After two years of process reviews at companies in Astana and Almaty, we've almost stopped hearing the question "why do we need AI." It's been replaced by another one: "we already tried, so why didn't it take off." The chatbot on the website answered off-topic, the pilot with a contractor dragged on for eight months, the subscription to a foreign service stayed a toy for two enthusiasts. If you recognize yourself here, let's get down to business.
Below is the sequence of steps we've refined in practice. Four steps: pick a process, gather sample data, get a prototype, measure the effect. The whole thing takes about a month. Now, in order.
Why "adopt AI" isn't a task yet
The phrasing "we need artificial intelligence" answers not a single working question. Which process are we changing? Who owns the outcome? By what number will it be clear in three months that things got better? Until there are answers, a pilot is doomed to end with the words "we gained an interesting experience."
Compare two framings. The first: "we want to apply neural networks in sales." The second: "managers spend 35–40 minutes a day maintaining cards in the CRM, fill in a third of the fields, and the manager can't see the real pipeline." You can already work with the second one: there's a metric – minutes and card completeness, there's an owner – the head of sales, there are data sources – telephony and the CRM. A process review exists precisely for this: to turn the first framing into the second.
Which process should you start AI adoption with?
Not the most painful one, but the most suitable. A painful process is usually chaotic, overgrown with exceptions, and held together by the memory of two employees – a failure there will bury the whole automation topic for a year ahead. A good first candidate looks different:
- it repeats dozens of times a day, not five times a quarter;
- the result is measurable in minutes, tenge, or units;
- data is already accumulating: call recordings, documents, rows in the accounting system;
- an error is visible right away and is cheap to fix, and the person has a checkpoint;
- the process has an owner who wants a result, not an experiment.
The seven questions we use to select a candidate during a review are laid out in a separate checklist – with it, the process gets identified in about half an hour. Most often the first one turns out to be one of three: handling customer inquiries, reviewing sales-team calls, and primary documents in accounting.
For Kazakhstan there's a mandatory caveat – language. Customers call and write in Kazakh and Russian, often switching within a single sentence. A system that understands Kazakh only through machine translation loses both the meaning and the intonation on live calls. Test it on your own recordings: the platform must work with Kazakh and Russian at a native-speaker level.
Where AI already delivers results for Kazakhstan companies
In clinics, agents fill in medical records from the doctor's dictation, assign ICD-10 codes, and check cases before submission to the fund – fewer defects, fewer rejections during the fund's audits. In telecom and service companies, the platform reads churn signals in calls and tickets weeks before a cancellation request. In the quasi-public sector, it prepares meeting minutes and tracks the execution of assignments. In accounting, it processes source documents, including handwritten invoices. Industry scenarios are collected in the solutions catalog; the logic is the same everywhere: the agent embeds into the systems where the process already lives, rather than forcing the company to move to a new one.
What data you'll need at the start
Less than people tend to think. For a prototype of a single scenario, 30–50 call recordings, or 20–30 sets of documents, or a CRM export for one quarter is enough. There's no need to put your databases in perfect order: gaps and duplicates are exactly the kind of routine the agent will take over first. What exactly to gather for each scenario, we've laid out in our article on data for a prototype.
A separate question for the security team is where the data goes. Nowhere. The platform is deployed inside the company's own perimeter, every step the agent takes is recorded in an action log, and permissions are configured the same way they are for employees. Agents connect to systems through APIs; if an old accounting system has no API, the agent works through its interface – the same buttons and forms an operator uses.
Why a prototype matters more than presentations
Your executives have already seen slides about AI's capabilities, more than once. Not one slide answers the question that decides everything: will this work on our data. That's why we start not with a commercial proposal but with a prototype: one scenario, your real examples, roughly a week to set up. For the client the prototype is free – a week of setup is more honest and cheaper than three months of negotiations, and that goes for both sides.
Here's what it looks like in real life. A wholesale distributor in Almaty, a review on Thursday at 11:00. The commercial director describes the picture: the CRM is a third filled in, the pipeline for the planning meeting is assembled by hand. We take 42 call recordings and an export of 300 deals. The next Wednesday we show the result: the agent processed every call, completed the cards, and flagged nine deals where the customer had asked for an invoice and the invoice was never issued. The conversation changes immediately – now they're discussing not "do we believe it or not" but how to roll the scenario out across the whole department.
How much does AI adoption cost and when does it pay off?
The prototype is free. Bringing a scenario into production takes from eight weeks: connecting live systems, access rights, checkpoints, training the team. The platform runs on a subscription starting at 12 million tenge a year, and subsequent scenarios are included in it – which is why the second process costs noticeably less than the first.
Monthly effect = runs per day × minutes per run × cost of an employee minute × 22 working days + gain in the process outcome (deals not lost, defects removed, requests closed overnight).
An example with rounded figures. A contact center handles 400 inquiries a day, the agent closes 60% without an operator, an average inquiry lasts 6 minutes, and a specialist's minute including taxes costs about 40 tenge. Direct savings: 400 × 0.6 × 6 × 40 × 22 – almost 1.3 million tenge a month. The second part of the effect is usually larger than the first, it's just counted less often: overnight requests stop getting lost, defective documents get caught before they're sent. For scenarios chosen by the checklist, in our practice the subscription pays off in 6–9 months.
A week-by-week plan for the first month
| Week | What happens | What we need from you |
|---|---|---|
| 1 | Process review: we pick a scenario, fix the metric and the owner | A 60–90 minute meeting: the process owner, the doer, IT |
| 2 | Handover of sample data, prototype setup, clarifying questions | Half a day for the export, two or three 15-minute calls |
| 3 | Prototype demo, a control run on fresh data, error review | A 60-minute meeting, a "go to production or not" decision |
| 4 | Production rollout plan: integrations, permissions, checkpoints, timelines | Access and sign-off on the plan |
The worst outcome of such a month: you spent three meetings and half a day on an export – and got an honest answer on your own examples, not on someone else's case studies.
Frequently asked questions
Do we need a machine-learning specialist on staff?
No. Our team configures the scenarios – both at the prototype stage and after launch. From you we need a process owner who answers the "what's the right way" questions, and an IT specialist for a couple of hours to open access and run an export.
We're a mid-sized company, isn't it too early for us?
Go by the volume of routine, not the headcount. If repetitive operations eat up 200–300 person-hours a month or more, the economics start to work out. Run the numbers with the formula above or come in for a review with your own figures – we'll work them out together, and it doesn't commit you to anything.
What happens to the employees the agent frees up?
In our practice there are no layoffs after launch – what changes is the shape of the day. The routine goes to the agent, people take on the exceptions and the human conversations: the doctor sees two or three more patients, the manager gets through more calls. The process checkpoints stay with people in any case.