The short answer: yes, the market is moving that way
We may not fully abandon general-purpose AI models that do not speak the user’s language, but enterprises will increasingly stop accepting them as the final layer of interaction. A model that performs well in English but fails in the language, terminology, tone, and local context of the user is not enterprise-ready for that user.
This matters far beyond translation. Language is how employees express judgment, how customers describe problems, how regulators define obligations, and how organizations encode knowledge. When AI does not understand the working language of the business, it does not merely lose fluency. It loses operational accuracy.
That is why the recent publication of Soro, a family of large language models built specifically for Tajik, is more important than it may first appear. Tajik is not a dominant AI language. It has roughly 10 million speakers, limited training data, few accepted benchmarks, and deployment constraints such as weaker connectivity and limited hardware in parts of the country. In other words, it represents exactly the type of environment that most global AI roadmaps tend to overlook.
Why Soro is a serious signal
Soro was developed by researchers at Cornell and collaborators as a language-specific model family for Tajik. The team did not train a model from zero. Instead, they used Google’s open Gemma 3 as a base and applied continual pretraining on a carefully curated Tajik corpus of about 1.9 billion tokens. The data included web text, PDFs, and educational materials aligned with local curricula.
They then added supervised instruction tuning using 40,000 teacher-student style examples in Tajik, allowing the model to behave more naturally in dialogue rather than simply predict the next text fragment.
The reported outcome is meaningful: Soro outperformed same-size Gemma 3 models in Tajik while preserving competitive English performance on standard benchmarks such as MMLU and HellaSwag.
The strategic lesson is not that every organization must build its own foundation model. The lesson is that AI quality improves dramatically when language, domain, data, benchmarks, and deployment reality are treated as one system.
That is the part many organizations still miss.
This is not a translation problem
A common executive mistake is to assume that multilingual AI means taking an English-first model and placing a translation layer before and after it. That can work for lightweight tasks. It is not enough for high-value work.
Translation layers often fail where enterprises need AI the most:
- Local legal and regulatory terminology
- Customer complaints written in informal language
- Dialects, mixed-language writing, and cultural references
- Internal abbreviations and professional jargon
- Educational or healthcare content requiring exact meaning
- Operational workflows where small misunderstandings create financial impact
A finance team, hospital, insurer, public agency, or industrial operator cannot rely on “mostly understandable” language handling. The issue is not literary elegance. It is risk, cost, and process integrity.
The real breakthrough: benchmarks for neglected languages
One of the most important parts of the Soro work is not only the model. It is the benchmark infrastructure.
Because Tajik did not have strong public evaluation sets on major platforms, the research team built new benchmarks covering general knowledge, language competence, and entrance-exam style domains for schools and universities. They released them openly, allowing future models to be compared with greater transparency.
This is exactly how a market matures. Without benchmarks, every vendor can claim fluency. With benchmarks, organizations can ask better questions:
- Does the model understand the user’s language or merely generate plausible text?
- Does it perform well in the domain where it will be used?
- Does it preserve capability in other required languages?
- Does quantization reduce quality below an acceptable threshold?
- Can performance be measured before procurement and after deployment?
For enterprises, benchmarks are not an academic luxury. They are procurement infrastructure.
Deployment economics: why FP8 and INT4 matter
The Soro researchers also tested quantization in FP8 and INT4. In practical terms, quantization compresses model weights so the model requires less memory. The key finding was that much of the Tajik improvement remained intact while memory requirements dropped significantly.
That matters because many real-world deployments do not happen inside ideal cloud environments. Schools, clinics, factories, field offices, and public-sector departments often have constrained hardware, unstable connectivity, or strict data-residency requirements.
A model that can run on cheaper devices is not merely technically elegant. It changes the economics of adoption.
For CFOs and operations leaders, the question becomes very concrete: can we deliver useful AI to the edge without turning every interaction into a cloud expense, a latency problem, or a security exception?
