AI

We design and integrate AI solutions into the processes of mid-size and large companies: from bots to automation platforms. Impact is measured in money and hours.

See case studies
Integrations: ERP · CRM · messengers
We work where your business works

Trusted by

Your companyAutomation level

Case studies by industry and process

Most of our projects are under NDA. So this board shows only part of the details: the task, the solution and the numbers, without company names or people.

Processes
Industries
Project board21 / 21

  • The brand stopped "vanishing" from AI answers

    Task
    The holding discovered that ChatGPT and other assistants confused its products with competitors, and in some answers the brand was missing entirely. Marketing had no way to see the scale of the problem, let alone measure it.
    Solution
    We ran an AIV Research audit across 7 models and 300+ queries. We restructured the corporate site and materials for AI crawlers, reinforced authoritative sources and set up 24/7 monitoring with alerts on answer changes.
    +19 pp
    share of voice in LLM answers over 5 months
    7
    models under continuous monitoring
    0
    critical fact distortions after month three
    AIV ResearchAIV FlowRAG feedsstructured markup

  • Source documents processed 4 times faster

    Task
    A food holding was drowning in invoices, acts and specifications from hundreds of suppliers. Operators retyped data into the ERP by hand, and errors surfaced during reconciliations.
    Solution
    We deployed recognition and validation of source documents with a double check: the model extracts details and line items, rules verify them against contracts in the ERP, disputed documents go to a human. Closed perimeter, local model.
    ×4
    document processing speed
    90%
    of documents pass with no operator involved
    7 mo
    project payback
    local LLM in the perimeterERP integrationaccounting systemOCR pipeline

  • 68% of requests closed without an operator

    Task
    A federal retail chain could not keep up with the flow of requests about orders, returns and the loyalty program. First response time reached 40 minutes, customers were leaving.
    Solution
    We built a first-line bot in Telegram and MAX with access to order statuses and the knowledge base. The bot resolves typical requests and hands complex ones to an operator together with a dialogue summary and the customer's emotional state.
    68%
    of requests without an operator
    4 sec
    first response
    3.9 → 4.5
    CSAT growth
    Telegram Bot APIMAXretail CRMRAG over the knowledge base

  • Mentors reach their targets 40% faster

    Task
    An education platform was growing fast, and new mentors took months to absorb internal regulations. Experienced staff spent hours answering the same newcomer questions.
    Solution
    We assembled an onboarding assistant built on internal regulations and training recordings. It answers in the corporate messenger, links to primary sources and escalates what it does not know to the process owner.
    −40%
    time for a mentor to reach the plan
    78%
    of newcomer questions closed by the assistant
    120+ h
    of senior staff time freed per month
    RAG over the knowledge baseTelegramLMS integration

  • Application scoring didn't stick in hiring

    Task
    A recruiting company manually sifted through hundreds of applications for mass positions. The best candidates accepted other offers before the first call.
    Solution
    We set up application and CV screening against the client's criteria: the model structured experience, checked stop factors, ranked candidates and assembled a shortlist for the recruiter.
    Why it didn't take off
    The model optimized for formal criteria: tenure, keywords, profile match. Strong unconventional candidates sank to the bottom of the list, recruiters kept re-checking the output by hand and stopped trusting it after two months. We shut the pilot down honestly: scoring without explanations does not work in hiring. We returned later with a different design: AI hints on every application instead of ranking.
    2 mos.
    the pilot ran before being stopped
    38%
    of strong candidates were lost in the tail
    0 RUB
    savings: the goal was not met
    LLM pipeline with rulesATS integrationCRM exports

  • Banner production sped up by 70%

    Task
    A media group needed hundreds of banner variants for different placements and formats. Designers assembled resizes by hand and could not keep up with campaigns.
    Solution
    We deployed a banner generation and adaptation service: brand templates, generative backgrounds, automatic resizes for placement formats, checks for contrast and logo safe zones.
    −70%
    time per banner set
    ×3
    more variants per test
    0
    manual resizes: designers work on concepts
    generative image modelsbrand templatesad platform integrations
1–6 of 21

Products we build for ourselves and for the market

Search, choice and decision-making are moving into AI assistants. We build tools for both sides of this shift: AIV shows what AI says about your brand, Influence Agents manage how your brand speaks through AI.

AI Visibility platform

AIV Apex. Manage what AI models say about your brand

People increasingly ask ChatGPT, Gemini or Perplexity for advice instead of a search engine. In those answers your brand is either present or not. AIV Apex measures brand visibility in AI answers and helps you grow it.

