Main

BTU researchers assess that Georgian companies should expand AI-agent activity in line with their teams’ capacity to review and use the results. Otherwise, faster preparation can create a growing queue of work awaiting a manager’s decision.

Key Takeaways

• BTU’s analytical synthesis identifies the allocation of review time, expertise and responsibility as an important condition for expanding AI-agent use in Georgian companies.
• BTU’s calculation from Geostat’s 2025 business-sector data puts small and medium enterprises at 60.6% of annual average employment within that statistical coverage; organizing human-agent work is therefore relevant beyond large organizations.
• BTU researchers assess that recommendations prepared by agents for Georgian businesses should carry the evidence needed for a decision, so the responsible person does not have to restart the entire investigation.

Introducing an AI agent into a company often begins with a straightforward promise. Employees will spend less time gathering information, preparing correspondence or comparing documents. Yet obtaining a text quickly and completing a business task are different achievements. Before an offer reaches a customer, someone still needs to check the price, delivery terms, underlying information and whether the proposed answer matches what the company can actually deliver.

If one manager continues to do all of this, several fast agents can become several new queues on that person’s desk. The sales proposal is ready, the procurement comparison is complete and the customer reply has been drafted. None may be ready to send without the appropriate check. The limiting resource has moved from text production to the time needed to understand the result, verify it and make a decision.

BTU’s assessment begins with that distinction. Alongside assigning work to an agent, a company should determine how the resulting material becomes usable. Review belongs inside the work process and requires resources of its own. Ignoring it can make productivity look stronger on paper: the company counts more material prepared while overlooking the accumulation of unfinished, unapproved or correction-intensive work.

A new signal from agent-intensive work

Anthropic’s study published on 16 October 2026 reports that meta-work, meaning work to organize and supervise agent workflows, rose from 10% of interactive sessions in autumn 2025 to 22% in spring 2026. The estimate uses automated classification of 2,000 short summaries. This is an observational company study, not a working-hours measure; organizational growth may also affect the findings.

A Georgian company does not need to apply that percentage to its own employees. The useful question is whether accelerating execution also creates work that nobody planned to accommodate. Clarifying an agent’s assignment, comparing inconsistent answers and investigating why something went wrong can all become additional responsibilities.

This does not imply that AI necessarily makes work worse. A capable agent can remove substantial routine effort. The opportunity is more fully realized when the organization also changes how it receives and examines the output. If an employee repeatedly has to find the original source and establish what changed between versions, some of the time saved in preparation returns as checking effort.

Why the issue matters in Georgia

Geostat’s detailed 2025 tables record annual average business-sector employment of 899,026. Small enterprises account for 367,456 and medium enterprises for 177,479. Together, the two groups account for 544,935, or 60.6% of the total. These figures refer to the coverage of the business-sector survey, rather than the entire Georgian workforce.

Enterprise size does not reveal how many managers a company has or whether any particular manager is overloaded. It does establish the economic scale of the question. AI adoption cannot be understood solely through a large bank, technology company or international chain. Its wider significance will also depend on whether smaller organizations can use additional capabilities with their existing human resources.

Where one manager simultaneously handles customer relationships, supplier decisions and employee support, a new stream of review work needs careful allocation. Such a business may not need a separate AI-control department. It does need a clear answer to a more immediate question: who will check each result, and what information must that person receive to do so?

BTU’s research framework indicates that technical difficulty alone is an insufficient basis for allocating attention. The consequence of an error, the ease of detecting it and the possibility of correcting it also matter. A missing detail in an internal meeting note and an incorrect commitment sent to a customer deserve different treatment. The organization’s use of human time should reflect that difference.

Reading is not the same as checking

Reading an AI response does not establish that the response has been verified. Fluent Georgian, a tidy table and a persuasive conclusion can create an impression that the underlying work has already been confirmed. In practice, one detail may determine the decision: whether a price is current, whether a tax has been included, or whether a condition actually applies to this customer.

