AI Agent Development and Implementation for Business
We connect AI agents to your company's workflows: an agent answers customers from the knowledge base, fills in the CRM, sorts incoming documents, prepares reports and drafts — inside your systems and with limited permissions. Actions with consequences run only after an employee approves them, and every step is written to a log. We work with companies in Almaty and across Kazakhstan.
- CRM & 1C
- Human approval
- Action log
What is an AI agent for business?
An AI agent is a program built on a language model that doesn't just answer questions but carries out the steps of a workflow: it finds data in the CRM and 1C, fills in fields, prepares a document, assigns a task to an employee. Unlike a scripted chatbot, the agent chooses the next step itself — but only among the tools and permissions it has been given. InterLAB (Almaty, Kazakhstan) builds AI agents for a specific process: with human approval for significant actions, a log of every step and control over model costs.
Who AI agents are for
Sales teams with a stream of inbound enquiries
Enquiries arrive from the website, WhatsApp and Telegram, and managers spend time on the same first questions and on copying data into the CRM. The agent answers from the knowledge base, clarifies the need and creates a deal with the conversation history.
Support and service teams
Questions about order status, terms and documents repeat every day. The agent answers them from your data and hands a non-standard case to an employee together with the context it has gathered.
Teams that work with documents
Invoices, contracts, requests and acts arrive in different formats and are checked by hand. The agent extracts details and amounts, checks them against the accounting system and prepares a draft for review.
Companies with a large internal knowledge base
Policies, instructions and price lists are scattered across folders and chats, and people ask colleagues for answers. An internal agent finds the answer in company documents and shows which document it came from.
What an AI agent takes on
Managers answer the same customer questions over and over
The agent answers in a messenger or on the website from the knowledge base and current data: prices, availability, terms, order status. If the question isn't in the base or the customer asks for a person, the conversation and its history go to a manager.
Data from conversations is copied into the CRM by hand
The agent extracts the contact, the need and the timing from the enquiry, creates or updates the CRM record and assigns a task to the owner. The manager starts with a filled-in record instead of re-reading the chat.
Incoming documents are checked by hand
The agent reads an invoice, contract or request, extracts details and amounts, checks them against 1C or the CRM and shows discrepancies. Posting and sending are confirmed by an employee.
Numbers for the meeting are assembled by hand
An analytics agent takes a question in plain language, builds a read-only database query and returns a table together with the query, so the numbers can be checked.
People ask colleagues for answers to internal questions
A knowledge-base agent searches the company's policies and instructions (RAG — meaning-based search over your documents) and answers with a link to the source, respecting which documents the employee is allowed to see.
What AI agent development includes
items
- Process review: where an agent helps and where ordinary automation is enough
- Agent design: goal, tools, limits, hand-off points to a person
- Connections to your systems: CRM, 1C, messengers, email, databases, internal APIs
- A knowledge base for answers from company documents, with a link to the source
- An approval screen: an employee approves, edits or rejects the agent's action
- An action log: requests, tool calls, answers and employees' decisions
- Role-based permissions for the agent — no wider than an employee's on the same task
- A test set of scenarios and answer-quality checks before launch
- Limits and tracking of language-model call costs
- A pilot, staff training and support after launch
How much an AI agent costs: InterLAB references
We start with a pilot of one scenario on your data: quality and costs are visible before a large rollout.
| Format | What is included | Reference | Timeline |
|---|---|---|---|
| Pilot of one scenario | For example: answers from a knowledge base or triage of requests and documents; a test set of scenarios and quality measurement before launch | from 800,000 ₸ | 2–4 weeks |
| Agent in production | Integrations with CRM, 1C and email; actions confirmed by a human; action log; data masking | 2–6 million ₸ | 1–3 months |
| Several agents or a model on the company's server | Several scenarios and departments, or a model deployed on your server | from 6 million ₸ | from 3 months |
| Monthly | Support and scenario improvements; model costs — based on actual usage | from 150,000 ₸ per month + model | after launch |
These are ranges, not a price list: we give an exact estimate after reviewing the process. Model costs are estimated during the pilot and capped with limits; if the model runs on your server, you pay for the server instead of model calls.