  • AIV Research. A deep audit: how 7 models see your brand across 300+ real queries. Tone, factual accuracy, competitor comparison, the sources models rely on. You get a map of opportunities with priorities.

  • AIV GEO. Ongoing visibility work: technical content preparation for AI crawlers, source reinforcement, content shaped for model answers.

  • AIV Flow. A live dashboard: metric dynamics, alerts when answers change, reports without manual assembly.

Metric system
AVIIntegral index of brand visibility in model answersSAVShare of voice: how often models recommend you, not competitorsMCIConsistency: do different models tell the same story about youAWRShare of wins in direct "which is better" comparisonsRVPresence in recommendations with purchase intentISVVisibility by query type: from "what is it" to "where to order"SAIAuthority of the sources models take facts about you fromSISInfluence of specific sources on final answers

For whom: brands with a long consideration cycle, retail, B2B companies, pharma, real estate. Anyone whose customers "ask the AI".

aiv.apex2.ru
Engineering tool

AI Visibility Checker. Does your website let AI in the door?

Before managing model answers, check the basics: can AI crawlers see your website at all. Checker visits your site as the real bots of OpenAI, Anthropic, Perplexity, DeepSeek and Google, and shows who gets the content and who hits the wall.

What it checks
  • Server responses to every bot: statuses, redirects, headers, TLS
  • robots.txt rules versus the site's actual behaviour
  • WAF, captchas, geo-blocks and anti-bot filters that eat AI traffic
  • How much text is available without executing JavaScript
  • Control measurements with a regular browser profile

The result: a reproducible bot × domain matrix with raw data, plus a report on what to fix first.

A hands-on tool of the RW+ and Runway AI ecosystem. Born from our own GEO audit practice.

Check your website
AI persona platform

Influence agents in social networks and media. A swarm of ambassadors for your business

A platform of transparent AI personas, ambassadors and community agents: they listen to the market, answer your audience, explain products, correct myths and test messaging. By default the agents disclose their AI nature; in selected safe and ethical scenarios they work undisclosed, modelling the behaviour of real users.

  • Community Agents. Live AI agents that take an active part in discussions: they comment on your category's topics across social networks, media and niche communities, keep the conversation going and answer the audience. Complex cases go to a human.

  • Social Radar. Monitoring of topics, questions, myths and competitor comparisons: it spots trends, gathers crowd opinion, scores and segments the audience and profiles each segment.

  • Persona Lab. An archetype builder and synthetic focus groups: segment reactions to messages, pricing and creatives are tested before anything goes public.

What the agents do
listen to the marketAgents read permitted sources and map the pains, objections, myths and language of your audiencespot trendsAgents catch trends and discussion spikes before they become obvious: topics, formats, news hooksanswer the audienceDisclosed AI assistants reply on Telegram, on the website and in the helpdesk, passing complex cases to a humanexplain productsVirtual experts with a knowledge base explain the product from technical, business and user perspectivescorrect mythsThe platform finds persistent myths and prepares precise answers with links to sourcestest messagingSynthetic audiences model segment reactions before launch: objections, confusion, crisis scenariosanalyse and score opinionsA map of opinions and sentiment: what the audience thinks, where the myths and objections are, and where the buying signals show
Methodology

AIV Metrics Pyramid

Our own methodology for measuring AI visibility: a pyramid of 13 metrics on five levels, from basic presence in model answers to a brand's readiness for the AI-first world. The lower levels answer whether AI sees you at all; the upper ones show how confidently models recommend you and which sources they rely on. The pyramid powers AIV Research audits, the source influence rating and public industry reports: every conclusion is tied to a specific metric and is reproducible.

Demo report on the smartphone market: 4,103 queries · 41,023 answers · 6 models

Metrics pyramidSource influence ratingDemo report: smartphones 2025
AIV Metrics Pyramid methodology page on aiv.apex2.ru

aiv.apex2.ru/metrics

These products grew out of our GEO service. If you need an outcome rather than a tool:

How we work, step by step

Consulting and audit are built into the process: first we calculate the impact, only then we write code.

01Intro and case review60 minutesfree

A call or a meeting. We break down the process that hurts, estimate the impact and the cost. An honest verdict right away: where AI will help, and where tidying up your current tools is enough. Sometimes our answer is "you do not need AI yet". That is a result too.

What you get:A short meeting summary with a perspective estimate.

02NDA and immersion2-3 days

We sign an NDA before we see your data. We get access, meet the process owners and collect context: systems, volumes, security constraints.

What you get:A signed NDA and an agreed audit plan.

