AI-Native Software Engineering Studio
AI agents for business — an employee for a single process.
Not a chat-window neural net.
We don't sell AI. We take one of your operational processes — leads, tenders, support, documents, reporting, engineering — and build an agent for it. We connect it to your data, set the rules, test on real load, hand it over to the team and maintain it.
8 directionsSales, Personal, Research, RAG, Support, Backoffice, Developer, Tender — we have deployed agents in each.
Pilot from 250K RUBPrice depends on the agent. Tender is harder — from 450K. Sales is simpler — from 250.
Enterprise is a separate projectNot a «more expensive tier», but an infrastructure contour: customer servers + local models + fine-tuning.
When an AI agent fits — and when it doesn't.
Honestly speaking, an agent doesn't pay off in every process. So the first thing we do in any conversation is check whether yours is the right case.
An agent works when four conditions all line up.
- 01
the process repeats — daily or weekly — requests, tenders, inbound emails, statuses, reports, customer enquiries.
- 02
data is accessible — the process already lives in your systems: CRM, 1C, email, chats, drives, spreadsheets, corporate APIs.
- 03
there is an owner — a person responsible for the process who can review the agent's output.
- 04
«correct» is defined — what counts as «done correctly» is clear — so you can measure whether the agent helps.
If all four are present — the pilot pays off on the first scenario. If even one is missing — we help assemble it first, and only then discuss the agent.
What we DON'T propose as an agent
Just an FAQ without system integrations — a regular scripted bot is enough here, cheaper and faster.
Unique one-off tasks — if the process is new every time, the agent has nothing to learn from.
Solutions without an owner — without a person who can say «this is right, this isn't», the agent becomes a noise generator.
Fully autonomous actions in production — critical steps always require human confirmation. Not a compromise, an engineering principle.
What we usually do next
If you recognized your situation in the first list — we move on: look at which of the eight directions is yours. If in the second — we'll say so on the first call.
Better to lose an hour now than three months on a pilot that shouldn't have started.
Eight directions, in each — from the inside
This isn't «eight types of AI agents». These are eight working contours in each of which we have already deployed and maintained agents. The pilot price for each is different — it reflects the real complexity of integration, scope and risks of the process.
Sales Agent
The lightest start. Takes one sales process: requests come in from different channels, the agent qualifies them, gathers customer context, updates the CRM, drafts follow-up. The sales lead gets a clear funnel picture without manual assembly. We connect to amoCRM, Bitrix24, HubSpot and one inbound channel.
Personal / Executive Assistant
For an owner or a key employee. Morning briefing, meeting context, follow-up on commitments, management reports. We connect to the calendar, email, messengers and CRM. Decisions stay with the human — the agent only prepares context and drafts.
Research / Market Intelligence
Monitors competitors, news, prices, market shifts across agreed sources. Prepares a weekly digest with conclusions «what changed and why it matters» — not just links, but short conclusions for decisions.
Knowledge / RAG Agent
Turns corporate documents into working search. An employee asks — the agent finds the answer in regulations, contracts, instructions, with a link to the source. Respects access rights, doesn't expose other people's documents.
Support Agent
First-line support: closes typical questions itself using the knowledge base, hands the hard ones to an operator with prepared context. Helps the support lead see not ticket counts, but where customers' real pain is.
Operations / Backoffice Agent
Requests, documents, acts, statuses, approvals. The agent processes inbound, prepares tasks, watches for overdue items, keeps an action log. Connects to 1C, EDI, task systems, email.
Developer / Delivery Agent
Engineering Governance: gives the owner and CTO development transparency without micromanagement. Project memory, weekly brief, release risks, architectural decisions captured from chats.
Tender Agent
The hardest to launch — and you can see it in the price. Monitors tenders on procurement platforms, reads the RFP and draft contract, highlights risks, prepares a bid/no-bid recommendation. Remembers your company profile and the history of past decisions.
Three levels of engagement — not three pricing tiers.
At most studios this is «light / standard / premium». For us it's different — three distinct project types by complexity of the task and format of responsibility.
Pilot — minimum working contour
Around one of your processes. Not a demo prototype — it actually works on your data, produces a result you can check by hand.