In many cases, smaller language-adapted models may beat larger general models on total business value.
What enterprises should learn from this
Soro offers a template that can be applied to Arabic dialects, Amharic, Bengali, indigenous languages, Israeli Arabic, Russian-speaking customer segments, and internal enterprise language that no public model fully understands.
The methodology is practical:
- Start with a strong open or commercial base model.
- Curate a serious corpus in the target language and business domain.
- Continue pretraining instead of relying only on prompts.
- Build instruction data that reflects real user interactions.
- Create benchmarks before scaling deployment.
- Test quantization and edge deployment early.
- Keep human review where judgment, risk, or ethics require it.
This is where AI stops being a technical toy and becomes organizational capability.
Human in the loop, but not human as the bottleneck
Language-native AI is especially powerful because it can handle non-deterministic processes: tasks where there is no simple rule engine, where judgment matters, and where human language carries the decision context.
But the right answer is not to put a person in front of every AI output forever. That simply recreates the old bottleneck with a new interface.
The real operating model is different: one professional who previously executed or supervised one process should now be able to supervise dozens or hundreds of AI-supported processes. The human remains responsible for exceptions, escalation, quality control, and policy interpretation. The AI handles volume, preparation, drafting, classification, routing, and first-level reasoning.
This is the difference between AI theater and operational efficiency.
The two adoption tracks: literacy and agents
Organizations should move on two tracks at the same time.
The first track is AI literacy. Employees must learn how to communicate effectively with models, evaluate outputs, protect sensitive information, and understand where AI is useful or dangerous. This is not optional training. It is becoming a core workplace skill.
The second track is AI agents. Agents can perform defined processes across systems, data sources, and workflows. Unlike generic AI tools, agents often require less behavioral change from employees because they can be embedded into existing operations. Technically, agents may look more complex. Organizationally, they can sometimes be easier to adopt.
That creates a new requirement: every serious organization needs an efficient platform for building, deploying, monitoring, and governing AI agents. Microsoft Copilot Studio is a reasonable option for companies deeply invested in the Microsoft ecosystem. Tools such as n8n are also entering larger enterprises in ways that would have looked unlikely a few years ago. Claude remains one of the strongest enterprise AI experiences in practice, especially with capabilities such as Claude Code, though security architecture must be handled carefully.
The broader point is vendor-neutral: information systems departments are likely to become a kind of human resources department for AI agents. They will onboard them, assign permissions, monitor performance, retire underperforming agents, and maintain governance.
Why academic depth still matters
Soro is also a reminder that serious AI depends on serious research. The strongest work in this field is multidisciplinary. It is not only computer science. It combines linguistics, education, management, business process design, evaluation science, data engineering, and applied deployment experience.
This is why organizations should be cautious with self-appointed AI experts who sell generic advice without deep technical understanding or real operational experience. Large enterprises usually have enough internal filters to reduce the damage. Small and mid-sized businesses are more exposed.
Good AI implementation requires education, field experience, and managerial judgment. It requires understanding the process before automating it. It requires knowing when a foundation model is enough, when retrieval is enough, when fine-tuning is justified, and when a smaller dedicated model is the better financial decision.
The strategic question for leaders
The question is no longer “Which AI model is best?” That question is too shallow.
A better question is: “Which model, data, benchmark, workflow, governance model, and human oversight structure produce reliable outcomes for our users?”
For some organizations, the answer will be a general model from OpenAI, Anthropic, Google, or Microsoft. For others, it will be a language-adapted model. In many mature environments, it will be a portfolio: broad foundation models for general reasoning, dedicated models for local language and domain work, and agents that operationalize both.
Soro points to a future where AI speaks closer to the user’s actual language, not only the dominant languages of the internet. That future is more inclusive, but it is also more efficient, more accurate, and more financially rational.
The companies that understand this early will not treat language adaptation as localization. They will treat it as infrastructure.
