The race for AI in the workplace is starting to shift. Previously, companies competed to find models with the highest capabilities. Now, the question has become more practical: which model is most suitable for a specific task, what is the cost, and when do humans need to approve the results?
This direction is evident from two announcements made in early October 2026. Google Cloud introduced the Gemini agent, a system designed to plan work, utilize various tools, connect to business systems, and select models that meet the needs. On the other hand, Anthropic released Claude Haiku 5.5 as a smaller, faster, and cheaper model for repetitive tasks with high volume.
Together, they suggest that the future of AI in the workplace may not be determined by a single super-powerful model, but rather by systems that can efficiently distribute work across multiple models.
From Chatbots to Work Organizers
Chatbots typically operate with a simple pattern: the user asks a question, and the model responds. Agentic systems work on a larger scale. They can break down requests into several steps, call tools, read data, update documents, and return the final results.
In its official announcement, Google Cloud described the Gemini agent as a universal agent for work. The system is designed to understand organizational context, utilize skills and tools, and operate within environments already used by companies, such as documents, inboxes, and development environments.
The most interesting aspect is not just its ability to handle multiple steps. The Gemini agent is also said to be able to select the best model for a specific job and maintain cost control. This means that simple tasks do not always have to be sent to the most expensive model.
For example, a request to “create a summary of the weekly sales report” might be adequately handled by a fast model. However, when the system needs to identify the causes of a sales decline, compare multiple data sources, and make recommendations, the task could be shifted to a model with higher reasoning capabilities.
Conceptually, this is similar to a company not asking the director to handle all tasks. There are staff who manage routine tasks, analysts for work that requires deeper examination, and specialists who are called in when the issues are more complex.
Smaller Models Are Not Unimportant
The focus on large models often makes smaller models appear as less capable versions. In fact, in AI systems that are used continuously, speed and cost can be just as important as the ability to answer difficult questions.
Anthropic refers to Claude Haiku 5.5 as a small model aimed at repetitive and cost-sensitive tasks, such as creating summaries, clustering data, running database queries, assisting customer service, and serving as a subagent for coding tasks. The company also stated that its average operational cost is about 75 percent lower than Haiku 4.5.
This figure is a claim from Anthropic, not the result of independent testing. However, the direction of its product is quite clear: not every part of a workflow requires the same expensive model.
In practice, an agent might use a small model to read 1,000 customer support tickets, categorize complaints, and flag recurring cases. Only tickets that carry high risk or require more complex reasoning would be sent to a more powerful model.
Real Benefits for Daily Work
- Faster responses. Simple tasks can be handled by lightweight models without waiting for lengthy processes.
- More controlled costs. Companies do not need to pay for premium models for every request.
- More flexible workflows. Systems can combine language models, databases, office software, and internal tools.
- Larger scale of use. Cheap and fast models are more suitable for tasks that need to run thousands of times each day.
For small teams, this approach can help create automation that was previously too expensive. For example, an online store can use lightweight models to classify customer inquiries, then forward refund cases or serious complaints to humans.
Risks: Automated Decisions Are Not Always Easy to Trace
The more decisions are delegated to systems, the more important it is for companies to understand what happens behind the scenes. If an agent selects a model, calls data, and then takes action, users need to be able to trace those steps.
There are several risks to be aware of:
- Choosing the wrong model. Tasks that seem simple can have significant consequences if sent to a model that is too limited.
- Data out of context. Agents may combine outdated data, different metric definitions, or unupdated documents.
- Costs can still escalate. Systems that conduct too many trials, read documents repeatedly, or call tools without limits can become expensive.
- Blurred accountability. When results are incorrect, companies must be able to determine whether the issue lies with the data, model, instructions, integration, or human approval.
Therefore, “automatically selecting models” does not mean companies can let AI operate without oversight. Systems still require cost limits, a list of actions that need approval, activity logging, and ways to test results.
What This Means for Users and Businesses
This change indicates that the use of AI will increasingly resemble the management of digital labor. Users will not only write prompts but also define goals, authority limits, data sources, and outcome standards.
For individual users, the principle is simple: choose tools based on the task, not the popularity of the model. Fast models are suitable for summarizing, reformatting, or creating initial drafts. More powerful models are worth using when tasks require comparison, reasoning, or checking multiple sources.
For businesses, a sensible first step is to create a task map. Record repetitive tasks, the data needed, the level of error risk, and when humans should check results. From there, companies can determine which tasks are suitable for automation with lightweight models and which should continue to use more powerful models.
What Can Be Done Now
- Separate routine tasks from high-risk tasks. Do not use the same automatic rules for creating summaries and approving transactions.
- Measure cost per outcome. Calculate not only the model price but also the time for review, failures, and the use of additional tools.
- Use tiered approvals. Agents can draft automatically, but submissions, data deletions, or financial changes require human approval.
- Keep a decision trail. Log the models used, data read, tools called, and changes made.
- Test with real cases. Evaluate agents using examples of daily work, including ambiguous and messy cases.
These recent developments do not mean that all work will immediately become automated. However, the direction is quite clear: AI at work is moving from “answering models” to “systems that organize how to complete tasks.” In this phase, the smartest models are not necessarily the best choice. The most useful systems are those that know when to use high capabilities, when it is sufficient to use lightweight models, and when to pause to ask humans to make decisions.
Sources & Further Reading
– Rio Yotto @rioyotto