Recent Posts

AI agents

How an AI agent frees people from routine?

Main BTU’s analytical assessment is that an AI agent reduces routine when people no longer have to repeatedly assemble

How an AI agent frees people from routine?

Main

BTU’s analytical assessment is that an AI agent reduces routine when people no longer have to repeatedly assemble and compare information and can focus on genuine changes. For Georgian businesses, realizing this opportunity requires a reliable case record, explicit exceptions and verification, because a shorter list of messages is not yet a measured saving in human time.

Key Takeaways

  • An AI agent can usefully retain the previous state of a case and compare incoming information against it, rather than repeatedly presenting unchanged work as a new decision.
  • In BTUAI’s single synthetic Georgian-language run on 6 November 2026, five new or changed cases were identified among 12 incoming records; seven records repeated existing information. This describes a constructed test stream, not Georgian business statistics.
  • Evaluating relief from routine requires measuring human review, corrections and service quality. This demonstration measured neither human time nor financial benefits.

A substantial part of a working day can pass without a new decision being made. Someone checks a message, finds an old file, compares two records and asks again whether anything has changed. These actions can be necessary: without them, a deadline disappears, a customer receives no answer or colleagues work from inconsistent information. Their necessity does not mean that people must reconstruct the information from scratch every time.

This intermediate work is a promising place for an AI agent. Here, an agent means a system that can choose steps toward a goal, use permitted tools and continue in response to their results. Its task need not be writing another isolated piece of text. It can establish the state of a case: what was known, what arrived, what changed and who needs to become involved. A colleague then receives material prepared for a decision rather than another collection of disconnected messages.

That promise needs a straightforward test. If the system returns the same ten cases every morning in different words, the reading burden remains. If it removes the one problematic case, its list becomes shorter while service deteriorates. A useful result therefore combines complete retention of meaningful changes with removal of unnecessary repetition. Brevity alone is an inadequate measure.

Routine often means reconstructing the state of work

Consider an illustrative procurement process. Yesterday’s record contained a quantity, delivery date and responsible colleague. Today, the same information arrives through another channel, followed by a corrected quantity and then a message specifying a different date. A person first has to connect each message to the right case, then distinguish a repeat from an agreed amendment or a contradiction. Only after that can the decision itself begin.

An agent can undertake this preparation if it has a stable case identifier, the previous state and evidence of where each statement came from. A stable identifier keeps references to one order together across messages. Previous state prevents every newly received sentence from being treated as a change. Source evidence establishes who supplied a different condition and when. Without these elements, the agent may be little more than a persuasive summarizer.

New information and a new message are different events. A customer resending the same request should not create another case. Conversely, almost identical text may carry an important single-number change: 85 units instead of 80, or delivery on Wednesday instead of Monday. Removing repetition must not conceal those differences. A smaller review queue is valuable only if decision-relevant information survives.

Not every step needs an agent

Finding an exactly identical record, adding numbers and standardizing dates can often be handled by ordinary software. A language model is more useful where the meaning of differently worded statements needs interpretation. “Please cancel the order,” for example, is a request, not evidence that cancellation has already happened. Preserving that distinction prevents premature closure.

Anthropic’s discussion of agent design distinguishes predefined workflows from model-directed actions and recommends beginning with the simplest adequate solution. BTU’s synthesis applies that principle to routine work: use language interpretation where it is needed and explicit rules or code for precise checking. Adding an agent is not, by itself, an improvement.

Nor must an agent finish everything independently. Stopping at an uncertain delivery date can be the correct outcome. When a supplier’s message conflicts with a procurement record, the system cannot invent the organization’s rule for which source takes precedence. Its useful contribution is to expose the conflict and prepare a specific question. The person no longer has to search an entire conversation for the problem, but still makes the final choice.

