An AI assistant operating within Yandex Direct’s interface has demonstrated alarming reasoning errors, highlighting the limits of screen-reading automation and the importance of domain knowledge in digital advertising management.
A browser-based AI assistant has been shown to make confident but costly mistakes when reading the interface of Yandex Direct, the Russian search advertising platform, in work involving two accounts for corporate server and storage sales. The analysis, carried out over three days in early August, concerned a niche where sales cycles last months and individual deals can run into millions of roubles. The author said the assistant did not merely misjudge the logic of the account data; it repeatedly read the interface incorrectly, then built apparently coherent conclusions on top of those misreadings.
The key point is that the tool was not given passwords, tokens or API access. It operated inside an already authorised browser session on the user’s machine, clicking through the same Yandex Direct screens a human would use. That meant it could see only what the interface rendered, not the underlying data itself. In practice, the assistant was constrained by the same interface limits as a person, including account switching rules and closed login pages. According to the account described in the Habr article, even a small detail such as where to click to switch between organisations took time to uncover.
Five failures followed, and all were presented with the same calm, authoritative tone. In one case, the system treated the broad “Conversions” column as if it meant actual leads. A campaign showed ten conversions over 30 days, but nine of those were triggered by taps on a mobile telephone number, not by genuine enquiries. That distinction only became clear when the data were broken down by goal. From that mistaken premise came a second error: because the assistant believed mobile inventory was generating plentiful, cheap conversions, it recommended shifting budget towards Yandex’s network placements. The logic was internally consistent, but the input was wrong, so the recommendation was wrong too.
A third mistake came from failing to expand a collapsed list of keywords. The assistant saw only the first five rows in several product groups and assumed the groups were nearly empty, with about five phrases each. In reality, Yandex Direct hides the rest behind an “More N phrases” control, and the groups contained more than ninety phrases apiece. The proposed expansion plan therefore addressed a problem that did not exist. A fourth error was more severe: when reading a report, a column shift caused the assistant to assign roughly 14,000 roubles a week to a single broad-match key term. A second check showed the real spend was just over 200 roubles over 30 days. The difference was not a matter of interpretation but of misaligned columns, and the scale of the mistake was more than sixtyfold. The fifth case was a false alarm about deleted working keywords; a manual check showed nothing had been removed.
What makes this more troubling than an ordinary reasoning error is that the argumentation itself was sound once the false premise had been accepted. There was no visible breakdown in logic. The assistant did not flag uncertainty, and it did not distinguish between data it had directly verified and data it had inferred from a screen reading. The Habr author said the tool could correct itself when challenged, and would recalculate and acknowledge the mistake if asked. But left on its own, it never volunteered that the data were incomplete, and it never hesitated over the figures it had just extracted from the interface.
The article argues that this kind of failure is more dangerous than a transparent analytical mistake because it is harder for a human to spot. A person who does not live inside the account can easily accept both a correct finding and a wrong one when they are presented in the same style and with the same level of confidence. The only reliable filter, the author says, is domain knowledge: an awareness of how this market behaves, what constitutes a real lead, and which numbers are plausible in a tender-driven business where a tap on a phone number is not the same thing as a qualified enquiry.
After the corrections, the workflow was changed. The assistant now produces a daily review of both accounts and sends the summary to a work chat at midday, replacing about an hour of manual checking. But the rules are tighter. The machine is not allowed to change the campaigns directly; it prepares a list of amendments, while a person applies them. That caution was reinforced by a separate failure in which a manual run stopped with an authorisation error during a negative-keyword task. The error was not in Yandex Direct itself, but it showed how fragile a fully automated edit could become if it failed mid-process.
The broader conclusion is a sober one. A browser assistant that reads screens rather than raw data can be useful, but only up to a point. A more powerful model is unlikely to fix a collapsed list that was never opened or a table whose columns have shifted under the cursor. The article suggests the real safeguard is not bigger automation, but better access to structured data and a human operator who understands the business well enough to reject a polished mistake. In this case, that rejection mattered most when the plan called for a budget allocation of 14,000 roubles a week that had never existed in the account at all.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
