A tiered framework for Agentic AI in localization
- Marina Pantcheva
- Jun 26
- 9 min read
The localization industry has long relied on automation to handle repetitive, time-consuming tasks. Traditional automation works well for structured tasks with clear input and output. However, certain tasks are beyond mechanical automation, like updating outdated translations of context-sensitive terms or handling complex morphological variations.
The emergence of agentic AI brought the promise of enhancing automation with creativity and flexibility enough to move it beyond the mechanical application of pre-defined rules. But what exactly is agentic AI, and how does it differ from traditional automation workflows? The term has been used so broadly that it's become nearly meaningless.
This article does not aim to advocate for agentic solutions, nor convince the reader that they are the answer to everything. The purpose is to bring clarity and define what agentic AI solutions are (regardless of how well and cost-efficiently they presently perform) by outlining varying levels of autonomy and complexity.
In discussing what types of agentic systems there are, the article follows established terminology in the AI domain and provides concrete examples of application. For simplicity, all examples are from the area of text-based AI, but the reader is encouraged to imagine multimodal applications as well.
What is Agentic AI, and how is it different from automation?
Traditional automation follows predetermined rules or instructions, performing exactly as programmed every single time. For example, a script that automatically removes duplicate entries in a translation memory will do exactly what it is coded to do – nothing more, nothing less. This type of automation is powerful for well-defined, repetitive tasks, but it cannot handle situations that fall outside its programmed code.
Agentic AI can largely overcome the rigidity of simple automation. Agentic AI systems are designed to act with a degree of independence, or “agency”, without step-by-step programming. Instead of just executing a fixed script, an AI agent makes context-dependent decisions and initiates actions to reach a pre-defined goal. In practical terms, this means that AI can retrieve information, analyze it, evaluate what needs to be done, and then carry out a plan, possibly adjusting along the way.
Agentic AI holds great promise for the localization field, but it's important to align the solution with the problem. In many cases, a simple script or a single-pass LLM can handle tasks more predictably and efficiently than a complex AI system.
The keyword here is "autonomous." Unlike scripts and basic automation, agentic systems can handle tasks for which they were not explicitly programmed. However, this autonomy comes with a trade-off: agentic AI does not always behave predictably. Ensuring that an AI agent makes the right decisions often requires careful design, testing, and human oversight.
The automation continuum: from rigid scripts to autonomous agentic systems
Automation exists on a continuum – from simple, rigid scripting to self-directed autonomous agents. To better define non-agentic and agentic automation solutions, it is helpful to map them across an autonomy scale of 5 levels, as shown in Figure 1.
On the left end of the scale, we have deterministic, rigid automation (Level 0 autonomy). As we move to the right on the scale, the autonomy of the systems increases, reaching Level 5, which is presumably the autonomy humans possess.
Let’s now zoom in on each level of autonomy on this continuum.
Level 0: Rigid rule-based automation
This is the most basic form of automation, using explicit rules or scripts (Figure 2). The system performs repetitive tasks following fixed logic, for instance, removing trailing spaces or converting time formats.
Such automation is a reliable solution for well-defined tasks where the results are consistent and easy to predict. Rigid automation is also relatively easy to implement and maintain if the task logic is straightforward. On the other hand, the deterministic nature of rigid automation makes it inflexible and brittle. If the conditions change or the input falls outside the expected parameters (for instance, the input data is not structured as expected), the automation might break or produce wrong results. It cannot learn or adapt on its own.
Level 1: Single-pass LLM
Single-pass LLM automation involves processing each request independently through a single prompt using a large language model (LLM) without iterative loops or memory (Figure 3). The flexibility of the system increases because LLMs can handle unstructured user input in the form of natural language.
This solution is quick to implement, requires simple infrastructure, and is ideal for high-throughput scenarios. However, it is susceptible to hallucinations, has limited consistency, and cannot reference previous interactions or context, making it suitable mainly for basic summarization, quick descriptions, instant translations of short and simple texts, and straightforward QA scenarios.
Level 2: Augmented LLM
Augmented LLM automation enhances a single LLM model with additional capabilities (Figure 4):
Retrieval-augmented generation (RAG): This enables the LLM to search for specific information in a database, for instance, check a Wikipedia page or query a database containing questions previously logged by translators and their corresponding answers.
Memory modules: This enables the LLM to store and recall relevant information across multiple interactions or sessions.
External APIs: They enable the so-called function calling (invoking external tools and services), which allows the LLM to access real-time data, perform calculations, and do more.
Augmentation enables an LLM to access external, real-time data, so its knowledge is no longer limited to what was contained in the training data. This improves the contextual understanding and reliability of the solution. However, its adaptability remains limited to the predefined enhancements. An augmented LLM solution is also much more complex from a technical perspective compared to working with single-pass LLMs.
Augmented LLMs are often used for first-pass translations of technical content with terminology and style guide rules being considered. Other popular implementations are knowledge-based chatbots, information retrieval systems, and translation assistant bots that answer translator queries.
Augmented LLMs are considered single AI agents (an “agentic unit”). They typically serve as components in more complex agentic workflows, like the ones discussed next.
Level 3: Agentic workflows
Agentic workflows integrate multiple LLMs, augmented LLMs, tools, and logic gates into an orchestrated sequence of tasks. These workflows follow a predefined, multi-step process, employing various components at each stage.
Agentic workflows have a high degree of autonomy, but their design is still somewhat fixed, which limits flexibility in unexpected scenarios. They also come in different configurations. The three most popular ones are detailed below.
