What AI Agents Cannot Do (Vendors Won't Tell You)

Every sales deck this year promises an AI agent that “handles everything” — sales, support, bookkeeping, even hiring. Business owners in Tashkent and Samarqand who buy into that pitch often end up disappointed three months in, not because the technology is bad, but because nobody told them what it structurally cannot do.

Understanding what AI agents cannot do before you sign a contract saves you money and saves your customers a bad experience. This isn’t a takedown of AI agents — Fera Tech builds them — it’s the honest limits list a vendor trying to close a deal usually skips.

---

AI agents cannot invent information they don’t have access to

An agent connected only to a product catalogue cannot tell a customer their delivery status if it isn’t wired into your logistics system. It cannot check a Payme transaction if nobody gave it access to your payment provider’s API. When an agent lacks data, the honest failure mode is “I don’t have that information” — the dangerous failure mode, which under-built agents fall into, is guessing an answer that sounds confident and is wrong.

This is why the integration work — connecting the agent to 1C, MoySklad, amoCRM, your Telegram bot, your payment gateway — is most of the real engineering cost. A vendor quoting a suspiciously low price is usually skipping this part, which means the agent will look great in a demo and fail on real customer questions.

AI agents cannot make final judgment calls on money, legal, or medical matters

Agents can draft a refund recommendation, flag a contract clause, or summarize symptoms a patient describes — but handing over final authority on a refund above a threshold, a legal commitment, or a diagnosis is a liability problem, not a technology problem. In regulated sectors like banking and healthcare — two of the sectors named in Uzbekistan’s AI Strategy 2030 push — a human needs to stay in the approval loop for anything with real financial, legal, or health consequences.

A well-built agent handles the 80% of routine cases end to end and routes the remaining 20% to a person, with full context attached so the handoff doesn’t waste the customer’s time repeating themselves.

AI agents cannot fix a messy process — they expose it

If your CRM has duplicate customer records, your product catalogue is outdated, or your team has three different ways of handling the same complaint, an agent built on top of that mess will inherit the mess. Automation reveals inconsistency faster than a human ever would, because it applies rules literally where a person would quietly work around a gap.

Before automating, it’s worth a short internal audit: which process are you automating, and is it actually the process you want repeated at scale? Our guide on choosing between an AI agent and a chatbot walks through matching the tool to the actual job rather than the hype.

AI agents cannot read tone perfectly, especially across languages

Uzbek businesses routinely serve customers who write in Uzbek, Russian, and sometimes a mix of both in the same message. Modern agents handle this far better than the chatbots of five years ago, but sarcasm, frustration expressed indirectly, or a regional turn of phrase can still be misread. A frustrated customer typing a short, clipped message might get a cheerfully generic reply that makes things worse, not better.

The fix isn’t abandoning automation — it’s designing clear escalation triggers (repeated negative sentiment, all-caps, specific keywords like “shikoyat” or “жалоба”) that hand the conversation to a human quickly, before the agent’s misread compounds the problem.

AI agents cannot run without ongoing maintenance

A launched agent isn’t a finished agent. Product prices change, promotions start and end, your CRM fields get renamed, Payme or Click update their APIs. An agent left untouched for months quietly drifts out of date and starts giving stale answers with the same confident tone as accurate ones — often the worst outcome, because nothing visibly breaks.

This is why realistic support budgets start around $50/month for monitoring and small updates, not zero. Vendors who don’t mention ongoing cost at the sales stage are usually the ones who disappear after launch.

Quick comparison: claim vs. reality

Vendor claim	What’s actually true
”Handles everything, no humans needed”	Handles routine 70–80%; complex/financial/legal cases need a human in the loop
”Works out of the box”	Needs integration with your CRM, payments, and Telegram to be useful
”One-time cost, done forever”	Needs ongoing monitoring and updates as your business changes
”Understands any tone perfectly”	Handles standard requests well; ambiguous or emotional messages need escalation rules
”Fixes your operations”	Only automates the process you already have — messy in, messy out

Before you buy, check the vendor can answer these

• What systems will this agent actually connect to (CRM, payments, Telegram, inventory)?
• What happens when it doesn’t know the answer — does it guess or escalate?
• Who approves refunds, contracts, or anything above a set so’m threshold?
• What’s the monthly cost after launch, not just the build cost?
• Can I see it handle a genuinely awkward customer message, not just the happy path?

If you want a realistic sense of what the build itself should cost before maintenance, see our breakdown of AI agent pricing in so’m.

So what should you actually expect?

A well-scoped AI agent, built by a team that’s honest about these limits, still delivers real value: many Tashkent food-delivery and retail businesses report Telegram-bot-driven automation covering 40%+ of monthly turnover, and AI-assisted sales teams often work roughly 40% more efficiently on routine tasks. That’s a strong return — but it comes from an agent scoped around what it does well, with humans handling the rest, not from a vendor’s “handles everything” slide.

For the fundamentals of how these systems are built and where they fit your business, our complete guide to AI agents for business in Uzbekistan is a good next read, alongside our plain-language explainer on what an AI agent actually is if you’re earlier in the research.

Frequently Asked Questions

Does this mean AI agents aren’t worth it for a small business? No — it means scoping matters more than the sales pitch. A focused agent handling Telegram orders and Payme payments for routine cases can pay for itself within weeks; an agent oversold as a full replacement for staff usually disappoints.

Can an AI agent ever handle 100% of customer interactions with zero human involvement? In practice, no reputable build promises this for anything beyond the simplest FAQ bot. Even mature agents route edge cases, financial approvals, and complaints to a human.

How do I know if a vendor is being honest about limits? Ask them directly what the agent cannot do, and watch for a specific, concrete answer versus a vague reassurance. A vendor who can’t name a single limitation is a red flag.

Is it cheaper to just build a chatbot instead? Sometimes, yes — a simple Telegram bot in the 500,000–1,500,000 so’m range covers basic FAQ and ordering flows well. The chatbot vs agent comparison helps you decide which fits your actual volume and complexity.

If you’re weighing an AI agent for your business and want a straight answer about what it can and can’t do for your specific case, take a look at our work and services, then get in touch — we’ll tell you plainly what’s realistic before you spend a single so’m.