The AI Production Paradox: Three Signals That Are Rewriting the Market
- Джимшер Челидзе
- 17 hours ago
- 6 min read
Three news items from the first half of 2026, taken individually, look routine. Taken together, they redraw the map of the AI product market.
Anthropic launches Claude for Small Business. Yandex publishes data showing that 86% of AI agent users in Russia are SMBs. Swedish company Sinch publishes a study showing that three out of four companies have rolled back AI agents in customer support — after launch. Read together, these signals tell one coherent story: the market is maturing faster than expected, and the shape of the products that will win on it will look nothing like what conferences promised in 2024–2025.
Signal 1. Anthropic finally pivots to SMB
On May 13, 2026, Anthropic launched Claude for Small Business — a bundle of approximately 15 ready-made agentic workflows with connectors to QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365. (3DNews report). Under a single subscription, an owner gets month-end close, invoicing, margin analysis, contract review, lead sorting, and content strategy. In parallel, an offline tour kicked off across ten US cities with free workshops for 100 local entrepreneurs per city.
This isn't a product launch in the usual sense. It's a repositioning by one of the two global market leaders. Anthropic publicly stated what I'd been arguing in my book AI: From the Clouds Down to Earth: SMBs generate roughly 44% of US GDP, yet AI adoption there lags significantly behind large enterprises. Industrial software was historically built for corporations, VC-backed startups, and mass consumers — but not for a 30-person landscaping firm or a 50-person real estate agency. Anthropic is closing that gap.
What matters as much is how they entered. Not with an "open chat," but with a box for specific daily tasks. Not with "first train the model," but with ready-made scenarios that work from day one. And one more thing: Claude doesn't act unsupervised — the user initiates each workflow, approves the plan, and confirms the result before anything is sent, published, or paid. This is exactly the paradigm of "formalized interfaces" and "minimization of direct prompting" I've argued for.
Signal 2. In Russia, the picture is the same — and it's also about SMB
On March 3, 2026, Yandex B2B Tech published data from its Yandex AI Studio platform:
Over 7,500 AI agents analyzed;
86% of AI agent users in Russia are small and medium businesses;
About 25% of agents work in tech support, banking, and HR consultations.
The scenarios are vertical and narrow: developers and realtors use agents for property selection, management companies for handling housing complaints, scientific labs for equipment diagnostics, and some customers for news analysis. This isn't "GPT for the office" — these are concrete tasks for concrete business profiles.
In parallel, Yandex introduced tools that allow companies to use AI in cloud infrastructure without internet access — a direct response to the SMB segment's main objection on data security. The Russian market is roughly six months behind the global trajectory — and here too, the winners are narrow boxed solutions with security properly packaged.
Signal 3. The Production Paradox: 74% rolled back
On May 14, 2026, Swedish communications provider Sinch published the AI Production Paradox study based on a survey of 2,500 AI project leaders across countries and industries. (Coverage in The Register). The headline finding was stark:
74% of companies have rolled back or completely turned off AI agents in customer support that had already been deployed. "Rollback" here means removal from live production, not abandoning a project before launch.
The most counterintuitive part: at companies that themselves describe their AI control system as "fully mature," the rollback rate is higher — 81%. Sinch explains it this way: mature teams don't fail less often; they simply see problems sooner and react faster. Notably, the number doesn't vary by region, industry, or company size — money doesn't fix the problem.
Further numbers paint the same picture:
84% of AI engineering teams spend at least half their working time on safety infrastructure, not on actual development;
75% of companies cite trust, security, and compliance among the top three priorities of their AI programs — above AI development itself (63%).
This validates an earlier 2025 Gartner forecast: half of the companies that had planned to seriously reduce customer support headcount through AI will abandon those plans by 2027. Gartner's VP put it bluntly: a fully unmanned contact center is technically infeasible and operationally undesirable at this point.
A methodological caveat: Sinch is an interested party — they sell communication infrastructure built specifically for AI agents and naturally frame rollbacks as a deployment-platform problem. But even adjusting for that, the 74% figure fits the broader trend of the past year — the gap between AI demos at conferences and AI's actual performance in production.