Level 3.1: LLM chaining
This is by far the most popular AI implementation scenario as it leverages the capabilities of LLMs while maintaining control over the output through a pre-defined sequence of steps.
LLM chaining decomposes a complex task into a sequence of individual subtasks (monotasks). It then uses multiple chained LLMs, each specialized (prompted) for a specific task. The output from the first LLM serves as input for the second, the output from the second becomes input for the third, and so forth.
An excellent use case for this workflow is a smart term alignment tool, which replaces outdated terminology by going beyond the mere lexical form of a term (i.e., blindly matching words in source-target segments against glossary entries). This solution can disambiguate polysemous terms (such as bank, flight, and view) based on context. It also considers all morphological variations a term might undergo; so it knows that mouse and mice refer to the same concept, even though the words are so different.
Such a workflow operates as follows (Figure 5):
First, an AI agent (an augmented Term locator LLM) evaluates a source-target pair against a source term glossary to determine whether the source segment contains any terms that match the glossary. If no glossary terms are found, the process ends here, as there's no need to verify if a term is translated correctly, given that no term is contained in the source in the first place.
If a glossary term is identified, the Term locator LLM passes to the next AI agent a prompt containing the source-target pair and the relevant source and target terms. This Term validator agent checks if the target segment contains correctly translated target terms. If the terms are correct, the process ends here.
However, suppose the Term validator agent identifies an incorrect target term. In that case, it passes to the next Term updater agent a prompt containing the target segment, the incorrect term, and the correct term. The Term updater LLM then replaces the incorrect term with the correct one.
This approach breaks down the complex task of term alignment into manageable subtasks, ensuring both linguistic accuracy and contextual relevance.
Level 3.2: Parallelization
Parallelization involves LLMs that work simultaneously on the same task. Their outputs are then aggregated programmatically into one single response. Each LLM can approach the task from a different perspective. A good example for this workflow is AI-driven LQA, where a single segment must be simultaneously evaluated across multiple MQM dimensions, such as accuracy, terminology adherence, style guide compliance, fluency, etc.
In this workflow (Figure 6), a segment is sent for evaluation to multiple specialized “expert” LLM reviewers. Each reviewer is specifically prompted to evaluate the segment according to their particular expertise. For example, one focuses exclusively on terminology adherence, another on accuracy, and a third on style compliance. The outputs from these specialized LLM reviewers are then collected and passed to an Aggregator LLM, which synthesizes their feedback and resolves any potential conflicts.
Level 3.3: Routing
A third implementation of agentic workflows is routing. This workflow is a further increase in autonomy and flexibility as it does not follow a predetermined path. Instead, the first AI agent (an Orchestrator – typically a reasoning AI model) determines which one of several possible paths to follow. Specifically, it selects from among multiple downstream AI agents to handle the task.
Consider a scenario where mixed-source segments can be in one of three languages: German, Spanish, or French. An initial LLM identifies the language of the incoming segment and then routes it to the appropriate "translator” LLM. For example, if the identified language is French, the segment is passed to the French translator LLM (indicated by a solid arrow in Figure 7), while the other two translator LLMs (German and Spanish translators) receive no segments (indicated by dashed arrows).
Level 4: Agentic AI system (autonomous agents):
This is the most advanced stage of agentic workflows with the highest level of autonomy. Agentic AI systems are capable of pursuing complex goals through multiple steps. They can take a given objective, break it down into subtasks, invoke various tools or services, and adjust their strategy based on outcomes. These systems typically use planning algorithms or iterative reasoning processes. For example, an AI agent might internally plan a course of action, execute tasks, monitor progress, and revise its plan as needed.
Due to their adaptability and scalability, agentic systems have the potential to manage novel or unexpected tasks. However, their complexity poses challenges in design and management. Without robust safeguards, they can produce unpredictable results, so they need significant oversight.
Examples of tools built as agentic systems include autonomous coding agents, deep research agents, and travel planning agents. As of June 2025, to the best of the author's knowledge, no fully autonomous agentic system has been deployed within the localization industry.
Start from the left and choose what’s right
Agentic AI holds great promise for the localization field, but it's important to align the solution with the problem. In many cases, a simple script or a single-pass LLM can handle tasks more predictably and efficiently than a complex AI system.
When greater complexity is chosen, it should bring clear, measurable benefits to justify the added effort and potential risk. There's no need to deploy an agentic workflow just to remove duplicate segments from a translation memory when simple automation can handle it effectively.
To gain efficiency and reduce unnecessary complexity, it is best to start by exploring solutions on the left side of the autonomy scale. Begin with basic automation. If that's not sufficient, incrementally move to the more sophisticated approaches on the right.
Weighing practical and human factors
The exploration above focuses primarily on the technical dimensions for selecting an appropriate automation or AI solution. However, success depends on many more factors, such as the effort required for implementation, the time investment and skills needed for setup and maintenance, the associated costs (particularly when involving multiple AI components), the potential risks involved, and the cost of things going wrong.
Equally important is recognizing scenarios where human involvement, with or without AI assistance, remains the best course of action. For instance, critical legal documents or highly nuanced marketing materials often require the understanding and cultural sensitivity that only skilled human translators can provide.
Ultimately, the goal is not to automate everything, but to apply the right level of technology, or none at all, based on what best serves the quality, purpose, and nature of the task at hand.
Recommended further reading
Building effective agents by Anthropic
OpenAI's Practical puide to building agents