AI Implementation in Business | Tantal

Companies that adopted AI early are gaining up to 8% EBITDA right now — not venture-backed startups, but ordinary B2B businesses working with the same budgets you have. Their competitors are still preparing a "why AI?" presentation for next quarter.

The gap is growing non-linearly. A year from now, catching up will mean closing a deficit not just in tools, but in data, processes, and team capabilities. Below — what the numbers say, where AI pays off fastest, and how to launch your first pilot in two weeks.

The market has already shifted — here are the numbers

Over the past year, the share of Russian companies applying AI to business tasks grew from 28% to 43% (Systema X, ComNews, Q2 2025). That is 15 percentage points in 12 months — a pace the enterprise software market has not seen since the shift to cloud.

The volume of AI and predictive analytics deployments in Q2 2025 grew 32% year over year. These are not R&D pilots — these are production deployments that appear in financial reporting.

Among large companies, 71% already use generative AI in at least one business function (Yakov & Partners jointly with Yandex, 2025). For enterprise, the question of whether to adopt AI is settled. The open question is how to scale what is already working.

For mid-market companies, this means one thing: the window in which AI delivers competitive advantage rather than mere parity is narrowing. Within 12–18 months, an AI assistant in the sales funnel will be as standard a requirement as a CRM is today.

What AI delivers in financial terms

A marketer going to defend a budget in front of the CFO needs three figures: investment, return, and payback period.

AI as a share of IT budget in leading industries — 13–17% of annual spend. A substantive line item, on par with infrastructure and licenses.

Impact — up to 8% EBITDA today, with projections rising to 13–21% within a year (Yakov & Partners + Yandex). For a company with EBITDA of 200 million rubles, that is an additional 16 million rubles in operating profit in year one — without expanding headcount.

Global ROI — 41%: every dollar invested returns $1.41 (Snowflake + Enterprise Strategy Group, "Radical ROI of Generative AI," April 2025, survey of 1,900 executives across 9 countries). 92% of early adopters confirm their investments have already paid off; 98% plan to increase AI budgets in 2025.

Time to positive return — 6–14 months depending on the scenario. Comparable to a CRM or ERP deployment, with a higher upside.

Where AI takes hold fastest

Five industries account for more than 60% of the total economic impact of AI in Russia: e-commerce, telecom and media, IT, construction and real estate, and healthcare. Yakov & Partners projects 7.9–12.8 trillion rubles per year by 2030, equivalent to up to 5.5% of GDP. Marketing runs through every one of these sectors: performance channels, content, database management, analytics.

A telling case: Rostelecom Contact Center. In its first year working with LLM and speech analytics, the share of projects with deployed AI reached 46%, 38 billion minutes of speech were processed, and the speed at which customers received answers increased sevenfold (CIPR, March 2025). A sevenfold acceleration in a customer-facing function is not a 5% optimization — it is a change in the operating model.

One scenario the market is still underestimating is automated video content generation. An AI video farm produces dozens of clips per week without camera operators or shoot days, fundamentally changing the economics of performance video on VK, Telegram Ads, and Yandex Direct.

Three entry points: how companies start in practice

71% of large Russian companies use generative AI "in at least one function" — that phrasing matters. No one deploys AI across the board all at once. Everyone starts with the single process that is causing the most pain.

Pattern 1. AI assistant for inbound inquiries. The shortest path from launch to a measurable result. The Tantal AI assistant handles up to 95% of inbound inquiries without operator involvement: applications, standard consultations, order statuses, service bookings. What you need to start: an FAQ database, CRM access, 2–4 weeks for integration. What you get: a lighter load on the contact center, higher conversion at first touch, and round-the-clock lead processing.

Pattern 2. Computer Vision for operations. The right fit for companies with physical processes — manufacturing, logistics, retail, security. Computer Vision reduces operating costs by 80% in tasks such as defect detection, shelf compliance checks, document OCR, and video analytics on live feeds. What you need to start: cameras and access to the video stream. What you get: automated quality control with no human error.

