"How much does it cost?" is the first question at every demo. And it is the one question to which the Kazakhstani AI market has no short answer: quotes for one and the same request range from 5 million ₸ for a chatbot to 150–200 million for a custom system with a team on the payroll for a year. That spread is not down to vendor greed. Under the word "adoption" three fundamentally different things are being sold: a script, a project built from scratch, and a subscription platform. At AISOL we regularly see other companies' quotes: they bring them to our reviews so we can compare. In this article we lay out the two competing payment models and calculate the full cost of ownership over a three-year horizon.
The custom project: what you are really paying for
The pattern is familiar: the vendor surveys your processes, writes a spec, names a quote and a deadline. For a system at the level of "parse incoming documents plus integrate with the accounting system," the typical market range is 45–90 million ₸ and 6–12 months. The figures are not inflated. A team of four to six people – an analyst, two developers, an ML engineer, a tester – costs 6–9 million ₸ per month. Nine months of such a team is already 55–80 million in cost, and the vendor still needs to make a profit.
The more interesting question is where that time goes. In our reviews we see the same proportion again and again: 70–80% of the quote is eaten by the plumbing, not by the "intelligence." Integrations with the CRM and the accounting system, permission controls, task queues, failure handling, logging, control interfaces. How smart the underlying model is makes no difference whatsoever to these cost items.
Then comes what the quote does not mention. Support is usually 15–20% of the project cost per year. Every process change is a paid addendum. Language models replace one another every six months, while a custom system is nailed to the one that was current when development started: migrating to a new one is a separate mini-project for separate money.
And the main risk, which also does not appear in the quote: the project may never reach production. Six months of development, acceptance testing, and the quality on real data is not what it was on test data. The money, meanwhile, is already spent – all of it, not part.
The subscription platform: a different cost structure
In the second model, no one writes the system from scratch. The integration blocks, access rights, agent orchestration, and action log are built in advance and maintained by the vendor: that is what the platform is. What gets configured for a specific company are the scenarios: which documents to parse, under which rules, where to stop and call in a human. We have listed what comes ready out of the box in the description of the platform's capabilities.
An AISOL subscription starts at 12 million ₸ per year. It covers the work of the agents, integrations, model updates, and support for the scenarios in production. A prototype on your own task is configured in about a week and free of charge; production takes from 8 weeks.
The key difference shows up not on the first scenario but on the second. In the project model, the second process is a second quote: another survey, more integrations, more months. On the platform, the integrations and permissions are already in place, so the second scenario costs noticeably less than the first and launches faster. We devoted a separate payback analysis to this arithmetic.
Calculating the cost of ownership over three years
You cannot compare a project quote with an annual subscription directly: they are different horizons. The honest frame is the full cost of ownership over three years: launch, support, the second scenario, updates. Below are the average figures for a company of 200–500 employees, as we see them in the quotes at our reviews.
| Cost item | Custom project | Subscription platform |
|---|---|---|
| Survey and spec | 5–10 million ₸, 1–2 months | 0 ₸ – process review and prototype are free |
| Launch of the first scenario | 45–90 million ₸, 6–12 months | within the subscription, from 8 weeks |
| Support / subscription, per year | 8–15 million ₸ | from 12 million ₸ |
| Second scenario | 30–60 million ₸, new quote | configured within the subscription |
| Migration to new models | a separate project | included |
| Total over 3 years | 95–190 million ₸ | 36–48 million ₸ |
Even if you cut the project figures in half, the gap remains a multiple. It is not about the margin but about how the costs are built: in the project you pay to build the plumbing just for yourself, while in the subscription it is built once and its cost is spread across all the platform's clients.
The cost that appears in no quote
The cost of waiting. While the system is being written for nine months, the process runs the old way – and every month costs real money. The typical parameters from our reviews: a department of 12 people spends 3 hours a day per employee on manual routine. That is 720 hours a month. At a fully loaded hourly cost of 3,000 ₸, a month of delay costs 2.16 million ₸.
720 h/month × 3,000 ₸ = 2.16 million ₸ per month → over 9 months of development ≈ 19.4 million ₸ of savings that never happened
No one will ever pay anyone those 19 million: they simply evaporate. In the platform model, the first scenario runs after 8 weeks, and the savings start to accumulate instead of being deferred. When comparing options, this line is almost always forgotten.
A mini-scenario: two quotes on one table
A situation we have reviewed more than once. A manufacturing company, around 300 employees, wants to eliminate the manual entry of source documents: acts, invoices, waybills, reconciliation with bank statements. The integrator's proposal is 68 million ₸ and 9 months of work. The platform option is a free prototype on real scans in a week, production in 10 weeks, a subscription of 14 million ₸ per year given the document volume.
The CFO lays out the horizons. The project: 68 million up front plus support from the second year at 12 million – 92 million over three years. The platform: 42 million over the same three years, and a second scenario already fits inside that sum – for example, replies to suppliers about payment statuses. We have reviewed similar processes on the page solutions for accounting.
When a custom project is justified after all
The project model is not evil. It is appropriate when the process is unique and forms the core of the business: your own scoring, proprietary pricing, analytics no one else on the market has. If you have your own ML team and are ready to keep it running for years, build your own. For standard tasks – documents, calls, reporting, customer requests – the blocks already exist, and paying to build them again just for yourself makes little sense.
A separate case: the security team bans the cloud, and from this people conclude "custom development only." The conclusion is wrong. The platform deploys inside the company's perimeter, and the data never leaves it – we have described how the security model works on a separate page.
Three questions for any quote
Whatever your comparison concludes, these questions are worth asking every participant in the tender:
- What share of the quote goes to integrations, permissions, and infrastructure – and why are we building this from scratch if ready-made blocks exist?
- How much will the second scenario cost after the first? If it is the same, you are paying for a project, not for a platform.
- Who migrates the system to new model versions in a year, and at whose expense?
The answers usually make everything clear faster than the presentations do.
Frequently asked questions
What is included in the subscription from 12 million ₸ per year?
The work of the agents on the configured scenarios, integrations with your systems, language-model updates, the action log, support, and the configuration of new scenarios. Separate quotes for each next process do not appear in this model.
Can the quality be checked before signing the contract?
Yes. The process review and a prototype on your scenario are free, and the prototype takes about a week. You look at the results on your own data, and only after that do you discuss money.
How long does the subscription take to pay off?
Our benchmark, based on our calculations, is 6–9 months, provided the first scenario is chosen for its economics rather than for fashion: a frequent operation, measurable hours, an owner of the metric. Run the numbers on your own figures: the timing of an operation, its frequency, and the hourly cost give you an estimate in half an hour.