A workflow will do
The path to the result is known in advance
A request always follows the same route, with the same fields and no decisions along the way. Ordinary automation will run faster, cheaper and more predictably, and it is easier to fix.
Services
AI agent development for business processes
A chatbot answers with words. An agent does the work: it reads the request, goes into your systems, does what an employee usually does — and stops where a person has to make the decision.
01 — The line
When you need an agent and when you don't
We start with this question, because the answer "you don't need one" saves more budget than any optimisation.
An agent is needed
The path depends on what turns up
Who the customer is, what was found in the database, what is missing from a document — each next step depends on the answer to the previous one. It's impossible to map out every branch in advance: there are too many.
An agent is needed
The input is free text or speech
Emails, requests, calls, chats. People phrase things in their own words, and the system has to understand the intent and extract the parameters from it.
Only with approval
Mistakes are costly and nobody checks
Money, commitments to customers, medical or legal conclusions. Here the agent prepares a decision, and a person makes it.
02 — What it does
Work, not just conversation
Understands what is being asked
Reads an email, a request or a chat message and extracts what can be worked with from free text: what is needed, for whom, when and on what terms.
Works in your systems
Reads and changes data where it lives: CRM, ERP, telephony, email, calendar, internal databases and spreadsheets. Not a copy of the data, but the same records.
Gathers what is missing
If there isn't enough data for a decision, the agent doesn't make it up but goes to get it: asks a person, searches the knowledge base, pulls up the request history.
Stops where it should
An irreversible action goes to a person for approval. You decide where the line between reversible and irreversible lies, not the model.
A separate case is when all the agent does is answer from company documentation: procedures, manuals, drawings. We have a dedicated page about that — RAG agent development.
03 — Control
What keeps the agent from doing harm
This is half the work. An agent with nothing to forbid doesn't go into production.
Permissions, not trust
The agent has its own access with the same set of rights as an employee in that role. It can't read what it shouldn't see, even if someone asks it to.
A log of every action
What was done, when, for which request and on what grounds. Reviewing a disputed case takes minutes instead of being reconstructed from memory.
A refusal instead of a guess
The set of cases where the agent must stop and hand over to a person is tested as strictly as the correct answers. Staying silent at the right moment is a measurable result.
Behaviour when systems fail
The CRM is down, the model takes longer than usual, an integration returns an error. What the agent does in each of these cases is defined before the case happens.
04 — Industries
Where it applies
Manufacturing and service
Handling requests, selecting spare parts and assemblies, answering from maintenance procedures, preparing repair orders.
Electronics and instrumentation
Support on specifications and compatibility, handling complaints, preparing answers to distributors' technical queries.
Healthcare and clinics
Booking and rescheduling appointments, answering questions about preparing for procedures, reviewing receptionists' calls and checking communication quality.
Sales and customer service
Qualifying inbound leads, answering questions on terms and availability, preparing quotes, running deals in the CRM without manual copying.
05 — Proof
Our agents at work
An agent for machine documentation
Answers from a library of technical documentation: rephrases the query itself, opens the right page, looks at the drawing. Can run on your own infrastructure.
Case studyCall review without listening
Call transcription, scoring against the client's standards, reports for management. Built for an allergy centre: about 36,000 minutes of calls a month.
Case studyA customer assistant for a barbershop chain
Understands questions in free-form chat and voice messages, answers about services and prices, helps book appointments and passes complex requests to a manager. More than 15,000 users.
Case study06 — How we work
Stages and timelines
/01
Process review
1–2 weeks
We look at how the work is done today: who makes decisions, which systems they use, where time is lost. At the end, we tell you whether an agent is needed here at all.
/02
Pilot on one process
3–4 weeks
One scenario, one or two integrations, a set of test cases and first measurements. The pilot shows numbers, not a demo.
/03
Production version
2 months or more
Access rights, logging, behaviour when systems fail, monitoring. As much work goes into what surrounds the agent as into the agent itself.