Why the distinction matters in Georgia

GeoStat’s enterprise ICT survey reports that 94.9% of enterprises had internet access in 2025 and 15.3% used a website. The release was published on 29 May 2026. Internet access does not measure workflow integration, and lacking a website does not establish technological backwardness. The figures nevertheless help distinguish access to a digital channel from a coherent digital work process.

The distinction has practical relevance for trade, tourism, education and professional services. Each may involve messages, spreadsheets and separate status records. These are potential applications, not sectors where an effect has already been measured. The relevant need is not simply another interface. It is connecting a Georgian-language request to the correct case and bringing unresolved conditions to a person’s attention.

In a small team, one colleague may answer customers, check orders and prepare management information. An agent’s economic role may emerge through continuity of preparation across those tasks. But if that colleague has to repair every update, the system creates hidden additional work. Adoption should therefore be judged against the quality of a specific process rather than a demonstration of software features.

Language itself belongs in that evaluation. A system must correctly interpret dates, quantities, negation and the status of an action. “Not yet cancelled” and “cancelled” describe opposite states. Fluent Georgian does not automatically establish that the distinction has been respected. A trial should include amendments, repeats and contradictions, not only clean records with obvious answers.

Less routine and higher productivity are different measurements

International research illustrates how workplace effects can be measured. A 2025 paper by Brynjolfsson, Li and Raymond, using data on 5,172 customer-support workers at one firm, associates AI assistance with an average 15% increase in issues resolved per hour. Effects varied with worker experience. This was a human-assistance system, not a Georgian trial of autonomous agents, so its estimate cannot simply be transferred to Georgian businesses.

The useful lesson lies in the outcome being measured. A case should be completed well, not merely closed quickly. An agent may produce a short brief that takes a colleague a long time to check. It may preserve an outdated condition whose consequences appear only when a customer complains. Excluding those costs makes apparent savings too favorable.

Relief from routine begins as a process change: less repeated searching, a clearer current state and more visible exceptions. Observation of real work is needed to establish whether that change becomes saved employee time, better service or additional income. The demonstration below examines an early part of that chain without claiming to measure the rest.

Part two The BTUAI pilot

What remained for review from twelve incoming messages

On 6 November 2026, a small Georgian-language demonstration was executed for BTUAI. Could an AI worker separate new and changed cases from a recurring information stream without concealing uncertainty? Twenty-four synthetic text records were created in advance. Twelve described the baseline and twelve represented a subsequent intake. These were not real company messages or field observations involving human participants.

The second intake contained six unchanged updates, three changed cases, two new cases and one retransmitted message. The changes included an agreed quantity correction, conflicting delivery dates and a cancellation request. The AI worker received both batches and instructions to preserve case history, identify changes and perform no external action.

The exercise ran once. The AI worker interpreted and compared the Georgian texts and produced a structured result; Python supported data processing and count verification. The object of the demonstration was an actually executed internal text-processing task, not a deterministic program relabeled as an agent. A separate AI reviewer checked the output against the original records; this was not human expert review.

Measure Single synthetic run
Baseline cases 12
Subsequent incoming records 12
New or changed cases 5
Repeated records excluded from the change queue 7
Distinct cases retained in the final state 14
Exceptions reserved for human decision 2

Source: BTUAI’s author-created Georgian test stream and executed AI processing, 6 November 2026. Units are records or cases as indicated in each row. Human time was not measured.

BTU’s calculation shows that five of the twelve second-intake records entered the change queue while seven repeated existing information. Compared with listing every incoming record, the queue had 7/12, or 58.3%, fewer entries. This is a compression measure for this particular constructed stream. It does not establish a 58.3% reduction in working time, labor costs or staffing requirements.

The difficult cases that remained visible

One case makes the boundary clear. The initial record scheduled delivery of 70 units for 16 November. A subsequent message said that the supplier now named 18 November, while the procurement record still specified 16 November. Neither source was given precedence in the exercise. A useful result retained both dates and flagged the disagreement rather than automatically treating the newest message as the final commitment.

