How to choose a reliable AI Implementation contractor: open data verification

The AI contractor can be verified before signing the contract.

We work in this field ourselves, so we offer an honest framework: all the criteria below apply to us. Good instructions on choosing a contractor should not be tailored to a single contractor. It should help the customer distinguish a real engineering team from a beautiful landing page, understand the boundaries of the project and not buy "AI" where the setup of someone else's designer is actually sold.

Briefly about SoftRest:AI agentsinternal systems

AI implementation services in business: who provides them today

The AI market looks big, but completely different performers live under the same banner. According to AIANA estimates for August 2026, the Russian AI market by the end of 2025 amounted to 316.1 billion rubles and grew by 29.7%; there were 883 companies in it, but only 386 supplied finished products for the final task. It is important to understand that the largest segment by revenue is AI equipment, 31.1%, and not the implementation of application solutions in business processes.

Therefore, the phrase "we are working in the AI market" does not prove anything by itself. One contractor sells a cloud platform, the second implements a voice robot, the third writes an agent for a specific process, the fourth resells the designer under the guise of development. The customer does not need to compare the volume of promises, but the type of performer and his compliance with the task.

The CNews Analytics rating of the largest players in the Russian AI solutions market for 2025 shows a different cross-section: 60 companies, 85 billion rubles of total revenue from AI projects, the top 15 - 75.6 billion rubles. This is a useful picture of the upper segment, but it does not answer the question of small businesses: who should be entrusted with the first application project for hundreds of thousands of rubles, and not a corporate transformation for years.

An AI agency, an AI bot development studio, or an integrator: what's the difference?

The terms on the market have not yet settled down. """"""""AI bot development studio", ""AI agency", "integrator" and "platform vendor" may look the same in the advertising output, but inside they are different models of work and responsibility.

AI agencyAI Bot Development StudioThe IntegratorPlatform Vendor

The practical conclusion is simple: the word in the name does not guarantee anything. It is important who is responsible for the result under the contract, how this result is measured, who owns the code, prompta and knowledge base after payment, as well as what happens if the model, platform or contractor becomes unavailable.

What do contractors call turnkey AI implementation — and why is it different every time?

The phrase "turnkey AI implementation" sounds convenient, but it is dangerous precisely because of its breadth. One contractor means setting up a ready-made platform on your documents. The other is to develop a solution for a specific process with integrations. The third one includes the preparation of a knowledge base, employee training, support regulations, and post-launch support. All three options may be fair, but these are three different projects.

the knowledge base

If the KP does not say who pays for the operation of the models, this is not an estimate.stages of implementation

How to distinguish a real AI agent from a renamed chatbot

agent washing

The difference is easier to explain through behavior. RPA follows predefined rules and does not adapt. The chatbot or assistant responds to the request and waits for the next one. The agent gets a goal, chooses the tools himself, plans the steps, checks the result, and reschedules the actions if unsuccessful. If the system only responds with text based on the knowledge base, it can be a useful assistant, but not necessarily an agent.

Four questions quickly reveal the substitution. Ask them to show a trace of the real dialogue: which tools the agent called, what they returned, and why the next step was chosen. Ask how the agent verifies that the action was actually performed, and not just the API was called. Specify what happens when the tool fails. Finally, ask them to show a behavior on a question that is not in the script.

There is also an honest counter-question to the customer himself: do you need autonomy? The agent is more expensive to develop, operate, and control. For many small business tasks, an assistant with a knowledge base and the transfer of dialogue to a person is enough. Selling an autonomous agent where an assistant is enough is also a form of agent washing, but not a technological one, but a pricing one.

A good free test demo on your data. If the contractor only shows slides and a sample scenario, you are seeing a presentation, not a solution. Teams that really know how to make applied AI systems usually quickly move on to one specific painful process: they take your documents, show you the first answers, discuss metrics, and only then talk about scaling.

