?
A product team should start not with an agent but with a log of previous revisions. First, identify recurring decisions, define critical violations and assemble a set of labeled examples. It is then useful to run the check in shadow mode, compare its decisions with those of experts and only afterward allow low-risk materials to pass automatically. There are limitations. The case was described by Sber's own team, so the time-savings and accuracy figures require independent verification.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.