What these three signals say together
Individually, three news items. Together, a map of the market.
First, the market has shifted its primary segment. The main buyer of AI products over the next 2–3 years is not the enterprise and not the "ChatGPT consumer" — it's the small and medium business. Anthropic and Yandex are entering from opposite ends of the world with the same playbook.
Second, the base product form has changed. What's winning is not the "open chat" but the box for a specific daily task — a workflow with a clear input, a clear output, and a clear zone of responsibility. Universal assistants are moving into the infrastructure layer; user-visible value lives in specific scenarios.
Third, the market's bottleneck is not the model — it's safety and trust. Sinch's rollbacks aren't happening because models got worse. They're happening because companies don't have the operational readiness to run autonomous systems in production. 84% of engineering teams spending half their time on safety is the new normal, not a temporary distortion.
Fourth, full autonomy lost the 2025 phase. The slogan "remove all employees and deploy an agent" has effectively been recognized by the industry as unrealistic in a 2–3-year horizon. Klarna and other high-profile 2024–2025 cases brought people back, boomerang style. Sinch's rollbacks are the same process at the technology level rather than at the headcount level.
Fifth, the human figure returns — not as an obstacle, but as a required architectural element. AI systems for 2026 and beyond are built as human-in-the-loop by design, not retrofitted because of regulation.
In short: the market is shifting from the paradigm of "AI replaces humans" to the paradigm of "AI amplifies a specific human under their control."
The profile of a winning AI product
These three signals reveal a clear profile of the AI product that wins with SMBs over the next 2–3 years. Six characteristics that all three signals validate:
Narrow but deep scope. An assistant for a specific role (owner, director, accountant, salesperson, realtor) or a specific task — not a "helper for everything."
Boxed packaging. Ready-made workflows and templates instead of "first train the model." SMBs pay for results from day one.
Human-in-the-loop by default. The user initiates the process, approves the plan, and confirms the result before publication, sending, or payment.
Security as the first message, not the fine print. "Your data doesn't go into training," "you see every step," "no write operations without confirmation."
Integration with the existing stack. Connectors to what the customer already runs — QuickBooks, HubSpot, Salesforce, Google Workspace, Microsoft 365 globally; 1C, amoCRM/Bitrix24, Kontur SKB for the Russian market.
B2C-first → B2B-expansion. Value to a specific person first, then a team version with roles and permissions later.
What this means for executives and entrepreneurs
Three practical implications for anyone working with AI right now.
If you're deploying AI in your company: Don't try to build a fully autonomous "turnkey" agent. Start with a narrow human-in-the-loop scenario — one department, one task, one user with an explicit rollback right. Sinch's 74% rollback rate is the lesson from teams that went straight to autonomy.
If you're building an AI product: Reassess your positioning. "Open chat for everything" is a 2023–2024 category. "A box for a concrete task with explicit human-in-the-loop and strong security communication" is the category that Anthropic and Yandex have entered. If you're closer to the first paradigm, you have 12–18 months to pivot.
If you're writing AI strategy for 2027–2028: Budget for safety, trust, and compliance to be comparable in size to development. 84% of AI teams are already working in that proportion. This is a structural shift, not a temporary spike.
Summary
Three news items from one spring describe a mature pivot in the market: AI is moving out of "magical agent that does everything autonomously" mode and into "boxed assistant that amplifies a specific human under their control" mode. The main segment is SMB. The main form is a narrow workflow with human-in-the-loop. The main competitive advantage is not model quality but safety and trust.
Whoever saw this turn earlier and started building products for this paradigm — rather than for the "autonomous agent replacing humans" paradigm — will ride the next wave on the right tracks.
Sources & References
Primary Sources:
Anthropic Claude for Small Business Launch
3DNews, May 13, 2026
How Business Uses AI Agents: Yandex B2B Tech Data
Vedomosti, March 3, 2026
https://www.vedomosti.ru/technologies/trendsrub/news/2026/03/03/1180334-ispolzuet-ii-agentov
AI Production Paradox: 74% of Companies Rolled Back AI Agents
Sinch Research Report, May 2026
"AI customer service bots get rolled back at 74% of firms"
The Register, May 13, 2026