What the pilot includes: one process, one or two key integrations, basic logic, testing on your real cases, training one person to work with the agent. What it doesn't include: the entire CRM, all channels at once, a large knowledge base, autonomous actions without approval.
The pilot price depends on the agent: Sales — 250K, Tender — 450K. That's not marketing — it's the real difference between connecting to a CRM and dealing with state procurement systems and tender documents.
The pilot's goal is not to sell you a full implementation. The goal is: in 2–3 weeks you understand whether the agent pays off in your process. If it doesn't — we'll say so.
Full implementation — agent for the department
Once the pilot proves value, the agent is rolled out across the whole department. It's not «we do the same thing but for more money». It's a qualitatively different project:
- ▪We connect all data sources, not just one
- ▪Multiple scenarios for different team roles
- ▪Recurring tasks, reports, reminders
- ▪Roles and access permissions
- ▪Training the team to work with the agent, not just one person
- ▪Process support — when something changes in the business, the agent adapts
Most often the implementation is deployed on our infrastructure — SaaS-style. It's faster, cheaper and easier to maintain. If your security policy requires something different — we have three other formats, see the next chapter.
Enterprise — this is an infrastructure project
A price list doesn't work here. First we do a technical scoping: what infrastructure you have, what data requirements, which models can be used, who's responsible for security, what SLA is needed.
Enterprise isn't «a more expensive version of the pilot». It's:
- Serverspurchased or provided by you
- Deploymentin your perimeter, on-prem or private cloud
- Local modelsours or yours; we can fine-tune on your data for quality and response speed
- Controlroles, audit logs, monitoring, backups, SLA in the contract
- Integrationsoften several internal systems — ERP, DWH, EDI, billing
Price is calculated after technical scoping. Not because we hide it — but because an enterprise contour at a bank and at an industrial holding are different engineering problems.
Footnote · maintenance
From 40K RUB/month — because the agent lives in production
Fair question: you paid for the pilot, paid for the implementation, the agent works — why pay every month on top? Because the agent isn't a static website. It lives in three constantly changing environments:
- Your business changes — a new customer category, the sales lead changes, a branch opens. Without maintenance the agent will be running on old rules in 3 months and the team starts bypassing it manually.
- The models change — OpenAI ships a new version, the old prompt breaks. Anthropic improves quality, we switch. Every 2–3 months something changes on the model side, and if you don't watch — quality drops silently.
- The infrastructure changes — if the agent runs on our SaaS contour, we update servers, apply security patches, renew certificates, monitor uptime. The customer shouldn't be doing this.
Included in 40K RUB/month
- · SaaS infrastructure (if hosted with us): servers, updates, backups, certificates, uptime
- · DevNeuroX local model within the agreed limit
- · Monitoring, incident analysis, service restarts
- · Adjusting prompts and scenarios for process changes
- · Adaptation to model API changes (new versions, new billing)
Billed separately
- · External model tokens (OpenAI, Claude, Gemini) — paid directly to the provider
- · Paid SaaS model subscriptions (GPT-4, Claude Opus, GigaChat Pro)
- · Major changes — new integration, new scenario, new role
- · Enterprise infrastructure (if on-prem) — servers and their administration on the customer side
Maintenance covers everything that should just work «by default» so that in 6 months the agent doesn't become a ritual artifact.
Eight agents surrounding one business
Each agent owns one process. Left — those working with the outside world: customers, leads, the market. Right — those holding internal processes. All eight can run in parallel — start with one.
Your business
CRM · ERP · EDM · your data and processes — agents plug into what already works for you
- One pilot in 2–3 weeks — no commitment to the whole stack
- On our SaaS infrastructure or on your servers
- Local models for sensitive data
- Bilingual — RU/EN from one interface
- From 250K RUB for a pilot, from 40K RUB/month for support
You don't need to deploy them all at once. Take the process where the pain is sharpest — a 2–3 week pilot will show real impact on your data.
Four formats of infrastructure.
When people say «deploy an AI agent» they usually imagine one thing — either «in your cloud» or «on our server». In practice there are four options, and they differ not in «expensive/cheap» but in control level, launch speed and the maturity of the customer's IT team.
On our infrastructure
- For whom
- SMB. Most pilots and first deployments.
- How it works
- Servers, backups, monitoring, certificates, updates — all on our side. External LLM/API costs are paid by the customer directly or capped by limits.
