How much AI implementation costs for a company

What makes up the cost of AI implementation

The cost of AI implementation is not the price of a “model”. It is the sum of the work around it. The model is only one part. For it to be useful, it has to be connected to your data and systems, tested on real cases and maintained after launch. That is why it is easier to read the budget by stage.

Below is the scope of work that the price is usually made of. Typical stage durations are listed next to each item. They show the scale, but they do not replace an estimate for your project.

• Discovery (3–5 business days): studying the process, the data and the constraints, setting the acceptance criterion, estimating the scope of work.
• Design (1–2 weeks): the solution architecture, integration points, and what a person does and at what moment.
• Development and integrations (2–6 weeks): building the solution for your process and connecting it to 1C, CRM and other systems.
• Testing on real data (2–4 weeks): running alongside the current process and comparing results with people's decisions.
• Launch (1–3 weeks): moving to the production environment, monitoring, a rollback procedure, team training.
• Ongoing support and improvement: monitoring, incident review, and changes when the process changes.

Ongoing costs: models, infrastructure and support

Development is the one-time part of the budget. But a solution with AI also has regular costs, and you should see them in the estimate before the start, not after launch. The first is calls to language models: the more requests or documents pass through the solution, the more model calls you pay for. The second is infrastructure: a server or cloud, data storage, monitoring.

Both items depend on the environment. You can deploy the solution inside your own infrastructure, or in a Russian cloud if you have none suited to the task. If personal data is involved in processing, the requirements of Federal Law No. 152-FZ (the Russian personal data law) are built in at the design stage, and this is also part of the scope of work. Some tasks can be solved with anonymized data, so it is worth deciding in advance which fields the solution really needs.

The third item is support. The process changes, new types of requests and documents appear, and the solution has to be retuned. Here it is convenient to use support packages measured in hours: ours are 10, 20 and 40 hours a month. Choose the volume based on how many changes you expect the solution to need.

Why the price is given after discovery

Asking “just name a price” in a first conversation is understandable, but an honest answer depends on what the contractor has not yet seen. A request to “automate request processing” can mean one queue in one channel. Or it can mean several sources, varied documents and several systems where the result has to be recorded. The difference in scope is fundamental.

There is also the other side. A good discovery can end with the conclusion that AI is not needed here: some tasks are cheaper to solve with a clear procedure, standard automation or an off-the-shelf service. A price given before such a review will either include a large margin or have to be revised during the work.

To get a more accurate estimate, prepare answers to the questions below. The price depends on them most.

• How many systems need to be connected, and how open they are to integration: 1C, CRM, email, messengers.
• What form the data is in, and whether you have examples of real cases, including rare and non-standard ones.
• Where the solution will run: your environment or a cloud, and whether personal data is involved.
• What the solution does on its own, and what it hands to a person and with what context.
• How the result will be accepted: what counts as “it works”.
• How quickly access is granted, and who makes decisions on your side.

How to lower the cost of the first step: a pilot on one process

The most reliable way to limit the budget of the first step is not to automate everything at once. Choose one process and one part of it: for example, sorting incoming requests or extracting data from documents of one type. Then you connect fewer systems, collect fewer examples and agree on one acceptance criterion.

Incoming requests show this well. While the process is manual, an employee reads the request, works out what kind of request it is, looks up data in several systems, finds the person responsible and enters the data into the CRM by hand. If AI handles classification, data extraction, routing and entry into the CRM, the person is left with oversight and the decision on the substance.

The pilot runs on your real data, not on a demo scenario. The solution is launched alongside current work: people keep working the old way, the solution runs next to them, and you compare the results. In the end you can see where it copes, where it makes mistakes, and whether to extend it to neighboring areas. You decide on the next step based on facts.

A pilot is not a stripped-down version for show. It is a complete piece of work with a clear result. So put the acceptance criterion in writing before work starts: for example, in which cases the solution's output must match the employee's.

Turnkey AI implementation: questions for your contractor and common mistakes

AI implementation proposals are hard to compare by the total: one includes integrations and support, another only a prototype. If you are offered turnkey AI implementation for your business, ask what exactly is included in “turnkey”. Compare the scope of work, not the figure.

Common mistakes look like this: choosing the lowest price without checking the scope of work; accepting work without an acceptance criterion; not budgeting for models and support; starting with the most complex process instead of a small one where results show quickly; testing the solution on prepared examples instead of real data. To avoid them, ask your contractor the questions below.

• What exactly is included in the price: discovery, integrations, testing on real data, launch, team training?
• Which costs are charged separately: model calls, infrastructure, third-party service licenses?
• How is the acceptance criterion set, and what happens if the solution does not meet it?
• Who owns the code, configurations, prompts and the model's operating rules?
• Where will the solution run, and how are the requirements of Federal Law No. 152-FZ met if personal data is involved?
• What support mode is offered after launch, and what does the contract fix?
• What does a person do, what does the solution do, and how are errors tracked?

Our price guidelines and where we start

Here are our guidelines. AI projects start at $2,100. AI agents and assistants start at $2,400 and 3 weeks. Quick start, an AI assistant on one channel, starts at $1,600 and 3 weeks. A first prototype on your data takes from 3 weeks, and from 4 weeks for 1C (a Russian ERP platform) and CRM integrations. These are lower bounds. We give the final figure after discovery.

Software product development that includes AI features is priced by estimate after discovery. Support after launch comes in packages of 10, 20 and 40 hours a month for $400, $730 and $1,300. Hours beyond the package cost $60 each. The support mode and SLA are fixed in the contract. Model calls and infrastructure are billed separately.

We start with a free discovery of one process. It takes 3–5 business days and ends with a written document: a process map, the chosen area to automate, the boundaries of the solution, an acceptance criterion, an estimate of the scope of work and a list of risks. If the task can be solved without AI, we say so plainly. The document stays with you, and you are not obliged to continue working with us.

Next, a pilot on real data is possible, with the acceptance criterion fixed in advance. If the solution does not meet it, we improve it at our own expense. After launch, you can keep support with your own team or hand it to us. The terms are put in writing.