Contract, payment and warranty
The same terms for every B2B project: payments are tied to accepted results, and a stage price does not change after it starts.
| Term | How we work |
|---|---|
| Payment for smaller projects | Projects up to 3 million ₸ and the MVP sprint: 50% upfront, 50% after acceptance |
| Payment for larger projects | ERP, CRM, SaaS, portals — in stages of 1–2 months: 30% before a stage, 70% after the stage is accepted and signed off |
| Stage price | Fixed before the start. New requirements are estimated separately and taken on only with your approval |
| Warranty | 6 months after acceptance: we fix defects in our code free of charge — whenever the system does not work as agreed |
| Not covered by the warranty | New features and changed requirements, as well as failures caused by changes in third-party systems (a 1C update, a new bank or service API) — that is support |
| Code and access | Owned by the client from the first commit: repository, servers and keys are on your accounts. NDA on your template or ours |
The terms are fixed in the contract before work begins.
AI agent scenarios: process, actions, control
Typical scenarios, not case studies of specific clients. Each one is built around your process and systems; companies usually start with one scenario and expand it after a pilot.
| Agent | Business process | What the agent does | Integrations | Human control | What changes |
|---|---|---|---|---|---|
| Sales agent | First response to enquiries from the website and messengers | Answers questions about products and terms, clarifies the need, qualifies the enquiry, suggests a meeting time | WhatsApp, Telegram, website form, CRM, catalogue or 1C | Discounts, individual prices and commercial proposals — only after a manager agrees | The customer gets an answer to a typical question without waiting for a free manager; the manager gets the enquiry with the conversation history |
| Support agent | Answering customers about order status, terms, returns and documents | Answers from the knowledge base and order data, registers the request, hands a complex case to an operator with a short summary | Messengers, email, helpdesk or CRM, order system | Returns, compensation and exceptions to the rules are decided by an employee; the agent only prepares the decision | Operators handle non-standard cases, while typical questions are closed by the same rules every time |
| HR agent | First screening of applications and answers to employees about internal rules | Compares CVs with the job requirements, asks candidates follow-up questions, suggests interview slots; answers employees from company policies | Email, Google Workspace (calendar, sheets), HR system or candidate database | Invitation, rejection and offer are the recruiter's decision; screening criteria are agreed in advance and visible in the log | The recruiter works with applications that are already sorted, and employees get answers by company rules without waiting for HR |
| Research agent | Gathering information about the market, competitors, suppliers and counterparties | Searches open sources and internal documents, compiles the data into a table, cites the source of every fact | Web search, internal documents, Google Workspace, company database | Conclusions and decisions stay with the employee; a fact without a source is marked as unverified | A review starts from a compiled table of sourced facts rather than an empty document |
| Analytics agent | Management questions about sales, stock and finance | Turns the question into a database query, builds a table or chart and explains which data was used | PostgreSQL and other databases, 1C, CRM, BI reports | Read-only access within the user's role; the query behind each answer is visible and checkable | Simple data questions don't wait in the analyst's queue, and every number can be checked against the original query |
| Document agent | Processing incoming invoices, contracts, requests and acts | Extracts details, amounts and dates, checks them against the accounting system, prepares a draft document or entry | Email, 1C, CRM, document storage, e-document exchange | Posting in the books, signing and sending to the counterparty — only after an employee's review | The employee checks prepared data instead of typing it in, and discrepancies are visible before posting |
| CRM agent | Maintaining customer and deal records | Fills in fields from conversations and call outcomes, sets tasks and reminders, finds duplicates and forgotten deals | AmoCRM, Bitrix24, custom CRM, email, messengers, telephony | Merging and deleting records, changing the owner or the deal stage — on a manager's approval | CRM data is more complete and up to date, and the manager sees the real state of the pipeline |
| Browser agent | Work in web portals without an API: supplier and partner portals, internal web systems | Opens pages, fills in forms, downloads reports and statuses following a set scenario | A controlled browser, a separate account with limited permissions, storage for downloads | Submitting forms, payments and signing — only with a person's approval; the session is recorded for review | Repetitive actions in web portals follow one scenario instead of being done by hand by each employee |
| Knowledge-base agent | Finding answers in internal policies, instructions and price lists | Finds the relevant passages in documents (RAG), answers and links to the source | Google Drive, Confluence or Notion, file storage, PostgreSQL with pgvector | The agent sees only documents the employee can access; without a source found, it says the answer isn't in the base | New employees find answers themselves, and experts are interrupted less by repeated questions |
| Voice agent | Calls: booking appointments, confirming visits, answering common questions | A voice bot built on a language model: recognises speech, answers by script and from the knowledge base, records the call outcome in the CRM | IP telephony, CRM, booking calendar, knowledge base | The caller can switch to an employee at any time; the agent says the call is handled by an automated system | Typical calls don't tie up operators, and call outcomes are kept in the CRM rather than in someone's memory |
| Orchestrator | Multi-step processes that involve several agents and systems | Splits the task into steps, distributes them among specialised agents, assembles the result and watches the limits | Task queues, the systems the agents are connected to, a shared action log | Step and budget limits, stop points for approval; on failure the process stops and goes to a person | A complex process is broken into checkable steps: you can see which agent did what and where a person is needed |
The outcome depends on data quality and how clearly the process is defined. If the task is solved by a fixed scenario without a language model, we'll propose that: it is simpler and cheaper to run. The browser agent is for systems without an API; such a scenario is less robust than an API integration.
