AI Transformation in 2026 is moving from pilots to real operations. Explore key trends, statistics, APAC growth, and Vietnam’s delivery opportunity.
According to MIT NANDA, 95% of enterprise GenAI pilots fail to deliver measurable impact on profit and loss. At the same time, 31% of enterprises have already put at least one AI agent into real operation.
These two seemingly contradictory numbers capture the state of the AI market in mid-2026. After the early excitement around chatbots, writing assistants and internal demos, many companies are beginning to see the distance between “trying AI” and turning AI into part of business operations. A tool may answer quickly, write well or summarize documents more accurately than people in certain situations, but its impact on revenue, cost, productivity or risk only appears when that tool enters the right workflow and changes how daily work is handled.
>> AI Agent vs Chatbot vs AI Assistant
From Digital Transformation to AI Transformation
Digital Transformation used to be associated with cloud, ERP, CRM, digitized records, data dashboards and projects to replace legacy systems. By 2026, a new layer of transformation is taking shape under the name AI Transformation, or AX, as AI begins to participate directly in how companies assign work, process information, make decisions and monitor risk. The move toward AX does not resemble the way many companies approached DX before, because with ordinary software, a company can buy licenses, train users and wait for productivity to improve over time; with AI, value appears only when the model is connected to real data, real workflows, real access rights and control mechanisms tight enough for real operations.
What Is AI Transformation?
AI Transformation, or AX, is the process of redesigning how a company works so that AI can take part in real business operations, rather than staying as a tool people use on the side. It means connecting AI to the company’s data, workflows, systems, access rights and decision points, then measuring whether it changes cost, revenue, speed, quality or risk. In that sense, AX is deeper than simply adopting more AI tools. A company only starts to transform with AI when the technology changes how work is assigned, how information moves, how decisions are checked and how outcomes are measured.
When Chatbots Enter the Operating Line
AI agents are changing how companies use AI, starting with a move beyond the familiar question-and-answer format. In the past, an employee would open a chatbot, type in each question and decide the next step based on the answer. In an agent model, a company can give AI a specific goal; the system then breaks the work into smaller tasks, calls the tools it needs, passes part of the work to a specialized agent, checks the result and sends high-risk steps back to humans for approval.
This way of working fits office tasks that repeat often but are not completely fixed. A financial reconciliation process, an internal ticket workflow, a sales report or a customer follow-up email all have a familiar structure, yet the input data, exceptions and handling requirements may change every day. Traditional RPA often struggles in these situations because it depends heavily on fixed interfaces and rigid rules, while an agent can read context, understand natural-language requests and choose a more suitable next step.
Connecting agents to strong models such as Claude or ChatGPT can help the system understand language better, but language understanding is only one part of the problem. For an agent to actually work inside a company, it needs to know: which tools it is allowed to use, which data it is allowed to read, which systems it can call and which other agents it can coordinate with.
- Model Context Protocol, usually called MCP, is being used as a standard for connecting agents with tools and data sources
- Agent-to-Agent, or A2A, opens another path for agents from different platforms to work together.
As these connection standards mature, enterprise agents have a better chance of operating with clearer order. Instead of each department installing its own chatbot and finding its own way to connect files, email, CRM or internal systems, a company can build a shared way for agents to access tools and coordinate with one another. This step matters if AI is to move from supporting individual employees to handling a chain of work that involves several departments.
The clearest signs of adoption are appearing in industries with large data pools, standardized processes and strong pressure to automate, especially banking and insurance. According to the report, 47% of companies in this group have at least one agent in production; JPMorgan is reported to have more than 450 agent deployments in its technology portfolio.
In software development, MIT Sloan research found that coding agents increased the number of shipped features by 14% per engineer-quarter, while sales development representative agents recorded a median payback period of about 3.4 months.
Once an agent is allowed to touch payments, source code, customer data or credit decisions, companies have to manage it the way they manage an account with permission to act inside the system. Which data an agent may read, which actions it may trigger, how its actions are logged and where it must stop for human review all become very concrete operating questions. Without these boundaries, the better an agent is at automation, the faster it can amplify mistakes.
Cheaper Models Change How Companies Buy AI
In the past, many companies assumed that using powerful AI meant renting APIs from a small number of major providers, with costs charged by each processing request. By 2025, that assumption had begun to change as the cost of running models fell quickly. From late 2022 to late 2025, the cost of achieving GPT-4-equivalent performance dropped from about $20 to $0.40 per million tokens, a decline of roughly 50 times.
