Where AI agents work faster than people

Main

BTU researchers assess that an AI agent’s most defensible speed advantage arises where finding information and moving it between systems takes longer than checking the result. For Georgian companies, the first automation candidate should therefore be a bounded part of a workflow rather than an entire occupation.

Key Takeaways

• BTU’s assessment for Georgian businesses measures agent speed through time to a verified outcome, including human involvement and repeated work.
• International customer-support research finds more issues resolved per hour with AI assistance, but that finding concerns support for people rather than autonomous agent superiority.
• Information retrieval, application preparation and exception identification are candidates for testing in Georgian services; a local effect requires local measurement.

An agent’s economic opportunity becomes particularly clear when an employee knows how to answer a question but gathering the information causes delay. The order number sits in one system, delivery status in another, and service conditions in a separate document. Connecting those records does not always require a difficult professional judgment. Nevertheless, these transitions can become a long wait for the customer.

An AI agent is a system that plans several steps toward a task, uses available tools and continues in response to results. Unlike an ordinary text response, its work can include retrieving and comparing information and preparing the next action. But the ability to write an answer does not confer authority to change an account. Finding a policy and approving a refund are different activities.

The title’s question therefore cannot be answered with a ranking of industries. Within the same company, an agent might quickly prepare an order history yet take longer to assess a disputed claim. Model capability is only part of the explanation. Task clarity, the location of authoritative information and the ease of checking the completed result also matter.

Where working time disappears

BTU’s service-management research distinguishes active processing time from the duration of the entire process. An employee may need little time to resolve a case, while the case waits a long time for attention. Agents address these problems differently. Automated retrieval can reduce active work; continuously available processing may reduce queues, provided that the final step does not still stop at a human bottleneck.

This distinction produces the article’s central contribution: task duration, issues resolved per hour and human attention are different measures. None alone establishes autonomous agent superiority. A speed claim must first identify the specific work and comparable outcome being assessed. If every prepared response accumulates in one manager’s queue, faster text production does not mean faster completion. The agent has changed the location of the wait.

A clear candidate is preparing an answer from a current, approved source. An order status, the existence of a document or completed application fields can be comparatively easy to verify. Where systems connect, an agent can gather information without requiring repeated employee navigation. The mechanism is reduced retrieval and transfer work, rather than a vague claim that the machine “thinks faster.”

Another candidate is the initial grouping of similar records. An agent could flag duplicates, missing fields or contradictory entries. Such processing is valuable when the human receives genuinely uncertain cases. If the employee must still reread every record from the beginning, the promised advantage becomes much narrower.

A third opportunity is gathering independent information in parallel. Several supplier documents could be retrieved simultaneously. Yet the answers must ultimately fit one decision. Parallel search does not eliminate reconciliation, version checking or an accountable decision. More agents are therefore not, by themselves, a measure of speed.

What workplace evidence establishes

The final published version of Brynjolfsson, Li and Raymond’s customer-support study, published in 2025, covers 5,172 workers. It reports an average 15% increase in issues resolved per hour with access to AI assistance. This is evidence from a particular company, not an estimate for Georgia’s service sector.

People retained responsibility for replies. The system suggested language and relevant information; employees could edit or reject its suggestions. The result demonstrates the potential of assistance, rather than the superiority of an independently acting agent. That distinction matters whenever an international percentage becomes part of a local purchasing case.

METR’s 2025 experiment with experienced developers found a different result: allowing the tools available at the time increased task duration by 19% on average. The finding does not describe all developers or today’s systems. It illustrates why a person’s familiarity with an environment and ability to check work can matter as much as the speed of generation.

METR’s February 2026 update is essential context. The researchers judged their later experiment an unreliable measure of the current effect because participant and task selection changed and concurrent agent use complicated time measurement. Neither the historical slowdown nor an uncertain later acceleration should become a universal promise.

These studies do not cancel each other out. They examine different work and different measures. Issues resolved per hour and time spent on a development task are not interchangeable statistics. BTU’s interpretation is that their joint lesson is to measure the actual process rather than select whichever foreign percentage suits a proposal.

Why the Georgian setting matters

For a Georgian company, information availability means more than internet access. The relevant condition must apply to the particular customer, product and date. Georgian correspondence, a supplier’s foreign-language instructions and local service rules may meet in one case. A response can be generated rapidly while reconciling their meanings still requires attention.

That creates a practical opportunity. Approved terminology, current conditions and clear source dates reduce uncertainty for both employees and agents. This preparation also improves the organization independently of AI. If a company reorganizes its information and introduces an agent simultaneously, attributing the entire improvement to AI would be inaccurate. A comparison should establish what the added system contributes to the already improved process.

In retail, an agent could retrieve delivery status and the current return policy before preparing a short response for an employee. In tourism, it could explain confirmed booking conditions in different languages. In education, it could assemble an application-document checklist. These are possible applications, not reports of measured Georgian pilots.

The shared economic mechanism is fewer information transitions and a result that is easier to check. Exceptions differ. Explaining a delay is not promising compensation, and preparing a document list is not deciding admission. Speed remains useful when the system performs the correct work within its authority.

For a small organization, oversight is a practical resource constraint. An employee who serves customers, checks records and handles exceptions may acquire additional work when an agent arrives. A useful system hands over a concise, evidence-backed case that is ready for a decision. A stream of long reports is not automatically assistance.

What a fair comparison measures

A fair comparison begins with the same completed outcome. Human and agent-assisted work should face equivalent quality requirements. If the human answer receives detailed scrutiny while the agent’s answer merely sounds fluent, the timing comparison is misleading. Errors that force customers to contact the business again are especially important.

BTU’s service framework emphasizes the cost and outcome of a resolved case rather than the cost of one contact. A quickly closed conversation can conceal an unresolved problem. If the customer returns, the work continues through another channel or employee. A shorter stage cannot then be treated as a saving for the entire organization.

Measurement should preserve the distinction between customer waiting time and employee attention. An agent may work in the background while the person does something else. Elapsed duration and human labor hours no longer describe the same thing. Both are useful, but combining them carelessly can create the appearance of counting the same saving twice.

Abandoned and transferred cases must remain in the record. Counting only easy, successfully completed tasks artificially improves the agent’s apparent performance. Escalation is often the correct behavior, but handoff and final resolution time still belong to the process. Assessment needs the complete workflow, not the most successful demonstration.

When quality is maintained and verification is brief, released time creates real possibilities: serving additional requests, reducing a backlog or giving employees more attention for complex problems. Time saved is nevertheless not automatically profit earned. Demand, organizational changes and a productive use for the released capacity determine whether the opportunity becomes economic value.

Where the advantage ends

An agent’s most credible speed opportunities involve clear information, verifiable preparation and fewer unnecessary handoffs. The advantage weakens where work depends heavily on tacit knowledge, conflicting objectives or irreversible decisions. Fast preparation can remain helpful, but it is not equivalent to a completed decision.

For Georgian companies, the starting point is therefore neither replacing an occupation nor buying the most elaborate agent. It is finding a segment of work where the correct outcome is known, the current source is visible and errors can be detected promptly. An agent works faster when the work it removes exceeds the checking and correction it adds. Measuring that boundary turns a technological promise into a business decision.

Questions and answers

Does a fast response mean a quickly resolved case?

No. Final resolution, repeat contact and corrections must also be checked.

Can an international productivity percentage be applied in Georgia?

It can provide context, but it is not an established expected result for a Georgian company.

Where should a company begin?

With retrieval from current sources and easily verified preparation, within clear authority boundaries.