AI assistant or regular bot: choose by the task

Start with the job: book an appointment, check an order, explain a condition or collect information for a quote. An open conversation alone does not make a service more useful. A business in Tashkent needs the right answer in the right language, a saved enquiry and a clear next step. Divide the workflow into predictable actions, interpretation of text and decisions that remain with an employee.

1. List jobs before model capabilities

Gather real enquiries from an ordinary working week, removing personal details from the evaluation set. Write down the expected outcome and an approved source for each answer. An address question should return an address; a complex project quote should collect requirements and reach a manager. Include Russian, Uzbek, English, mixed languages, spelling mistakes and voice messages. Voice input adds a transcription check: a fluent reply to an incorrectly recognised request still fails the task.

2. Choose a route for each scenario

A hybrid bot handles recognised commands first and uses AI where interpretation helps. For an order enquiry, the model may identify intent, but the order system supplies the result. Never ask it to invent a status from the conversation. The customer can still use one chat: routing changes how the work is completed, without forcing the customer to learn several separate applications.

3. Separate answering from permission to act

Define approved sources such as a maintained catalogue, delivery terms and an answer library. Treat user messages and uploaded files as input, not instructions to change the operating rules. Access checks, stock and final prices belong to ordinary business logic. Initially, AI can prepare an enquiry draft; cancelling an order, granting a discount or taking payment needs a separately authorised workflow and confirmation. Explain these limits to both the customer and the operator who inherits the conversation.

4. Design the handoff to a person

Escalate when no supported answer exists, records conflict, the customer asks for a person or misunderstanding repeats. Low confidence can be a signal, but model self-assessment cannot replace a source check. Give the employee a summary, language, collected fields and the escalation reason. Show the customer an honest queue state and a way to continue. Do not promise a response within five minutes unless the team actually operates that response schedule.

5. Evaluate against a fixed test set

Illustrative test: prepare 60 enquiries, 20 in each language. Include successful tasks, unavailable products, contradictory information and attempts to access another customer order. Set the expected route and required fields before testing. Score factual correctness, enquiry completeness, correct escalation and prohibited actions separately. Polished wording cannot compensate for a wrong address or invented price. After a correction, rerun the whole set and keep some new enquiries outside the tuning process for an independent check.

6. Price a correctly completed scenario

Include model processing, retrieval, repeated calls, maintenance and operator time. In an illustrative load of 1,000 enquiries, 400 might follow rules and 600 use AI. If each AI enquiry needs two calls, estimate 1,200 calls rather than 600 chats. Compare expenditure with correctly completed tasks. Updating the catalogue and investigating mistakes remain real work after the software is written. A lower model bill is not a saving if the operator must reconstruct most conversations.

7. Run a bounded pilot

• Name the owner of the answer library and the employee responsible for escalations.
• Start with one task and allow AI to be disabled while ordinary buttons keep working.
• Test the manual queue, external-system downtime and repeated customer messages.
• Agree expansion criteria: accuracy, completed tasks, errors and team time.

Finish the pilot with a decision about a specific workflow, rather than an open promise to automate the department. Automation scope: https://samsa.agency/en/services/ai-automation. Published projects: https://samsa.agency/en/cases. Use projects to discuss implementation choices and scope; do not transfer another process outcome to your own without checking the conditions.

Sources and scope

Telegram commands and buttons: https://core.telegram.org/bots/features. Qualified and later-stage leads: https://support.google.com/google-ads/answer/11459091. These documents cover platform capabilities. The routing, evaluation set and workload example above are planning methods, not promised business results.