AI Integration in Business: Why 75% of Projects Fail to Deliver Results | Tantal
71% of Russian companies use generative AI — but only one in four gets the desired result. The rest paid for someone else's experience: bought licenses, assembled a pilot, showed it at a meeting — and got stuck.
Why the market is growing, but results are not
The Russian market for generative AI grew from 13 to 58 billion rubles in 2025 — five times over. Companies connect assistants, set up chatbots, test document recognition. The average number of business functions with AI per company rose from 2.4 to 3.1. By the numbers — it's a boom.
Only 16% of AI projects are scaled up to the level of the entire company (McKinsey). The rest live in pilot mode: one department, a nice presentation, and that's it. Money spent, process remains manual as before.
The reason is one: pilots are launched to "try out AI" — but technology doesn't pay for itself. It pays off when a specific process is accelerated or made cheaper. When the goal is vague, there's nothing to measure — and the project quietly dies after the enthusiast who promoted it leaves.
Three entry points where AI gives measurable results
AI enters the company through three doors — into areas with the most routine and money.
Sales
AI Assistant handles primary lead processing: responds instantly, qualifies leads, schedules meetings, works outside office hours. The manager gets a prepared contact and doesn't waste time on those who just asked the price and left. Companies that implemented AI-powered agents first see a 22.6% increase in productivity (NVIDIA) — and sales feel it first.
Customer Service
AI handles routine inquiries: order status, delivery conditions, returns, frequent questions. Up to 95% of routine requests are removed from operators — they're left with complex cases. The client gets an answer in seconds, not hours in a queue.
Operations
Document processing, invoice analysis, data synchronization between systems. The effect is quieter but more stable: a 15.2% reduction in costs (NVIDIA). Operations often give predictable ROI — the volume of work is measurable, and errors cost specific money.
Bitrix24 and 1C as a platform: what works without relocation
The main myth about AI implementation — that you need to buy a new system and train your team on unfamiliar software. Most medium-sized companies already work in Bitrix24 or 1C — AI sits on top of these systems as an add-on, not a replacement.
What's realistically achievable without changing the stack:
- Automated funnel. AI leads deals through stages, reminds about stalled contacts, suggests the next step based on history.
- Incoming processing. Website, messenger, and email requests are processed in CRM already analyzed — with topic, priority, and responsible person.
- Lead scoring. The system evaluates deal probability based on client behavior and moves hot contacts up.
26% of managers call "unclear return" the main reason they cut AI budgets (vc.ru). When AI works within that CRM where the team sits every day, the return is visible immediately: same cards, same reports — lost requests are gone, managers get more done. Adding performance marketing closes the "ad — request — CRM — deal" loop in one manageable circuit.
ROI without illusions
PwC analyzed 200 AI projects: median ROI was 159% with a payback period of 6.7 months. Half of the projects returned more than 1.5 times invested within seven months.
"Median" — key word. It's not a guarantee, but the middle value in distribution. The same document recognition process gives an 80% reduction in operational expenses in one company, while it doesn't pay off in another — volume of documents is too small. According to "Yakov and Partners," the economic effect of AI for Russia by 2030 will be 7.9-12.8 trillion rubles. These trillions are made up of projects where economics were calculated before launch.
Three factors that depend on your ROI:
- Process volume. More repetitive operations — faster automation payback.
- Error cost. Where manual error is expensive, AI pays off even at low volumes.
- Data quality. If data is scattered, part of the budget will go to preparation before launch.
Three mistakes that kill a project
Only 25% of AI initiatives give expected ROI (McKinsey). Three reasons why most end up in the other 75%.
Automating chaos instead of process. If the process isn't described and each manager does it their way, AI will speed up disorder. First, describe the process — then automate, not vice versa.
Picking a tool before setting tasks. "Let's implement AI" — that's not a task. The company buys a platform, then looks for what to apply it to. The tool is there, but no effect.
No owner of implementation inside. Without a specific person with authority, the project dissolves between departments. The contractor does their part, but without an internal champion, the process won't take root.
When you need a contractor
Three scenarios where an external team is cheaper than doing it yourself.
1. No AI expertise. Hiring an ML engineer for one project is expensive and takes time. Tantal gathers a team of 253 specialists — without hiring in-house. 2. The task intersects multiple systems. An error at the CRM, phone system, and warehouse interface costs more than working with a contractor. Here you need integration experience, not experiments. 3. You need results faster than six months. In-house pilots drag on — people have their main work. An external team works focused on the task.
What to prepare before meeting with a contractor:
- Process description — what happens now, step by step.
- Operation volume — how many requests, calls, documents per month.
- Current stack — Bitrix24, 1C, amoCRM, phone system.
- Desired KPI — response time, processing cost, conversion rate.
A full list of directions — from assistants to mobile development — is in the Tantal services section.
How a normal pilot looks like
A normal pilot is a limited-time experiment on one business function, not a multi-month project with an inflated budget.
One process — for example, processing incoming sales requests, not "all AI in the company." 4-8 weeks — enough to test a hypothesis and get measurable results, but too little for budgets to grow. A small team: setup person, integrator, data specialist. If the pilot is successful, the payback period fits within those 6.7 months from PwC data.
The point of a pilot is to test one hypothesis at a known cost. If it's confirmed — scale up to adjacent processes. If not — you lost eight weeks, not a year's budget. That's how projects end up in the 25% that give expected results.
If you understand which process you want to automate but are unsure where to start — discuss a pilot with the Tantal team. We'll break down the task and honestly say if the result is realistic within four-eight weeks and at what cost. With 253 specialists in-house, we can gather a team for your specific process without long-term hiring. Leave an application to assess the project — no further commitment required.