Pattern 3. Predictive analytics on data you already have. Less obvious, but often the fastest to show a return. Most companies have accumulated data in their CRM, billing system, and web analytics — and it sits unused. An AI model built on that data predicts churn, LTV, purchase propensity, and the optimal moment to reach out. The budget is lower than building from scratch; the result is targeted improvement in conversion and retention.

The guiding principle: choose not the trendiest scenario, but the one where you already have data and a clearly defined problem with a number attached to it in rubles.

Why AI implementations fail

98% of companies plan to increase AI budgets in 2025. Money without process does not become results — it becomes a report on money spent. Three failure patterns we encounter most often in projects.

Mistake 1. Buying a tool instead of solving a problem. A company subscribes to an AI platform because "competitors have one," then spends six months looking for somewhere to use it. The correct order is the reverse: start with the metric you need to move, then identify the process that drives it, and only then choose the tool.

Mistake 2. Poor-quality data. An AI model is only as good as the data going into it. If the CRM is maintained inconsistently, if customer calls are not recorded, if 1C exports run manually once a week — no assistant will rescue that process. Data preparation typically takes longer than building the model itself.

Mistake 3. Deploying without the operational team. When a solution is handed down from above without involving the people who work in the process every day, it either gets quietly ignored or worked around. Successful deployments always come in pairs: the technology, plus process adaptation and team training.

Over 10+ years and 281 completed projects, we have run into these mistakes repeatedly — in clients' cases and our own. That is the core value of an experienced partner: not writing the code, but guiding you past the pitfalls that have already been encountered.

How to choose an AI partner — five questions

The market is growing, more vendors are entering it, and quality varies by an order of magnitude. Five questions worth asking before signing a contract.

1. Show me three projects in my industry or one with comparable process characteristics. Not "we've done AI for various companies," but specific cases with before-and-after figures. If a vendor cannot produce relevant case studies within ten minutes, that is the first warning sign.

2. Who will be on the project team? An AI project requires a combination of an ML engineer, a backend developer, a data analyst, and a project manager with domain experience. A single generalist assigned to the project does not scale.

3. How do you handle client data? Where is it stored, who has access, what does the security perimeter look like. This is especially critical when data includes personal information or trade secrets. The answer should be ready at the first meeting.

4. What does post-launch support include? AI models degrade over time, data changes, and business logic gets updated. Without a clear SLA covering support and model retraining, the solution will stop performing within six months.

5. Do you cover the full stack — from process audit through CRM integration to production launch? An AI deployment touches the CRM, telephony, billing, and analytics. A vendor that only delivers the model leaves you to handle integration on your own.

Additional filters: Skolkovo resident status, a real in-house development team, and published case studies with named clients.

How to launch a pilot in two weeks

The main obstacle to launching AI is not budget or technology — it is inertia. A company spends months debating "strategy" while a competitor is already reporting initial results.

Week 1. Choose a scenario. Bring the team together for two hours, list 5–7 processes where there is a real problem, and score them on two axes: volume (how much time it consumes per month) and measurability (is there a figure in rubles attached to it). Pick one — the most painful and the most measurable.

Week 2. Data audit and MVP. Check whether the data for your chosen scenario exists at the required quality level. Launch an MVP in a limited scope — one channel, one team, one customer segment. The goal is not a perfect solution, but a measurement of the effect at small scale.

After two weeks you will have three things in hand: a hypothesis validated with real numbers, an understanding of the bottlenecks, and a scaling plan. That is worth ten times more than an 80-slide presentation on an "AI transformation strategy."

—

At Tantal, we run a free process audit and help you identify a pilot scenario in a single meeting. 253 specialists on staff, Skolkovo resident status, 281 completed projects across Russia and Belarus. Write to info@tantal.ai or call +7 962 996 00 66 — Anton, Director of Marketing, will respond personally.