The sharpest cost drop happened in the stage after a model has already been trained. Every time a company sends a request to AI and receives a result, it pays for the amount of data being processed; the market usually calls this stage inference. In 2025, competition in this exact stage pushed token prices down by more than 90%, making some use cases that were once too expensive to run regularly become more feasible.
DeepSeek R1 became a milestone that forced companies to rethink how they buy AI. The model was released with a much lower stated training cost than comparable Western models at the time, while also offering a low cost for running the model per million tokens. The training-cost figure still needs to be read carefully because it does not include the full cost of research and experimentation, but DeepSeek’s market impact is clear: companies began to look more carefully at whether they should rent all reasoning capability from an external API.
As costs fall, companies gain more ways to allocate work to AI. Difficult tasks that require deep reasoning or carry major impact can still use the strongest models from large providers. Repetitive, lower-risk tasks can move to cheaper models. For sensitive data, a company can choose to run the model on its own infrastructure to control where data is stored and processed. Buying AI is therefore beginning to look more like a problem of allocating cost, data and risk than choosing a single tool.
Sovereign AI comes from a very practical concern: where does the data go when a company uses AI? For a bank, that data may be transaction history and credit records; for a hospital, medical records; for a factory, blueprints, production formulas or operating data. When this data has to be sent outside for processing, the company faces security risks while also depending on the provider’s prices and policies. That is why many organizations are starting to want models to run inside infrastructure they control.
The rise of DeepSeek, Qwen, Llama and other open-weight model families makes this option more realistic. If a capable enough model can be deployed in a private environment, the company gains more control over both data and cost. Competitive advantage therefore shifts toward what is harder to copy: how well transaction history, technical documents, operating processes, customer knowledge and internal experience are organized so AI can actually use them.
Data Is the Biggest Bottleneck in AI Production
Many AI demos look convincing until they meet a company’s real data. In a test environment, AI is often given clean documents, clear questions and a narrow scope. Once it enters operations, it has to deal with old contracts, wrong file versions, customer data with missing fields, outdated technical documents and important decisions scattered across email or personal files.
This gap appears clearly in market numbers. According to surveys summarized in the report, 78% of companies feel unprepared for GenAI because their data foundations are weak, while only 22% rate their data as “very ready.” McKinsey found that 71% of organizations use GenAI regularly, but only 17% say GenAI contributes more than 5% of EBIT. In other words, many companies are already using AI, but they have not yet created value that is clear enough in business results.
The problem often starts with how knowledge is stored. Contracts sit in the legal folder, customer data sits in the CRM, technical instructions sit in the wiki, issue histories sit in the ticketing system, and part of the operating know-how exists only in the heads of long-time employees. When AI only retrieves isolated text passages, it may find a sentence that looks close to the question, but it struggles to understand the relationship between customers, contracts, clauses, responsibilities, risks and exceptions in the real workflow.
RAG was once seen as a quick way for AI to look up documents before answering. This approach is like finding the passage that seems closest to a question and giving that passage to AI before it writes a response, so it works reasonably well for simple requests such as finding a clause in a contract or summarizing a document. But in a company, many questions do not sit inside a single paragraph. To know whether one clause conflicts with a master agreement, AI has to understand who the customer is, which contract is still valid, which clauses are related, which department is responsible and which internal rule applies.
GraphRAG and knowledge graphs emerged to handle this type of relationship. Instead of only finding text similar to the question, the system tries to organize knowledge through links between objects such as customers, contracts, products, suppliers, workflows and risks. In sectors such as finance, law, pharmaceuticals or supply chains, the ability to follow these links matters more than a fluent-sounding answer, because the final decision often depends on several data sources at once.
The operating rhythm of data also changes when AI agents take part in workflows. Companies are used to checking data monthly or quarterly, while an agent needs to know almost continuously which source is still trustworthy, which version is latest, who has permission to view it and which information has become outdated. With a dashboard, one wrong number often stops at a wrong report and can be caught when a person reviews it. With an agent, a small error can travel much further.
A wrongly updated receivables record, for example, may cause an agent to misread a customer’s status. From that wrong judgment, the system may send a payment reminder at the wrong time, move the case to the next handling step, create a ticket for the customer service team and leave employees spending hours undoing a chain of actions that has already run automatically. The risk of bad data is therefore not just one incorrect number, but the fact that this number can become the starting point for many actions that follow.
