2026-11-12

AI Agent Implementation Cost: A Pilot Budget in KZT

Build an AI agent pilot budget in Kazakhstan: setup, runtime, human review and rework. Use a KZT TCO worksheet to separate capacity value from cash payback.

AI Agent Implementation Cost: A Pilot Budget in KZT cover illustration

The owner has agreed on a task for the pilot, but the proposal lists only “development” and “model access.” That is not enough to understand the AI agent implementation cost. Who will prepare the data, how much time employees will spend reviewing results, and whether the company’s cash outgoings will change are still unknown.

Calculate three amounts separately: setup, monthly operation, and the cash impact. Freed-up hours are useful, but they do not pay for the project on their own. Below is a calculation in KZT that you can copy into a spreadsheet and populate with your own data. Every amount in the example is hypothetical: this is neither GolemWorkers pricing, a market estimate of salaries, nor the result of a client implementation.

This article is for an owner or CFO of a company in Kazakhstan who has already selected one operation. If you have yet to define the task, start with the matrix for choosing your first process. If proposals promise different outcomes, first bring them to a comparable scope of work.

What the budget must include before you calculate payback

The initial cost goes beyond the developer’s work. The company also spends time documenting rules, preparing examples it is permitted to use, arranging access, and carrying out acceptance checks. Include these hours in the economic cost even if the pilot does not change payroll payments.

On a phone, scroll the table horizontally.

Category What to include What to clarify in the proposal
One-time setup Process analysis, data, configuration, integration, testing, training, and work by the client’s team The specific deliverable and acceptance criterion for each line item
Fixed costs Server, storage, logs, support, and an agreed scope of changes What is included, who is responsible for failures, and what costs extra
Variable costs Model usage, tools, recognition services, retries, and external services Cost across all incoming tasks, including unsuccessful ones
Human work Review, corrections, and manual completion of rejected tasks Time to reach a completed result, rather than merely produce an impressive answer

Do not automatically treat an employee’s ChatGPT subscription as payment for the agent’s API usage: ChatGPT and the API have separate billing [1]. API charges may account separately for tokens, tools, and storage; check the specific rates for the chosen setup [2]. Our example uses no currency conversion and is not tied to a model: the variable amount in KZT is an explicit assumption that must be replaced with a measurement or confirmed supplier terms.

For Kazakhstan, agree in advance whether each line item includes applicable taxes, which currency will be used for invoicing, and which exchange rate and date apply to any conversion. The same “KZT” label does not make a tax-inclusive amount comparable with a tax-exclusive one. This table is neither tax advice nor confirmation that data processing is legally compliant.

Three cost layers: initial setup, ongoing system operation and human review
Setup, system operation and human review belong to separate cost lines. Conceptual illustration, not a measured project.

One example: an email becomes a reviewed draft

Synthetic scenario. A trading company handles 600 requests per month in Russian and Kazakh. Preparing a draft manually takes 12 minutes per request. The pilot only prepares the draft: an employee reviews it, and nothing is sent to the customer. This example assumes no ready-made connection to email, a CRM, or 1C.

For the calculation, assume that 80% of drafts are usable without rework after a routine review. An employee spends 3 minutes on each such result. The remaining 20% are rejected: review and full manual completion take 12 minutes in total, rather than 12 minutes in addition to review. If a rejected task takes 15 minutes in your company, enter 15 in the table. Quality here means the share of usable tasks, not an abstract measure of “AI accuracy.”

On a phone, scroll the table horizontally.

Input Symbol Hypothetical value
Requests per month N 600
Manual time before the pilot, minutes per request t0 12
Share of usable drafts q 0.80
Review of a usable draft, minutes tr 3
Total time for a rejected task, minutes tf 12
Hourly value used to assess labour, KZT W 4,000
One-time setup cost, KZT I 900,000
Fixed monthly costs, KZT F 80,000
Variable cost per incoming task, KZT v 100

The hypothetical KZT 900,000 comprises 600,000 for external work, 200,000 for the client team’s involvement, and 100,000 for training and preparation for acceptance checks. This is an illustrative budget allocation, not a quote received from a supplier. For the cash calculation below, we make a separate conservative assumption: all 900,000 represents additional cash payments. In an actual budget, internal hours may not increase cash payments, so the cash outlay for setup and its economic cost can differ.

Calculate hours first, then cost, without double counting

Average human time after launch is the weighted time for the two outcomes:

ta = q × tr + (1 − q) × tf
H = N × (t0 − ta) / 60
V = H × W
O = F + N × v

Here, H is the hours freed up, V is their calculated value, and O is the additional monthly cost excluding the remaining human work. In this example, ta = 4.8 minutes, H = 72 hours, V = KZT 288,000, and O = KZT 140,000.

People still have work to do in the process. Before the pilot, that work took 120 hours; afterwards, it takes 48 hours, valued at KZT 192,000 under our assumption. Compare two valid approaches:

  • Monthly economic benefit: 288,000 in freed-up labour value − 140,000 in additional costs = KZT 148,000.
  • Full economic cost after launch: 192,000 in remaining labour + 140,000 to operate the system = KZT 332,000, compared with 480,000 before launch.

Both give the same result. Deducting 192,000 again in the first formula would be a mistake: review and rejected tasks have already reduced H. The 148,000 is the value of freed-up capacity less additional costs, not demonstrated cash savings.

Over six months, full economic total cost of ownership (TCO), including setup, is 900,000 + 6 × 332,000 = KZT 2,892,000. Without the pilot, the labour valuation is 6 × 480,000 = KZT 2,880,000. Even if the favourable assumptions remain unchanged, the six-month economic outcome is still KZT 12,000 worse than the baseline. A good monthly figure cannot hide an expensive setup.