/04
Maintenance
under contract
Processes change, and the set of test cases changes with them. Without this, the system quietly drifts away from reality within a few months.
07 — Budget
What the estimate depends on
The cost is calculated after the process review: before it, any number would be made up.
- How many processes are automated and how different they are from each other
- How many systems need to be connected and whether they have a usable API
- Whether your own infrastructure is required — it changes both the model and the hardware requirements
- How strict the requirements for approvals, logging and access separation are
- Whether you have people ready to build the set of test cases with us
08 — Questions
Frequently asked questions about AI agents
How is an AI agent different from a chatbot?
A chatbot answers with words; an agent does the work. A bot will say "3:30 pm is free"; an agent will check the schedule, create the booking, send a confirmation and add the customer to the CRM. The difference is not the model but the fact that the agent has access to your systems and is allowed to change things in them. That is why an agent is designed differently: the main question becomes what it may do without a human's approval.
When don't you need an agent?
When the steps are known in advance and don't change. If a request always follows the same route, a regular workflow will run faster, cheaper and more predictably — and it is easier to fix. An agent pays off where the path to the result depends on the answer: who the customer is, what was found in the database, what is missing from a document. We'll tell you plainly if a script is enough for your task: an agent for the sake of an agent benefits neither side.
What happens if the agent makes a mistake?
That is designed before development, not after. Actions are split into reversible and irreversible: the agent drafts an email or adds a calendar entry on its own, but hands money transfers, messages to customers or contract changes to a person for approval. Every step is logged, so you can always see what exactly the agent did and why. Any system can make mistakes; the goal is to keep the cost of a mistake low.
Do we have to send our data to someone else's cloud?
No, if that is unacceptable. The agent can be deployed on your own infrastructure: the model, the database and the logs stay on your server, and no request leaves it. Companies choose this when the work involves trade secrets, engineering documentation or personal data — in the last case, no third-party AI provider processes that data, which simplifies the GDPR assessment. Cloud models are also an option if there are no restrictions: they are cheaper to start with.
How do you know the agent works correctly?
Before development, we build a set of scenarios with the correct outcome fixed in advance — together with the people who do this work by hand today. After that, every change is run against the whole set, and you can see whether it got better or worse. We test refusals separately: a set of cases where the agent must stop and hand over to a person. Without such a set, "it got better" remains a matter of taste.
Which systems can the agent work with?
Any system with an API: CRM, ERP, telephony, email, messengers, spreadsheets, internal databases. If there is no API, there are workarounds — exports, shared folders, an integration layer. In practice the limit is not technology but access rights: the sooner it is clear what the agent may read and what it may change, the faster the work goes.
Does the EU AI Act apply to such an agent?
It depends on what the agent does. Most business agents — handling requests, working in a CRM, drafting documents — are not high-risk systems, but if people talk to the agent directly, they must be told they are dealing with AI. Uses in areas such as hiring, credit or access to essential services carry much stricter obligations. We build in what the regulation expects in practice — transparency for users, logs of every step, human oversight of irreversible actions — and help you describe the system for your assessment. The legal classification itself remains with your counsel.
How long does implementation take?
A pilot on one process takes three to four weeks: understanding how the work is done today, connecting to one or two systems, a test scenario set and first measurements. A production version with access rights, logging and failure safeguards takes two months or more. Exact timelines depend on the state of the integrations: most of the time goes not into the agent but into access to data.
What stays with us if we stop working together?
Everything: source code, configuration, the test scenario set and deployment documentation. We don't lock clients in with a closed platform — the system runs on your infrastructure without us. The same rule protects us too: there is no dispute over whose solution it is if the answer is written down from the start.
Tell us about the process
It's enough to describe how the work is done today: who makes decisions, which systems they use, where time is lost. That already shows whether you need an agent or ordinary automation will do. If it will, we'll say so.
The first step is a free consultation: a call of up to 30 minutes with the company's founder. You don't need a ready specification for it.
[To be completed: WhatsApp phone number]Documentation agents