AI agents for business
An AI agent is not a chatbot with pleasant answers—it is a member of staff with access to your systems and a clear area of responsibility. It reads the knowledge base, goes into the CRM, files the request and hands a human whatever it cannot do.
Scope of work
point by point.
- Process review and the choice of where an agent pays for itself
- Knowledge base collected and cleaned up
- Document search that names the source of each answer
- Tools: CRM, database, calendar, email, 1C (Russian accounting system)
- Escalation rules: when the agent must call a human
- Protection against prompt injection and data leakage
- Quality measured on a labelled set of questions
- Conversation panel: what was asked, and where the agent got it wrong
- Token spend control and limits
- Further tuning on real conversations after launch
Situations
where this pays for itself.
First-line support
Answers from the knowledge base with a link to the source, opens a ticket, escalates anything contested to a human.
65–75% of requests with no operator
Qualifying inbound enquiries
Clarifies the task and the budget, fills in the CRM record and assigns the task to the right manager.
Customer answered in 40 seconds at any hour
Assistant for internal documents
Policies, contracts, instructions. Someone asks in plain words and gets an answer with a quote and a link.
−80% of the time spent finding a document
Processing incoming documents
Pulls the data out of invoices and delivery notes, checks it against the order and flags discrepancies.
200 documents an hour
Four steps
from brief to handover.
- 01
Choosing the section
We work out where an agent gives a measurable saving. If the process carries legal liability, we say do not automate it.
- 02
Knowledge base and tools
We gather the documents, clear out the contradictions and connect the systems. 80% of answer quality is decided here, not by the model.
- 03
Pilot on real conversations
The agent answers in the shadows: a person sees the suggestion and decides whether to send it. That builds up labelled data without the risk.
- 04
Launch and supervision
We switch it on for part of the traffic, watch the metrics, and go through the errors and fix them every week.
What we
build it with.
Three tiers.
The exact estimate follows the brief.
from ₽300,000
15–25 working days
- One scenario and one knowledge base
- Document search with a link to the source
- One integration: CRM or email
- Escalation rules for handing over to a human
- Conversation panel and metrics
from ₽700,000
25–45 working days
- Several scenarios and roles
- Tools: CRM, 1C, calendar, database
- Quality measured on a labelled set
- Prompt injection testing
- Further tuning on real conversations
- Two months of support
on request
45–90 working days
- Several agents that hand tasks to each other
- Deployment inside your own network
- A local model where data cannot leave your side
- Decision log for audit
- SLA and on-call
Prices are the lower bound. What pushes an estimate up is set out on the pricing page
About this
service.
01Will the agent make answers up?
It answers only from your documents and has to cite the source—otherwise the answer is not shown to the customer. If the base has no answer, the agent says so and calls a human. Before launch we run a set of 100–300 labelled questions and show you the share of correct answers as a number.
02Will our data end up at OpenAI or Anthropic?
That depends on what we agree. By default we use the API mode with no training on your data. If the data is sensitive, we deploy an open model inside your own network—quality is lower, but nothing leaves the perimeter. This is decided at the brief, not after launch.
03What does it cost to run?
Usually ₽3–15 per conversation, depending on length and model. At 3,000 requests a month that is ₽10,000–45,000. We set limits and alerts so the bill holds no surprises, and show the spend by day in the panel.
04Why is this better than a scripted bot?
A scripted bot is cheaper and more predictable—if there are ten questions and they do not change, take the bot. An AI agent is for when questions are phrased freely and the knowledge base is large and changes constantly. Often the right answer is both: a script for the routine, an agent for the rest.
Get an estimate:
AI agents for business
The brief takes 5–7 minutes. In working hours we reply within two hours, and the estimate is free.