Released time leads to more work capacity while a separate gate controls the path to cash
Available time and cash benefit are different measures. Cash realization needs evidence. Conceptual illustration.

When freed-up time becomes cash

If payroll remains unchanged and the time freed up does not yet generate verified additional margin, there may be no cash saving on labour. That does not make the 72 available hours worthless: they can be used to work through a backlog of requests or improve service quality. The effect simply needs to be described accurately.

Introduce r: the share of V that becomes a verifiable additional cash benefit, such as overtime payments or contractor costs that have actually been eliminated. For an additional sale, count the margin after associated costs, rather than revenue. Do not count the same benefit as both a cost reduction and an increase in margin.

B = r × V − O
Simple cash payback period = cash setup outlay / B, only if B > 0

At r = 0, the example’s cash result is −KZT 140,000 per month. At r = 0.50, it is 144,000 − 140,000 = KZT 4,000: formally, 900,000 / 4,000 = 225 months. This is a warning about weak economics, not a credible long-term forecast. At r = 1, the result is KZT 148,000 and payback is approximately 6.1 months after steady-state operation begins.

To achieve a positive monthly cash benefit in this example, more than 140,000 / 288,000, or 48.6%, of the value of freed-up time must be realised. Recovering the setup outlay over six months of steady operation requires an even higher amount: O + I/6 = KZT 290,000 per month. That exceeds the entire V of 288,000. Therefore, even r = 1 does not deliver six-month payback with these inputs.

This is a simple, undiscounted calculation. Development time, the ramp-up period, changes in workload, taxes, the cost of capital, and additional risks may lengthen payback. State the date from which you measure it: “6.1 months of steady operation” does not mean “6.1 months from the first conversation.”

Test what could change the decision: volume and quality

Only N and q change in the table below; the other assumptions remain the same. The variable cost of KZT 100 already includes average retries and stays constant for this test. In practice, poorer quality may also increase that cost, in which case both values must be recalculated.

On a phone, scroll the table horizontally.

Scenario N q Hours H Value V, KZT Costs O, KZT Cash B at r = 0.50, KZT
Low volume 300 0.80 36 144,000 110,000 −38,000
Baseline 600 0.80 72 288,000 140,000 4,000
More rejections 600 0.50 45 180,000 140,000 −50,000
Higher share of usable drafts 600 0.95 85.5 342,000 140,000 31,000
High volume 1,200 0.80 144 576,000 200,000 88,000

At baseline quality and r = 0.50, the monthly cash benefit becomes positive at 572 tasks: F / (r × (t0 − ta) × W / 60 − v) = 571.43, rounded up. If the denominator is zero or negative, increasing volume alone will not cover fixed costs. And 572 tasks is the threshold for a positive operating cash result, not recovery of the one-time setup outlay.

This sensitivity analysis suggests a practical decision: measure human time on usable and rejected results before commissioning a large project. A small change in quality can turn a nearly zero cash surplus into a loss.

A template for your company

Copy the input table above and add three columns to each row: source of the value, responsible person, and verification date. Leave unknowns as unknowns, rather than entering an optimistic zero. Keep a separate line for the cash setup outlay; do not substitute the economic value of internal hours for it.

You will need data from comparable Russian, Kazakh, and mixed-language requests. Do not derive an overall q from a handful of convenient Russian emails if other languages matter in the actual workload. Record the operating scope: drafts only or writing to a system, the number of sources, permissions, permitted actions, and the manual fallback. Checking the specific CRM/1C version and access is a separate task; no production integration is claimed here.

A general-purpose agent and reusable components may reduce repeated configuration work. For example, a verified mechanism for reading files or saving a draft can be used for the next task. However, each new process still requires checks of data, permissions, and quality. Ask the supplier to show in the budget which existing component will be reused and which work will actually be eliminated. Employee training must also deliver an outcome: staff should be able to review a draft, spot an error, and return to manual work. The mere availability of a general-purpose agent does not make I equal to zero.

Anthropic recommends increasing complexity only when necessary and highlights the trade-off between cost and quality in agentic systems [3]. For budgeting, this means paying for complexity that solves a measured problem, rather than for a separate “intelligence” for every job title.

Before starting, agree on the decision criteria: how much can be spent testing the hypothesis, which values need to be measured, and what result would trigger stopping the project. If correction time and the route to a cash benefit are still unknown, the next step is a bounded measurement exercise, rather than a promise of payback.

Discuss the calculation for your process

For an assessment, prepare one process, its monthly volume, the time spent on manual work, examples in the required languages, the systems used, and the permitted actions. Do not attach customer documents or secrets to an open enquiry: agree on a transfer method separately.

The contact form is in Russian, including when you follow it from this English article.

Describe your process to discuss a GolemWorkers pilot. The team can clarify the scope of work and the missing calculation inputs; pricing and terms are agreed after reviewing the task [4]. This article promises no fixed price, timeline, ROI, or compliance with any regulatory requirement.

Sources and calculation limitations

Checked on 12 November 2026. The formulas, tables, and all amounts are the author’s synthetic example; the sources do not substantiate these values.

  1. OpenAI: separate billing for ChatGPT and the API.
  2. OpenAI: current API pricing categories. No specific rates were carried over into the calculation.
  3. Anthropic: Building effective agents. Used for the principle of the simplest sufficient level of complexity, not as a current product catalogue.
  4. GolemWorkers: the approach to a pilot.