The fate of the entire rollout is decided at the moment people think about least: choosing the first process. In review after review, we see the same mistake: a company picks its most painful area. The logic is clear – treat it where it hurts. The trouble is that the most painful process is usually chaotic, riddled with exceptions, and leaves no data behind. A failure there shuts down the whole topic of AI for a year or two: "we tried it, it doesn't work for us." The cost is higher than it looks: rallying a team for a second "let's try again" is far harder than the first time.
What you need first is not the most painful process, but the most suitable one. Below are the seven questions we ask in every review. Take three or four candidate processes, run them through the list, and score a point for each "yes." It takes about half an hour.
Question 1. Does the process repeat at least 20–30 times a day?
The economics of an agent are built on volume. A task that comes up 15 times a month won't even pay back the setup, no matter how unpleasant it is. But 300 requests a day or 60 sets of documents give you a basis for the math. Frequency matters more than complexity: a simple high-volume process delivers more impact than a complex rare one.
Question 2. Is the result measurable in minutes, tenge, or units?
"It'll be more transparent" and "we'll take load off people" are not metrics. A metric is "response time dropped from four hours to ten minutes," "the share of completed records rose from 30 to 95%," "the amount clawed back on defective cases was cut in half." If you can't name the number before the start, the argument about the result after the start will never end. One metric. Two at most.
Question 3. Is the process data already accumulating somewhere?
The phone system records calls, email stores correspondence, the accounting system holds entries, the clinical information system holds appointments. If traces of the process settle somewhere, the agent has something to learn from. Perfect cleanliness isn't required: 30–50 real examples are enough for a prototype. It's worse when a process lives entirely in verbal agreements and messengers on personal phones – a candidate like that goes to the back of the queue.
Question 4. Is a mistake easy to spot and cheap to fix?
A good first process is set up so that an agent's slip is visible at once and costs pennies: a draft reply before sending, a pre-filled record that a person confirms. A bad one is where the error goes to the customer or into a report to the fund without review. A checkpoint for a person is built into the scenario from day one, and every step the agent takes is written to an action log – you can always trace why a decision was exactly what it was.
Question 5. Does the process have an owner who needs the result?
Not an IT curator and not a "working group," but a manager who lives with this metric: the head of sales, the chief accountant, the head of a department. Pilots without an owner are the first of five reasons projects never reach production. If no one stands up for a process during the review, we don't take it, however promising it looks on paper.
Question 6. How much does one run of the process cost today?
Estimate it roughly, on a napkin. The formula is short:
Cost of the process per month = runs per day × minutes per run × employee rate per minute × 22 working days.
Say accounting enters 60 documents a day at 12 minutes each, and a specialist's minute costs 45 tenge: 60 × 12 × 45 × 22 – about 710 thousand tenge a month on data entry alone. Then compare that with the cost of automation. If the number impresses neither you nor the director, that's an honest "no" for this candidate.
Question 7. Will the effect be visible within a month or a month and a half?
The metric has to move while the team still has attention and belief. Processes with a quarterly cycle – annual budget approval, seasonal purchasing – make poor first candidates: by the time the effect can be measured, the pilot has already been forgotten. Daily processes show movement in the second week, and that movement sells the project inside the company on its own.
What a review looks like live
A clinic in Astana, a meeting room, Wednesday, 3:00 p.m. At the table: the head physician, the deputy for quality, and an IT specialist. The pain is named right away: after a shift, doctors sit for 40–50 minutes finishing patient records, and the fund returns some cases over paperwork defects. We run it through the seven questions. Frequency – yes: 25–35 appointments per doctor a day. Metric – yes, two even: minutes per record and the amount clawed back during the fund's audits. Data – yes: appointment audio can be recorded, and record templates and ICD-10 codes sit in the clinical information system. Checkpoint – yes: the doctor confirms the entry before saving. Owner – yes: the head physician, both numbers are in their reporting. Cost and speed of effect we work out right on the whiteboard. Twenty-five minutes – the candidate is found. And this isn't "all of AI in medicine," but one concrete scenario from the solution for clinics.
Example: three candidates from one review
| Process | Frequency | Metric | Points out of 7 | Verdict |
|---|---|---|---|---|
| Replies to routine requests | 320 a day | share closed without an operator | 7 | first prototype |
| Entry of source documents | 60 a day | minutes per set | 6 | second scenario |
| Preparing commercial proposals | 15 a month | no metric owner | 3 | postpone |
The picture is typical: the winner isn't the most-discussed process, but the high-volume, measurable one. Meanwhile the third candidate isn't "bad" – it's just not first.
What to do with the result
Six or seven "yes" answers – the candidate is ready for a prototype: we set it up on your data in about a week, for free; how that week works we described separately. Four or five – the second tier: come back once the first scenario pays off. Three or fewer – the process first needs order in its own steps: automating chaos only speeds up the chaos.
If you're torn between two candidates, come to a review with both – in an hour we'll go through the questions together and calculate the economics of each.
Frequently asked questions
Not a single process scored six "yes" answers. What next?
Most often it's the metric and the owner that fall short – that's fixed in a week: assign someone responsible and agree on a number. If the data falls short, start accumulating it now: turn on call recording, save scans to a single folder – in a month you'll have enough examples for a prototype.
Can you launch two processes at once?
Technically – yes, the platform runs dozens of scenarios in parallel. At the start we still advise just one: the owners' attention is the scarcest resource of a pilot. The second scenario connects faster and cheaper than the first, once the integrations and permissions are already built.
Who from the company should be at the review?
Three people: the process owner, a strong doer who does the work by hand and knows the exceptions, and an IT specialist who can say where the data lives. A 60–90 minute meeting is usually enough to walk out with a chosen candidate and a plan for the prototype.