- Pros
- Fast start. Low upfront cost. Simple maintenance.
- Limitations
- Not suitable for strict security requirements — banking, government. Some data may pass through external services (subject to agreement).
Dedicated cloud environment
- For whom
- Mid-market with sensitive data, but without a hard on-prem requirement.
- How it works
- Dedicated cloud environment — isolated access, own storage and logs. Can use paid external models or local/private ones.
- Pros
- Balance of control and speed. Suitable for most production deployments.
- Limitations
- More expensive than format A. Requires cloud environment architecture sign-off.
On your servers
- For whom
- Companies with a mature IT perimeter and their own infrastructure.
- How it works
- The customer provides servers, access and network. We deploy the agent inside. Responsibilities are fixed in advance.
- Pros
- Maximum control over data. Easier to pass internal security requirements.
- Limitations
- Harder to coordinate. Depends on the customer's IT processes. Slower launch.
Fully closed perimeter
- For whom
- Large enterprises, regulated industries, projects with closed data and a ban on public models.
- How it works
- Servers are purchased by the customer. Agent in a closed perimeter. Only local models — ours or fine-tuned. Roles, audit logs, SLA, monitoring, backups.
- Pros
- Full control. Compliance with internal security requirements. Model fine-tuning for the customer is possible.
- Limitations
- More expensive and slower. Requires the customer's IT team. Separate technical design phase.
Cheatsheet · how to pick a format
In reality we discuss the format on the first call — after we understand the process and requirements. Not the other way around.
Four model options — and tokens aren't hidden.
The most frequent question on the first call: «Are you on ChatGPT?». The answer — «it depends». In production deployments a single model is rarely used: more often it's a combination of 2–3 for different tasks.
External paid models
Fast start, maximum quality
OpenAI, Anthropic Claude, Google Gemini. A working option when you need launch speed, there's no ban on external LLMs and data is processed under agreed rules.
- Billing
- Customer pays the provider directly
- Control
- Limits, spend monitoring, prompt optimization
- Jurisdiction
- US, EU
Different tasks within one agent may use different models: GPT-4 for complex reasoning, Claude Sonnet for documents, Gemini Flash for fast classification.
Russian corporate
Jurisdiction and a clear perimeter
GigaChat, YandexGPT and similar. Fit when there are jurisdiction requirements and you need a corporate contour with a clear contract.
- Billing
- Russian invoicing, pay-per-token
- Quality
- Comparable on typical tasks, weaker on complex ones
- Jurisdiction
- Russia
Critical for corporate customer accounting and for compliance with Russian data law 152-FZ.
DevNeuroX local models
No dependency on external APIs
Open-source models (Llama, Qwen, Mistral and similar) that we deploy and maintain on our infrastructure.
- Billing
- May be included in the 40K/mo maintenance within a limit
- Control
- Predictable cost — no token billing
- Jurisdiction
- Our infrastructure (RU)
Big practical advantage: counted in requests per month, not tokens per week. If volume grows — exceeds the limit — we revisit maintenance, without surprises.
Local models at the customer
Enterprise and fine-tuning
When format D (Enterprise on-prem) is chosen, the model is deployed on the customer's infrastructure. Requires GPU servers or an agreed private cloud.
- Billing
- Only our work on the agent layer
- Control
- Full independence, the model stays with you
- Jurisdiction
- Customer's perimeter
Customer fine-tuning is possible here — see below.
Enterprise · model fine-tuning for the customer
Fine-tuning on your data — faster response and higher quality in your domain.
In format D fine-tuning is possible — training a model on your historical data. We fine-tune both our local model and open-source models inside your perimeter.
Faster response
the model is adapted to your domain, no need to explain context every time
Higher quality
internal jargon, standard templates, corporate terminology — the model already knows
Fewer «hallucinations»
the model has seen your documents during training, it doesn't make things up
Full independence
the model and the data stay with you
This is a separate engineering stage within an enterprise project: data preparation, training infrastructure, usually 2–4 weeks of work.
What we DON'T do — and why it's important to know upfront.
The industry is full of promises that sound great on a landing page and end in disappointment two months after the pilot. So here are several important «don'ts» — better to hear them now.