Chatbot, AI assistant or AI agent
Three different tools. For some tasks you don't need an agent — a scripted business chatbot is enough.
| Criterion | Scripted chatbot | AI assistant | AI agent |
|---|---|---|---|
| How it works | Buttons and pre-written dialogue branches | A language model answers questions, often from the company's knowledge base | A language model chooses steps and calls tools in your systems |
| What it can do | Take an enquiry, show a menu, answer common questions | Explain, find an answer in documents, draft a text | Find data, fill in the CRM, prepare a document, hand over a task |
| Actions in systems | Only those built into the script | Usually none — just the answer | Yes, within the tools and permissions it has been given |
| When to choose it | Questions and steps are predictable | You need answers from documents and free-form dialogue | The process has many steps and needs data from several systems |
| How it is controlled | The whole script is checked before launch | Answer quality on a test set of questions | A test set, action approval, a log, step and budget limits |
Boundaries: what the agent does on its own and what needs a person's approval
We set the boundaries before development, together with the process owner. The agent gets the minimum permissions its task requires.
The agent does on its own
- Searches and reads data within its role
- Answers typical questions from the knowledge base
- Classifies enquiries and documents
- Fills in drafts: CRM records, documents, replies to customers
- Sets tasks and reminders for employees
- Hands a conversation or task to a person
Only after a person approves
- Payments, refunds and any operations with money
- Sending documents and commercial proposals to customers and counterparties
- Changing prices, discounts and contract terms
- Posting documents in the books and signing
- Deleting data and bulk changes
- Disclosing personal data and granting access
- Decisions on hiring, rejection and sanctions
The list is tuned to the process. Approval can be removed from an action only by the process owner's decision — after the pilot and a review of the log.
How control works
A language model can make mistakes, so control is built into the agent's architecture rather than added after the first incident. The agent can be turned off with a single setting — the process carries on manually.
Action log
Every request, tool call, model answer and employee decision is recorded with time and author. The log shows why the agent acted as it did and what it changed in your systems, and a mistake can be traced step by step.
Role-based access
The agent has its own account with permissions no wider than an employee's on the same task. An internal agent respects the permissions of whoever asks the question: other people's documents don't end up in the answer.
Client data is not for model training
InterLAB's working principle: we connect models via APIs and provider settings under which submitted data is not used for training, send the model only what a step needs and mask unnecessary personal data. The provider's terms are recorded in the project documentation.
Model call costs
Every agent step is a paid model call. Before launch we estimate the cost of a typical scenario, set limits on the number of steps and the budget, route simple steps to cheaper models and show spending by scenario.
Quality checks before launch
We build a test set from anonymised real enquiries and run the agent against it before launch and after every change to instructions, tools or the model.
Protection from injected instructions
The agent treats text from emails, documents and websites as data, not as commands. The arguments of every action are checked by the server, not by the model.