How to check an AI implementation company using open registries

Contractor verification does not begin with a call, but with open data. They won't tell you how talented the team is, but they will help you eliminate obvious risks: a dubious legal entity, a mismatch between the copyright holder, confusion between accreditation and the software registry, and unsubstantiated license claims.

The accreditation of an IT company and the entry of a product in the Russian software registry are two different things. The presence of a product in the registry does not automatically grant the company accreditation. In the software registry, you need to verify the exact name: the commercial name may differ from the registry name, and the origin is confirmed by the serial number of the entry and the copyright holder.

The customer is often confused with FSTEC licenses. An ordinary AI agent developer does not need them. A license to develop and manufacture security products is required if the product itself is a means of protecting information. A TKKI license is needed if the contractor designs secure systems, implements and configures security features, monitors security, or prepares the facility for certification. If the contractor simply writes an agent, there is no need to request a license. If he deploys an agent in a secure loop and takes over the protection of information, the conversation is different.

See the VAT in the invoice separately. Starting in 2026, the total rate is 22%. The privilege for transferring rights to software from the registry applies to the object, not to the status of the company: custom development, implementation, technical support and training usually go in separate lines and are taxed differently. The different rates in KP are not necessarily a trick; sometimes it is a sign that accounting understands the subject.

What to ask before ordering the implementation of AI: nineteen questions

Most of the failures are not due to fraud, but due to a mismatch of expectations. The customer buys "AI", the contractor sells the development, but no one has recorded what is considered a working result. Therefore, questions should test not only integrity, but also measurability.

About the task

1. Which process do you suggest automating first, and why choose it?
2. What metrics of this process do you want to see before starting? A good answer is a specific list, not a "demo" one.
3. What do you consider the pilot's success and in what numbers?

About technology

4. What models will you use and where is the inference physically performed?
5. What happens if the provider of the model becomes unavailable or raises prices?
6. Do you use third-party platforms and constructors?
7. How do you measure the quality of responses and how often?
8. What does an agent do when he doesn't know the answer?
9. Who cleans and sorts the documents before uploading them to the knowledge base, and is this included in the price?
10. How does the knowledge base work, in what format will the customer receive it, and will they be able to edit it without a contractor?
11. What happens when an agent transmits a dialog to a human, and how does the system decide what to transmit?

The last three questions are particularly important. According to practitioners, a significant part of the quality problems lives not in the model, but in the data: headers and footers, page numbering, recognition artifacts, and hidden characters from Word end up in the knowledge base and spoil the search. If there is no cleaning of documents in the estimate, you pay for a system that will confidently respond to garbage.

About money, risks and rights

12. What is included in the price and what is paid separately?
13. Who pays for the work of the models after the completion of the project and what is the limit?
14. How much does post-launch support cost and what does it include?
15. Who owns the code, prompta, and knowledge base after payment?
16. What will the customer receive upon termination of cooperation?
17. personal data
18. Who is responsible if the agent gives the client incorrect information, and what is written about it in the contract?
19. What changes for the project from March 1, 2027?

The eighteenth question shows whether the contractor considered the risks at all. The Nineteenth is checking whether it is following the regulation of the industry: from March 1, 2027, the rules related to the notification of users about the rights to the results of the application of large fundamental models will come into force. In practice, this may mean editing user agreements and interface notifications. The contractor, who cannot say anything about this at the end of 2026, is not following the topic closely enough.

AI Development Services Contract: three formulations that solve everything

The first formulation is the right to the result. If the subject of the contract explicitly provides for the creation of a program or database, the logic of Article 1296 of the Civil Code of the Russian Federation works: the customer has the default right, unless the contract provides otherwise. If the program was created while fulfilling a contract that did not explicitly provide for its creation, there is a risk of Article 1297 of the Civil Code of the Russian Federation: the contractor may retain the default right.

Therefore, the subject of the contract should explicitly speak about the creation of a computer program or other specific result. The wording "providing AI implementation services" is too vague. As part of the transmitted result, you need to list the source code, prompta, configurations, knowledge base datasets, integration schemes, deployment instructions, and a list of third-party components with their licenses. Promptness and knowledge base are often the main value of an AI project, but they are the ones that are forgotten to explicitly convey.

