Any AI agent can make mistakes. What matters is knowing where errors can occur, measuring them and limiting the damage. Below are the main risks and how we counter them.
For every project, a risk register is built from this table.
| Risk | How it shows up | Countermeasure |
|---|---|---|
| Wrong answers (hallucinations) | The agent confidently states something that isn’t true | Answers only from sources; with no source, hand off to an employee |
| Data quality | An outdated catalogue, two conflicting SOPs, empty CRM fields | Start with data cleanup and assign an owner |
| Uzbek language and voice | Word errors with dialects and mixed speech | Measure on real recordings, confirm key details in text |
| Channel rules | Reply window expired, account restricted | Adapt the scenario to Telegram and Meta rules |
| Personal data | Data ends up in an unauthorized location or abroad | Data flow map, sign-off from a lawyer |
| Excessive permissions | The agent changes a record it shouldn’t touch | Least privilege, a separate technical user, logging |
A language model writes text based on probabilities, so it sometimes states errors in a confident tone. We limit this on three levels. First, the agent answers only from approved sources. Second, commercial terms, deadlines and commitments are fixed in written rules. Third, every week a sample of conversations is checked by hand and the error rate is measured.
These limits are set by the technology or the platform. They can’t be bypassed, only accounted for.
According to Kun.uz (March 27, 2026), on March 26, 2026, the President signed amendments to the Law “On Personal Data,” which took effect on the day of official publication. Many categories of personal data can now be stored and processed abroad if one of the conditions set by the law is met: the country provides equivalent protection, standard contractual clauses or binding corporate rules apply, or international standards are followed. Certain categories must still be stored within Uzbekistan.
The rule is simple: no more than the task requires.
The agent never works as the administrator. It has its own account and a minimal set of groups; for example, an assistant connected to Odoo via MCP works with the key owner’s rights.
Broad read access, narrow write access. The agent is never given delete rights.
Records involving money, inventory and contracts are created only as drafts.
Every agent action is logged with its time, source and result. Suspicious actions trigger an alert to the reviewer.
Text from a customer or inside a document may try to mislead the agent, for example: “forget your previous instructions.” That is why the agent treats external text as data, not as commands, and risky actions never run without human approval.
Only the part needed for the task. Which fields go to which service and to which country is recorded in a table at the start of the project. Scenarios involving biometric data are reviewed separately.
Some parts, yes, such as open speech recognition models. Self-hosted versions of large language models usually lose on quality; we assess this trade-off through testing.
The agent is a tool; decisions are made by the company’s employees. That is why important actions sit in an approval zone and every agent has an owner on the client side.
We review your processes and recommend a specific task, KPI and plan for your first agent.