03Process and data audit1-3 weeks

The consulting phase. We map processes, measure the volume and cost of routine work, check data quality and availability, calculate the economics of scenarios. We prioritise: what pays off fast and what needs preparation.

What you get:An opportunity map with impact, time and cost estimates per scenario. The document stays with you whether we continue or not.

04Pilot with metrics2-6 weeks

We take one process from the map. Before the start we fix the success metrics: the numbers at which the pilot counts as a win. We build a working solution on real data with real users.

What you get:A working pilot and a "metrics before and after" report.

05Rollout and integrationfrom 4 weeks

We take the pilot to production: integrations with the ERP, CRM and channels, a closed perimeter where needed, load testing, operating procedures and team training.

What you get:A solution inside your perimeter, documentation, trained staff.

06Support and growthongoing

We watch answer quality and compute costs, update models, add scenarios from the opportunity map. A monthly report: what was done, what was saved, what comes next.

What you get:An SLA, monthly reports, a growing list of automated work.

The average path from the first meeting to a working pilot:weeks. It depends on data readiness, and at step 03 we tell you about it straight.

Safe AI within the Russian legal field

We build solutions your lawyers and security team can sign off without a fight: data inside your perimeter, processes aligned with personal-data law, architecture with headroom for upcoming AI regulation.

Server rack inside a protected perimeter

Fine figures are provisions of Russian law. We do not give legal advice: we work together with your lawyers and security team.

Data never leaves your perimeter

We deploy models and agents on-premise or in your cloud. Personal data, trade secrets and customer conversations do not go to external providers. For tasks without sensitive data we use a hybrid: a local model for the private part, a cloud one for the rest.

The cost of a mistake has grown. We design for it

of annual revenue for a repeated leakturnover fine ceiling

Since 30 May 2025, repeated personal-data leaks are subject to turnover fines. So we design by the principle of minimum: AI sees only the data a specific task needs, access is logged, sensitive fields are masked before they reach the model.

Local models, done properly

We work with GigaChat and YandexGPT through Russian APIs, and for a closed perimeter we deploy open-weight models on your hardware. We pick the model for the task, not for fashion: classifying requests does not need a model with hundreds of billions of parameters.

Ready for the upcoming AI law

The state is preparing AI regulation: the draft law includes labelling of generated content and certification of higher-risk systems. We design with headroom: model decisions are journaled, a human stays in the loop on critical operations, content is labelled where appropriate. When the requirements take effect, your system will not need a rebuild.

Special competence

Local AI without bills for excess hardware

The main risk of a closed perimeter is not security, it is the budget. An unoptimised local model eats the GPU budget and buries the project economics. We bring the cost per request down to a sensible figure and show it in reports.

  • Right-sizing. We match model size to the task. A compact model with good context often beats the biggest one.

  • Quantisation. We compress models with quality control on your test sets: less memory, a cheaper card, the same accuracy on the target task.

  • Efficient inference. Request batching, caching of repeated prompts and answers, speculative decoding. One server serves more users.

  • Hybrid schemes. Sensitive workloads stay in the perimeter, bulk ones go to the cloud. A router decides where each request goes, by rules your security team approves.

  • Cost monitoring. Cost per request, GPU load and quality degradation are visible on a dashboard. A spending anomaly is caught the day it appears, not at the end of the quarter.

In plain words: the same task runs on smaller hardware. The difference shows up in the equipment budget and the electricity bills.

GPU cards and stacks of coins: the economics of compute
Perimeter economics: an example20,000
Without optimisation270,000 RUB / month
With optimisation90,604 RUB / month
≈ ×3.0 savingsAn illustrative comparison of approaches, not a quote.

Platform independence. We integrate with anything

The solution is built around your process, not around a single platform. The bot lives in Telegram, MAX and on your website at the same time, data flows into the ERP and CRM. If one platform hits blocks or changes its rules, the business keeps running on the rest.

Messengers1C / ERPCRMTelephonyMarketplacesAnalyticsRunway AI
What we work withintegrations in production

Customer channels

Business systems & accounting

Telephony & contact centres

Commerce & traffic

Infrastructure & models

Your system is not on the list? The list is not complete, it is just what we have built hands-on. Ask us:

We will break down your case. Free and honest

60 minutes with engineers, not salespeople. You leave with an impact estimate and a straight answer on whether you need AI. If you do not, we will say so.

What brings you here?

We reply within one business day. No newsletters, no "just checking in" calls.

Not a fan of forms? Write to us directly: a.matroxin@runway-agency.ru