Useful before we talk
How an AI agent is implemented
- 01
Process review
We walk through the process with its owner: steps, systems, data, exceptions, who makes decisions. We decide where an agent helps and where ordinary automation is enough.
- 02
Design and boundaries
We fix the agent's goal, its tools and permissions, the actions that require approval, the hand-off points to a person and the quality criteria. The timeline and cost estimate come at this stage.
- 03
Prototype on test data
We build the agent on anonymised examples of your enquiries and documents, run the test set and show the answers and the step log.
- 04
Integrations and approval screen
We connect the CRM, 1C, messengers and other systems via APIs with minimum permissions and build a screen where an employee approves or edits the agent's actions.
- 05
Pilot
We launch on a share of enquiries or with one team: significant actions go through approval, and we review mistakes in the log and refine the scenario.
- 06
Launch and support
We widen the scope, monitor answer quality and model costs, and update the knowledge base and the test set when the process changes.
Get your project assessed before development starts
Describe the task in the form below — we reply within a business hour and suggest the next step: a call, a review or an estimate.
How it works technically
An agent is a backend service, not a separate "neural network": the language model chooses the steps, and the actions are carried out by your server's functions, which check permissions and arguments. We connect language-model APIs from different providers and choose the model for each step by its quality on your test set, speed and cost. Answers from documents are built on RAG with vector search in PostgreSQL (pgvector), long processes run through queues, and every step is logged.
- Language-model APIs
- Tool calling
- RAG
- PostgreSQL + pgvector
- Agent orchestration
- Task queues
- Logging and tracing
- Speech recognition
- Node.js
- TypeScript
What we connect the agent to
The agent works with the systems your company already has: it reads data and performs actions through their APIs. Where there is no API — through exports and the database, and as a last resort through a browser scenario.
AmoCRM, Bitrix24, 1C, WhatsApp, Telegram, Email, Google Workspace, Databases, Internal APIs, IP telephony
Common questions about AI agents
InterLAB references: a pilot of one scenario — from 800,000 ₸ and 2–4 weeks, an agent in production with integrations and confirmed actions — 2–6 million ₸, several agents or a model on the company's server — from 6 million ₸. The cost depends on the number of scenarios and systems, the state of your data and knowledge base, and data-storage requirements. Separately — support from 150,000 ₸ per month and model call costs, which we estimate during the pilot and cap with limits.
It depends on the process and the number of integrations. First comes a pilot of one scenario on a share of enquiries; expansion happens only after reviewing the pilot's log. We name the timeline for the first stage after reviewing the process.
The knowledge base, the log and the agent's settings are stored in the company's infrastructure or cloud account. The language model receives only what a specific step needs. InterLAB's working principle is to use provider APIs and settings under which submitted data is not used to train models; the chosen provider's terms are recorded in the project documentation. If data can't be sent to an external service, we discuss a model deployed on your own server.
That is built into the architecture: significant actions go through an employee's approval, the agent's permissions are limited, and every step is written to the log. We trace the mistake in the log, add the case to the test set and fix the instructions, the tool or the scenario — after the fix the agent runs through the whole set again.
A chatbot follows pre-written dialogue branches. An AI agent understands free text, chooses the next step itself and performs actions in your systems — finds data, fills in the CRM, prepares documents — within the tools and permissions it has been given. If an ordinary business chatbot does the job, we'll propose that: it is simpler and cheaper to maintain.
Yes, and that's how we recommend bringing AI into a business: one scenario, a clear quality criterion, a pilot on a share of enquiries. After the pilot, decisions about the next scenarios are made on the log and the results, not on expectations.
No. To start, the documents your employees already use are enough: policies, price lists, answers to common questions. During the process review we flag what is outdated or contradictory — the quality of the agent's answers depends on the quality of these materials.
The client: the code, the agent's instructions, the test set, the knowledge base and the API keys are registered to the company's accounts, and documentation and access are handed over after launch. We design the architecture so the model can be replaced with one from another provider without rewriting the scenario.
Related services
We'll review your process and show where an agent helps
Describe the process, the systems and what is done by hand today — no passwords, keys or customer data. We'll propose an agent scenario, its boundaries and the next step.
- The first short conversation is free — we choose the format
- NDA — yours or ours
- Pricing after a brief and an agreed scope