We don't take processes without an owner on the customer side. If there's no person on the team who can say «this is right, this isn't» — the agent will turn into a noise generator. This isn't about customer laziness, it's an engineering necessity: machine learning requires human feedback to improve quality.
We don't do «just AI» without a business task. A request like «let's deploy AI and see what happens» isn't our format. We start only from a defined process with a measurable result. If it's unclear what counts as value — we don't launch the pilot.
We don't promise efficiency percentages before the pilot. No «we'll boost sales by 30%» in a commercial proposal. We agree metrics before the pilot and measure them in reality. In real cases the effect varies: in some places the agent closes 80% of typical requests, in others 30% — and both can be a success depending on the starting point.
We don't enable fully autonomous actions in production. The final decision on critical operations always stays with a human — sending an email to a customer, processing a payment, changing a deal stage. The agent prepares, the human confirms. This isn't a compromise, it's policy — there are too many market examples of how full autonomy without supervision ends.
We don't process personal data via public LLMs without an agreed contour. Russia's 152-FZ isn't a formality, it's a real risk. If the task requires PII processing — we discuss format C or D with a contour inside Russia and Russian models or a local model.
We don't present enterprise as a «more expensive tier». As we said, it's a separate engineering project, not a subscription. Asking for «the same thing the neighbor has, but enterprise» won't work. First — infrastructure and requirements scoping.
Where to start.
Depending on where you are right now — one of these three entry points fits better. They differ in depth, cost and format of responsibility.
01 · Pilot
Process is clear but it's unclear if AI will help
We start with a pilot for one specific agent — pick one scenario, connect the needed data source, show how the agent works on your real cases.
2–3 weeks · from 250K RUB
Discuss a pilot →02 · Implementation
The task is already clear — needs to be done properly from the start
We move to implementation scoping — review process, systems, data, roles. Prepare architecture, working contour and launch plan.
1–2 months · from 500K RUB
Get an implementation estimate →03 · Enterprise
Closed perimeter, multiple systems or local models
This is an enterprise contour. We start with technical scoping of infrastructure and security requirements. Based on the scoping we propose architecture and an implementation roadmap.
No «N thousand» estimate — this isn't calculated that way
Request technical scoping →If you're not sure which one to start with — message us on Telegram, we'll figure it out together in 5 minutes. Better to ask now than to guess.
We don't just deploy — we teach people to use AI properly.
In 2–3 years AI tools will be the standard in engineering and business teams. Whoever starts learning now will gain a competitive advantage that latecomers won't have. That's why we have a separate direction: training client teams, so that AI tools and agents don't become «a toy for 3 months», but turn into part of the working process.
Corporate programs for adopting AI in development
For company engineering teams that want to switch to the AI-native approach. A 2–4 week program: review of the team's stack, selecting AI tools for the tasks, training on using them, ramp-up to regular production use. Not a «theoretical course», but a real shift in how the team works.
CTO, tech leads, developers · teams 5–50+
Client team onboarding for working with AI agents
Once we've deployed Sales, Personal, Support, RAG or another agent — we train your team to work with it well. How to adjust scenarios, write prompts for typical tasks, handle complex cases, give the agent feedback for improvement. Without this even a well-deployed agent gets «abandoned» in 3 months.
Client operational teams · typically after agent deployment
«Where to start with AI» mentor sessions for executives
Short individual sessions with CEOs, CTOs, business owners: which AI tools actually work, where to start in your company, how to link AI to business goals, how not to burn budget on «let's try something trendy». Sessions — 90 minutes, with a concrete roadmap as the output.
CEO, CTO, owners · 90-minute sessions
Open learning materials in the Telegram channel
A free column in our Telegram channel @dxaiblog. AI tool reviews, practices from real projects, mistakes and how to avoid them, reviews of new models and platforms. Anyone can subscribe — engineer, owner, product manager. This is the public part of our education, no commitment.
Anyone exploring AI · public channel · free
Training cost is discussed individually — depends on team size, format (online/hybrid/corporate visit), program depth, and whether you need a process built from scratch or accompaniment of an existing one.
If you want to start small — subscribe to the Telegram channel @dxaiblog. All materials there are free. That's our «step zero» of training.
Describe the task — we'll come back within 2 hours
We reply within 2 hours during business hours. In the pilot call we'll show where the agent pays off and where it shouldn't be tried.