The second exception concerned cancellation. The customer’s request was recorded without representing the order as already cancelled. Separately, an agreed quantity amendment from 80 to 85 units was recorded as a confirmed change. These cases explain why shortening text cannot substitute for processing a case. One amendment changes a record, another leaves conflicting sources unresolved, and a third requires an authorized decision.

The number of retained cases also matters. Two new cases were added to the initial twelve, leaving fourteen in the final state. Cases absent from the second intake were not deleted. Otherwise, filtering noise would also have hidden outstanding obligations. A concise update and a complete work record serve different purposes; a useful system preserves both.

What the demonstration cannot tell us

The embedded changes were identified in this run, but success does not establish general accuracy. The records were small in number, clean and constructed for the task. They did not include scanned documents, ambiguous names, missing attachments, changed case identifiers or tool failures. There was no comparison with ordinary software, so the exercise did not establish an advantage for an agent over simpler automation.

There was also no human control group or timing study. How long an employee would take to verify the output, and how helpful the brief would be during a real working day, remain unknown. Absence of evidence on those questions must not become a productivity promise. A useful next step would be observation of quality and total labor requirements in a real, authorized workflow.

Conclusion

An AI agent frees people from routine when they no longer have to reconstruct the state of work and can concentrate on exceptions. The small demonstration showed a possible internal outcome of this mechanism: repeated messages were filtered, meaningful changes retained and uncertainty not misrepresented as a decision. For Georgian businesses, delegating that verifiable work is a useful objective. Savings in time and money remain a separate outcome to establish in practice.

Questions and answers

Which routine is a suitable starting point

Comparing a case’s previous and current state, provided the source records are available and changes can be verified.

Does filtering seven records mean seven fewer cases

No. The demonstration filtered repeated information while preserving every distinct case in the final state.

Why were two issues left for a person

One involved conflicting delivery dates and the other a cancellation request. Final resolution exceeded the information and authority supplied.

Was a time saving established

No. The exercise demonstrated record processing and change identification. Human time, costs and service outcomes were not measured.

Research basis

Research question: Can an AI worker separate new and changed cases from a recurring Georgian-language stream while preserving case history and unresolved issues?

Method: On 6 November 2026, a single small synthetic demonstration was executed for BTUAI. An AI worker compared constructed texts and produced the final case state, change queue and exception queue. Python supported processing and count checks; a separate AI reviewer compared output against the source records. This was not human expert review, a real-company field study or a blinded test.

Data and period: The author-created stream dated 6 November 2026 contained 24 records: 12 baseline and 12 subsequent records. The second intake comprised six unchanged updates, three changed cases, two new cases and one retransmission. Context comes from Geostat’s 2025 enterprise ICT indicators, released on 29 May 2026, the 2025 QJE workplace study and Anthropic’s 2024 design guidance. These contextual sources do not independently validate the pilot result.

Main result and units: BTU’s calculation places five cases in the change queue: three changed and two new. Seven repeated intake records were excluded from that queue, while 14 distinct cases remained in the final state (12 + 2). The two exceptions, conflicting delivery dates and a cancellation request, are a subset of the five queued cases, not additional cases. Queue-entry reduction is (12 − 5)/12 × 100 = 58.3%. The denominator is second-intake records, not distinct cases or working hours.

Limitations: One small constructed stream used stable identifiers and clean text. Human time, review effort, costs and service outcomes were not measured; there were no repeated runs or ordinary-software comparator. The 58.3% result is not time, cost or staffing savings and does not estimate general accuracy or Georgian business productivity. Geostat’s indicators do not measure AI adoption.

Primary sources

BTUAI author-created Georgian synthetic records, execution output and separate AI review, 6 November 2026: no public URL.

Geostat 2025: https://www.geostat.ge/media/79412/Use-of-Information-Communication-Technologies-in-Enterprises—2025.pdf

QJE 2025: https://academic.oup.com/qje/article/140/2/889/7990658

Anthropic 2024: https://www.anthropic.com/research/building-effective-agents