BTU therefore assesses that an agent should deliver information that supports review, rather than just an answer. The basis for the proposed decision should remain visible: the source of a fact, its reference period, what is unresolved and the action being proposed. This need not create additional bureaucracy when it reduces the need for the reviewer to begin searching again.

Consider an illustrative Georgian distribution business. An agent compares two suppliers and recommends one as more attractive. The conclusion alone does not equip the manager to decide. A lower purchase price may come with later delivery, a larger minimum order or less favorable payment terms. A useful comparison brings those conditions together and shows which ones have been confirmed.

The human task in this example is not to rewrite every document. It is to evaluate the actual choice: whether the offer serves the company’s needs and whether the remaining risk is acceptable. The example explains the organization of work; it is not a measured company result or an estimate of savings.

Different outputs require different checks

Source: BTU analytical synthesis of its research materials. Qualitative, illustrative comparison; outcomes have not been measured.

The distinctions in the table show why a universal rule that one person must read every answer is inadequate. Different actions require different checks. Attention should concentrate where an error has a significant consequence, while suitable parts of repetitive technical verification can follow a previously defined rule.

A short answer must retain the evidence

A shorter report often looks like an obvious solution. It can reduce reading time, but brevity helps only when the essential questions remain answerable. If compression removes the source date, conflicting evidence or a material condition, the manager is simply reading less reliable material faster.

BTU researchers assess that a good summary should conserve attention without sacrificing the information needed to verify it. The headline conclusion can be short while its basis remains accessible. A response to a customer complaint, for example, need not reproduce the entire conversation. It should preserve the correct sequence of events, the applicable rule and the specific circumstance that makes the case different from a routine request.

The Georgian language adds a practical requirement. An internal policy written in another language, a request received in Georgian and a proposed customer response may be combined within one workflow. The reviewer needs to see whether translation has changed the meaning of a condition. Dates, amounts, exceptions and promises are particularly important because an ambiguous phrase can create a different expectation.

A second agent can sometimes help by comparing figures or flagging a missing source or contradiction. But two persuasive answers do not constitute independent confirmation when both rely on the same incorrect information. The value of verification comes from testing a specific question against appropriate evidence, not simply obtaining another text.

Protecting time for a decision

A company can introduce AI while expecting employees to retain all their previous responsibilities and also review agent output. The new responsibility then becomes invisible. Performance assessments may count proposals prepared without recognizing the time spent preventing an error.

BTU’s framework suggests treating review as professional work requiring time and relevant expertise. A person should have more than an approval button. The reviewer needs the ability to ask a question, return work for correction, request further evidence and, when necessary, stop the process.

This does not mean moving every decision indefinitely upward. If each minor change reaches the director, the delay simply accumulates higher in the hierarchy. Where an existing rule resolves a case, the employee with the relevant competence should be able to decide within their authority. The leader’s attention should remain available for more difficult choices in which objectives or interests conflict.

Training should follow the same logic. Alongside learning to write a useful prompt, employees need to distinguish a fact from an assumption, evaluate the reliability of a source and recognize information missing from a decision. For the company, these skills can be as important as knowing the functions of a new application.

Conclusion

The speed of an AI agent is an opportunity for business, but human attention is not unlimited. BTU researchers assess that Georgian firms should connect the expansion of agent activity to their teams’ ability to understand, appropriately verify and responsibly use the results.

In well-organized work, the agent reduces repetitive search and comparison while the person concentrates on consequential choices. Without that connection, the company receives more material while decisions continue to wait for a person who no longer has time to check it.

Questions and answers

Does 60.6% mean that AI will replace this share of work

No. It is the small-and-medium-enterprise share of annual average employment in Geostat’s business-sector coverage. The statistic measures neither AI adoption nor job replacement.

Can another agent take over all verification

A second agent can assist, but its conclusion must also be grounded in evidence. Responsible human judgment is particularly important for decisions with material consequences.

How can a company recognize superficial review

A warning sign is an approval whose author cannot explain what was checked and what evidence supported it. Review requires the necessary information and genuine time for examination.

When is adding more agents useful

When an additional agent reduces a specific constraint and the team can use the result. Increasing the quantity of prepared text alone is insufficient justification.