There is no shortage of writing about automation ROI, and almost all of it is Western. The problem is that its arithmetic rests on the cost of an employee-hour in Europe or the US. In Tashkent that number is different — and it changes the conclusions, though not always in the direction you would expect.
Below: how to do the maths on local numbers, and three worked scenarios. A caveat up front — these are calculation models with open assumptions, not reports on other people's projects. Every number can be swapped for your own to produce your own answer.
No one else's "450% ROI" deserves your trust, ours included. What deserves trust is your own four numbers: volume, share of routine, cost per hour, cost of the solution.
What makes the local economics different
Two factors pull in opposite directions, and both have to be accounted for.
Against automation: labour is cheaper. A specialist in Tashkent costs less than one in Europe, while the cost of a call to a language model is the same everywhere in the world. So the payback threshold is higher: a process that justifies automation at five hours a week in Berlin will need fifteen here.
For automation: fewer off-the-shelf options. The Western market is saturated with niche SaaS products for almost any task. Here they either do not exist or do not handle the Uzbek language and local specifics. Often the choice is not "custom versus off-the-shelf" but "custom versus manual work" — and that second comparison is far more favourable.
Plus one factor of its own: language. Bilingual support usually means a double set of staff. A single knowledge base covers Russian and Uzbek without a second team, and that is a saving absent from Western calculations entirely.
The formula
Annual saving = hours per week × the share the agent takes × cost per hour × 50
Payback in months = build cost ÷ (annual saving ÷ 12)
Cost per hour is calculated from the full cost of an employee — salary plus taxes, workspace, equipment — not from take-home pay. The difference is usually 30–40%, and omitting it systematically understates the benefit.
Scenario 1: collecting and qualifying leads from messengers
The task. The company finds clients manually: staff monitor relevant chats and channels, pick out relevant messages, note down contacts and transfer them into the CRM.
Assumptions: two people spending three hours a day each, full cost per hour of $5, the agent taking 70% of the work (selection and initial classification) with the rest verified by a human.
| Figure | Value |
|---|---|
| Hours per week | 30 |
| Agent's share | 70% |
| Annual saving | 30 × 0.7 × $5 × 50 = $5,250 |
| Build | $1,500 |
| Running costs | around $20–40 a month |
| Payback | roughly 4 months |
What is easy to overestimate here. Collected contacts are not the same thing as leads. If conversion from such a contact into a conversation is low, the time saving remains but revenue does not grow. What you count is hours saved, not "contacts collected".
Scenario 2: tender monitoring
The task. An analyst reviews tender portals daily, selects those matching the company's profile, reads the documentation, assesses fit and passes the shortlist to management.
Assumptions: one analyst, four hours a day, full cost per hour of $8 (higher qualification), the agent taking 60%. Less than in the first scenario: reading terms requires judgement and cannot be handed over completely.
| Figure | Value |
|---|---|
| Hours per week | 20 |
| Agent's share | 60% |
| Annual saving | 20 × 0.6 × $8 × 50 = $4,800 |
| Build | $1,500 |
| Payback | roughly 4–5 months |
The main benefit is not in the hours. A different effect is stronger in this scenario: suitable tenders stop being missed because the analyst was on holiday or ran out of time. A single won tender that would otherwise have been missed outweighs the entire time saving. But that is a one-off event, and putting it into the ROI figure would be dishonest — better shown separately.
Scenario 3: handling inbound enquiries
The task. Two operators answer customer questions in Telegram and Instagram. Most questions repeat. At night and on weekends nobody answers at all.
Assumptions: two operators at $500 a month (full cost around $650), the agent closing 70% of routine enquiries.
| Figure | Value |
|---|---|
| Time saving | 2 × $650 × 0.7 × 12 = $10,920 a year |
| Build | $1,500 |
| Payback on time alone | around 2 months |
The critical caveat. That $10,920 becomes real money only if you avoided a hire as volume grew, or moved people onto revenue-generating work. If both operators stayed and simply became less busy, you saved hours, not budget. More on this trap in measuring AI ROI.
What gets underestimated here: enquiries outside working hours. Their share is usually larger than expected, and right now they are not handled at all. That is not a saving but lost revenue, and it is calculated separately — through your conversion rate and average order value.
When not to automate
If the formula gives a payback of more than a year, the answer is usually one of three:
- the volume is too small — the process takes a couple of hours a week and automation will not pay for itself;
- the process is not described — you cannot name the share the agent would take, because you do not know what the process consists of;
- you picked the wrong process — there is another one nearby that hurts more.
All three are legitimate outcomes of the calculation. Finding out in an hour of arithmetic is cheaper than finding out over three months of a project.
Frequently asked questions
Where does "the share the agent takes" come from?
From a week's export. Tag your enquiries or tasks: how many are of one type, with the answer available in documents, requiring no decision. That share is the upper bound. In practice an agent reaches it not immediately but over several weeks of refinement.
Why not 90% in the examples?
Because 90% does not happen on real processes. Every one of them retains a portion of exceptions and judgement calls that a human takes. A calculation built on 90% is a calculation that will not add up.
Should we calculate in dollars or in som?
Whichever is convenient, but consistently. API costs are denominated in dollars and salaries usually in som — for an annual calculation that should be stated as an explicit assumption rather than papered over with a single exchange rate.
Should our own time be counted?
Definitely. Data preparation runs to five or six hours, then roughly an hour a week keeping the base current. Over a year that is a meaningful amount, and leaving it out makes the calculation more optimistic than reality.
What if the number is borderline?
Start with a pilot rather than a large project. A pilot $1,500 on a single process exists precisely for cases where the calculation gives no obvious answer: in two weeks you get actual figures instead of assumptions.
Where to start
Take one process and fill in four fields: hours per week, share of routine, full cost per hour, expected cost of the solution. Substitute into the formula above. If payback comes to less than six months, the process qualifies.
Want to run the numbers together on your own figures? We do a free process assessment and will say plainly if automation does not yet pay off.