Business process automation with AI means handing the machine the steps where something has to be understood — free text, a document or a conversation: reading an incoming email, pulling the amount and bank details out of an invoice, transcribing a call and writing the outcome into the CRM. Operations that arrive as ready fields — stage, date, amount — are cheaper to cover with ordinary rules, and AI goes where no rule can be written in advance.
Key takeaways
The difference is in what arrives at the input. Rules and scripts work when the data already sits in fields: deal stage, amount, deadline, request source. AI is needed where the input is free text, a scan or a recording, and nobody knows in advance what shape the information will take.
That gives a working order: first check whether a rule can do the job. A rule is cheaper, behaves the same way every time and can be verified with a test. AI costs more and needs its quality measured, but it takes on what a rule cannot describe.
| Criterion | Rules and scripts | AI |
|---|---|---|
| What arrives at the input | Filled-in fields: stage, amount, date, request source | An email, a scan, a photo, a call recording |
| Typical tasks | Change the stage, create a task, send a reminder, assemble a report from numbers | Sort enquiries by topic, pull details out of an invoice, transcribe a call, draft a reply |
| Behaviour in a new case | It stops: no rule, no action | It always answers, including incorrectly |
| How quality is checked | With a test: the same input gives the same output | On a sample of your own examples: the share of correct fields and answers |
| What the price is made of | A one-off setup, after which running it costs almost nothing | Setup plus a charge for every request to the model |
The scale of the task shows in the estimate by McKinsey: the automation potential grew precisely because models learned to work with natural language.
This is an estimate of potential, not the outcome of a rollout: it shows how many activities can be automated at all, not how much time will be freed up in your company. The order of steps before choosing a tool is covered in our article on where to start with automation.
The first candidates are operations that repeat every day, arrive as an unstructured stream and do not end in a decision about a person's rights. Check your process against the list: the more matches, the better the odds that a pilot will work.
If half the points do not match, the model is not the problem. The process is described and simplified first and automated second: the software the company already uses does not need to change, what changes is the order of steps and who is accountable.
In enquiries AI works out the topic and the substance, in documents it pulls out the fields, in calls it turns the conversation into text and agreements. These three areas produce most of the effect: that is where the unstructured stream arrives.
Everything that happens inside a CRM with records that are already filled in — stage changes, reminders, tasks triggered by events — runs without a model. Those scenarios are covered separately, in our article on CRM automation.
AI works unsupervised where a mistake shows up at the next step and costs little: a draft reply, a preliminary topic for an enquiry, a prompt for a manager. Decisions about money, rights and obligations are prepared by the system and taken by an employee.
That is a legal requirement, not a contractor's caution. Law No. ZRU-1115 of 21 January 2026 added article 7-1 to the law on informatisation: when taking legally significant decisions that affect a person's rights and freedoms, it is not permitted to rely exclusively on the conclusions of information resources and information systems built on artificial intelligence technologies.
The same law added part two of article 46-2 to the Code of Administrative Liability: unlawful processing of personal data using artificial intelligence technologies and its distribution in the media, telecommunications networks or on the internet carries a fine of 50 to 100 base calculation units together with confiscation of the items involved in the offence.
Important A model has no default "I do not know" mode: given an unfamiliar document or an odd question it will still produce an answer. That is why a scenario is built with a confidence threshold and an explicit route to a human, otherwise the mistake travels down the process in silence.
The result is counted per operation rather than across the company: take the process you automated and compare a handful of numbers before and after launch.
The measurement before launch is taken on the same operations and in the same form, otherwise there is nothing to compare against. The second reference point is frequency: rare operations give a small effect even at perfect accuracy.
One process is rolled out at a time, and every step ends with a result you can check.
How long each step takes depends on the shape the data is in, on whether the systems are reachable through an API, and on who inside the company reviews the doubtful cases.
A project usually stumbles over the way the task is framed, not over the model.
The last point is not an abstract risk: the duties of anyone who processes personal data are set out in law No. ZRU-547 of 2 July 2019. What exactly goes to an external service is agreed before launch, not after the first incident.
The price has two parts: a one-off setup and the charge for requests to the model. The one-off part depends on the number of processes, the number of integrations and the shape the data is in; the per-request charge depends on how much text passes through the model each month.
The other packages — a bot for requests and an AI assistant that answers from your materials — are listed with their starting prices on our AI and automation page.
Syntra Systems starts by working through the detail: what an employee's day is made of, which enquiries repeat most often and where data is moved by hand. The AI audit produces an automation diagram with the order of steps and a budget estimate for each; you can use it with us or on your own.
Example Not every task needs a model. At the SIMACH restaurant, bookings for dinners were covered by a Telegram bot running a scenario: a guest leaves a name and a phone number, the request goes straight to the manager, after the dinner the bot asks the guest to rate the evening from 1 to 5, and at a rating of 3 or lower the guest's number goes to the manager for a call back. No AI was needed here — the data arrives as ready fields. The details are in the SIMACH case study.
Where the input is free text, a document or a conversation, we add a model — and keep a route to a human and a log: what the system did, on what data, and who approved the result. Access is separated by role, and what goes to an external service is agreed before launch.
Let’s discuss your project
Tell us what you need, and we will estimate the timeline and cost and suggest a solution.
Standard operations within written rules, yes: identifying the topic of an enquiry, filling in a record, drafting a reply. Legally significant decisions that affect a person's rights and freedoms cannot rest on the conclusions of an AI system alone — law No. ZRU-1115 of 21 January 2026 prohibits it. The system prepares the option, an employee signs it off.
A few dozen real cases of one type is usually enough: the share of correct fields and answers without edits is counted on them. With fewer than ten examples a pilot produces a random result. The examples should be your own, from recent weeks, not from a contractor's demo.
They stay with people. A rare document type turns up a few times a month, and configuring for it costs more than handling it by hand. The system has to be able to say that a case is non-standard and pass it to an employee — that is what the confidence threshold is for.
Count the frequency. If an operation repeats several times a day, the effect shows even in a team of ten. If it repeats a few times a month, the time saved will not cover the setup and the charge for requests to the model: such tasks stay with people or get a simple scenario.
Before launch you decide which fields leave your systems at all: some data is anonymised, some replaced with identifiers, some processed on your own servers. The duties of anyone processing personal data are set out in law No. ZRU-547, and the set of data being transferred is recorded in writing.
A process audit answers the question of where the company loses time. An AI audit adds a second question: which of those areas can be covered by rules, which need a model, and what each option would cost. The output is an automation diagram and a rollout plan with priorities.
Setup is a one-off payment; every month after that you pay for requests to the model, and the bill grows with the volume of documents and conversations passing through the AI. That's why the pilot calculates the cost of processing a single document or conversation up front, and that figure is used to forecast the monthly budget.