An important caveat: the legal status of the prompta as an independent work remains a gray area. Therefore, it is safer not to argue about the nature of the object, but to specify in the contract that prompta, knowledge base, scripts, routing rules and test suites are included in the transmitted result. It is also worth separately prohibiting training based on customer data, if this is critical for you.

The second formulation is personal data. If the contractor processes your clients' data, you need an order for processing in accordance with Part 3 of Article 6 FZ-152: with the obligation to comply with the principles of the law, confidentiality and security measures. The key thing is that you remain responsible to the customer as an operator. The contractor may violate, but the claim will often come to the company that installed the bot on the website or service.

acceptance criteria

A separate line is the exit conditions. The contract must include the export of the data in a workable form, the return or deletion of the data within the agreed period, the distribution of the deletion to subcontractors, and written confirmation. Owning the source code by itself does not save you from being tied to a supplier: you need a provider-independent layer, portable prompta, quality tests, and a proven project transfer scenario.

How much it costs: market forks and what should alert

There is virtually no independent study of AI implementation prices in Russia. Contractors' public price lists and their own articles are available, that is, the lower limits stated by the sellers, rather than the median of actual transactions. Therefore, the only protection of the customer is not the "average market price", but the pilot, the contract and a transparent estimate.

someone else's constructor

The most reliable alarm signal is the absence of a line for operation. Models, storage, infrastructure, monitoring, knowledge base updates, and support cost money after launch. If there is no support in the estimate, it will appear later and suddenly. It is better to analyze the price structure in detail separately; there is a rule here: a price without operating conditions does not show the full cost of ownership.

What does a normal first project look like?

process metrics

The pilot's working format is 2-6 weeks before the first result and the budget is in the range of hundreds of thousands of rubles, not millions. The success criterion should be recorded by the number before the start: for example, the proportion of offline resolved requests, reduced processing time, increased classification accuracy, or reduced manual operations. The stop criterion is also needed in advance: if after six weeks the increase is less than the agreed threshold, the project is closed or reassembled.

There is a practical threshold of applicability: one or two repetitive processes with clear logic and at least 50-100 events per week — applications, applications, resumes, documents. Below this volume, AI often does not have time to produce a noticeable effect, even if it technically works. Less than 30 days is usually not enough data for conclusions, and more than 120 days the pilot begins to turn into a project without a host.

The pilot is needed not to "try AI", but to get a figure in one significant process, according to which a decision is made: scale, redo or stop. That is why an honest stop is not a failure, but a normal part of risk management.

AI companies in business: how to put together a short list of three

Start with 5-7 candidates of different types: platform vendor, studio, integrator, specialized conversational AI team. If you compare only similar landing pages, you will get similar promises. Different types of contractors will give different answers to the same request, and that's what's useful.

Next, sort out the candidates by registry and legal grounds. Then send everyone the same short task description with the original metrics. Do not change the introductory questions from call to call: the differences in responses will be the main material for comparison. Remove those who do not ask counter-questions about the process, data, limitations, and acceptance criteria.

Ask the rest to show a demo on your data. Not a full-fledged project for free, but a short test of the ability to work with your material: documents, sample questions, sample applications. After that, you don't compare presentations, but rather contracts: rights to the result, personal data, responsibility of the model, exit conditions, and support costs.

The final short list of three should include not "the most convincing sellers", but those who passed an open check, honestly named the limitations and were able to show how the result would be measured. It is better not to sign a large contract right away: start with a pilot for one contractor and write down in advance the condition under which the project is considered successful or closed.

What to do next

If you are choosing an AI contractor, do not start by asking "how much does an agent cost". Start by checking who will be the legal executor, whether there are verifiable products and competencies, what is written in the contract on rights, data and responsibilities, whether the contractor is able to show the work on your data, and not just on the presentation.

an assistant from an agent

Minimum checklist before the contract:

Автор Антон Амосов

Поделиться Скопировано!