Client brief
The previous rule-based bot relied on exact wording and frequently fell back. The new assistant had to understand free-form Russian and Hebrew, keep service data isolated, and never quote an unverified price.
Client context
An Israeli company specialising in glazing and custom-built structures; the client’s brand and internal data are not disclosed.
Outcome
According to the project author, the assistant was deployed in the client’s Kommo workflow and accepted by the client. The public case does not disclose the brand, correspondence, commercial terms, or internal identifiers.
04 · DATA AND PROCESS
From CRM audit to a controlled launch
COMMISSIONED IMPLEMENTATION · RU + HE · PRODUCTION
- CRM auditBots, triggers, and fields
- Knowledge structureSources and catalogue
- Bounded actionsFields, tags, and handoff
- QAReference questions
- Controlled launchLive test and acceptance
- Russian + Hebrew
- 18 knowledge sources
- 51 CRM fields
- 20 actions
- 100 reference questions
- Caption
- Controlled implementation flow: CRM audit → knowledge structure → bounded actions → QA → launch.
- Provenance
- Project configuration and QA artefacts; production status confirmed by the project author on 13 August 2026.
- Caveat
- Production status was not independently read back from the Kommo account in this session. The public case makes no claim about measured conversion or revenue uplift.
05 · IMPLEMENTATION AND BOUNDARIES
What was implemented
The native Kommo AI Agent was configured as a controlled layer on top of the official WhatsApp channel: 18 structured knowledge sources, a service catalogue, 51 CRM fields, 20 bounded actions, handoff rules, and a reproducible QA set.
Artifacts
- 18 structured bilingual knowledge sources
- A service catalogue as the single source of price data
- 51 CRM fields and 20 bounded assistant actions
- A set of 100 reference questions for repeatable QA
- Handoff and AI-stop rules for complaints, hazardous situations, and requests for a manager
Quality control
- Answer quality and real CRM actions are verified separately.
- Preview validates content and routing, while live testing validates fields, tags, tasks, and AI stop after handoff.
- Activation follows an isolated control conversation.
Public case boundaries
- The client’s name and internal data
- Correspondence and commercial terms
- Internal account, agent, and catalogue identifiers
- Changes in conversion, revenue, or other business KPIs
06 · CHALLENGES AND DECISIONS
Why the setup required an engineering approach
The assistant had to do more than hold a conversation: it needed to work safely with bilingual knowledge, prices, and CRM actions.
The challenges I encountered
Free-form bilingual conversations cannot be handled reliably with exact-match rules.
Different services have different prices and technical constraints; mixed sources could produce an incorrect consultation.
Existing bots and triggers could duplicate replies, while the AI Agent action set is bounded by Kommo capabilities.
Preview validated response content but could not prove deal updates, tags, tasks, or AI stop after manager handoff.
The decisions made and why
Native Kommo AI Agent for the first release
Avoiding an external n8n/OpenRouter chain reduced components, failure points, and integration latency.
Strict source hierarchy
Technical facts come from the knowledge base, while the catalogue remains the only source of prices.
Bounded actions and mandatory handoff
The assistant can use only approved fields and workflows, and hands risky situations to a manager.
Repeatable testing before activation
Reference questions and isolated conversations reveal routing errors before real customer interactions.
Why this approach is stronger than a typical setup
Stronger than a scripted Salesbot: it understands the intent of a free-form message instead of one predefined phrase.
Safer than a generic chatbot: knowledge, prices, actions, and manager handoff have explicit boundaries.
Easier to maintain: the catalogue and knowledge sources can be updated without rebuilding all business logic.
More reliable at launch: answer quality and CRM actions are verified separately before a controlled live test.
For this workflow, the approach is optimal: native Kommo integration combines minimal external dependencies, managed risk, and content the business can update independently.
AI ASSISTANT FOR CRM
Discuss a similar implementation
Describe the channel, CRM, languages, knowledge base, and manager-handoff rules, and we will scope a safe first release without unnecessary integrations.