The data foundation for AX cannot be just a large storage repository. A company needs to know where data comes from, who can access it, when it changed, which workflow it relates to and whether it is trustworthy enough for an agent to use. As data governance platforms are repositioned as infrastructure for AI-ready data, the market is moving toward a more practical question: does the company have data that is clean enough, permissioned correctly and placed in the right context for AI to act safely?
AI Moves Beyond the Chat Window
AI is moving beyond the chat window and into voice, images, video and software interfaces. Real-time speech-to-speech models have reduced latency to a level close enough to natural conversation, allowing AI to take part in calls instead of only replying in text. Multimodal models can process text, images, diagrams, audio, video and screenshots in the same flow, which means many tasks that once required people to look, listen and compare manually can now be partly automated.
Another direction, often called computer use, allows agents to operate directly on software interfaces. Instead of waiting for a company to build APIs or integrate systems, an agent can click, type, scroll, switch screens and move through steps the way an employee would on a computer. The technology is still early, but it opens a noteworthy approach to legacy systems.
The meaning of this direction becomes clearer in companies that still run on ERP, CRM, internal portals or specialized software with weak APIs. In the past, automation often got stuck at the integration stage because technical documentation was incomplete, data lived in many different forms and workflows depended on the user interface. With a computer-use agent, part of the work can be handled directly on the screen employees already use every day, as long as the task is clearly bounded and the output can be checked.
The first applications usually appear in contact centers, healthcare, manufacturing and IT operations. Voice agents can handle part of customer calls in place of rigid IVR trees. Ambient AI can help record conversations in clinical settings. Vision models can inspect defects on production lines. Computer-use agents can move through ERP or ticketing screens to carry out some repetitive actions. What these applications share is that AI is not only answering; it is beginning to take action.
For a market such as Vietnam, where many companies use fragmented systems and lack APIs, this direction is practical. A company does not have to complete an entire system modernization project before trying to automate part of a workflow. The key conditions are to limit access rights, define the task clearly and keep human review at points where a mistake could cause damage.
AI Production Needs to Be Audited Like a Business System
The failure rate of GenAI projects is forcing companies to look at AI in the language of operations. MIT Project NANDA found that 95% of enterprise GenAI pilots do not deliver measurable P&L impact; IDC and Lenovo reported that only 4 out of 33 AI proofs of concept reach production. Gartner also found that more than half of GenAI projects are abandoned after the POC stage because of poor data quality, weak risk controls, rising costs or unclear business value.
A mature AI project cannot be judged only by the number of trial users, the number of prompts run or the estimated hours saved by employees. Companies need to see impact on cost lines, revenue, processing time, error rates, departmental productivity or service quality. When AI enters risky workflows, requirements such as logging the processing trail, explaining why the system produced a result, controlling prompt injection, preventing data leaks and proving compliance need to be considered from the design stage.
The LLMOps layer appears from that need. Model observability platforms such as Langfuse, Arize, Weights & Biases and Datadog LLM monitoring help track what data AI received, which steps it went through, what result it returned, how much it cost and whether there are signs of error. For an AI system that operates every day, this monitoring is like attaching gauges to a production line: if the company does not know where an error appears, it cannot fix the problem before it spreads to customers or other workflows.
Risk frameworks also become more important when agents have the right to act. OWASP LLM Top 10 puts prompt injection among the key risks because a malicious instruction can trick the system into doing something unintended if access rights are not controlled. AI TRiSM and the EU AI Act are also making risk governance a more familiar requirement in technology procurement, especially for systems that affect hiring, credit, education, healthcare or public services.
Companies that create value from AI often start with a narrow but painful problem. It may be a workflow that consumes many hours each week, has reasonably stable data, has a clear owner and has metrics that can be measured after a few months. This approach is less attractive than demos that appear to do everything, but it is closer to where AI can create revenue, save cost or reduce errors in everyday operations.
For peole who are paying attention, HBLAB is providing access to our AI Roadmap Whitepaper, which gives leaders a practical structure to assess where AI can move from pilot activity into controlled production, and where stronger governance is needed before scale.
APAC Becomes a Major AI Deployment Region
Asia-Pacific, or APAC, is becoming one of the fastest-growing AI regions in the world. IDC estimates that the APAC AI market reached $102 billion in 2025 and could rise to $175 billion by 2028, equal to a compound annual growth rate of 33.6%. GenAI investment in the region is projected to reach $54.5 billion by 2028, while Microsoft recorded that 53% of APAC organizations have used agents to fully automate some workflows, higher than the global average of 46%.
APAC does not operate as a single, uniform AI market. China has major influence through its open-weight model ecosystem. India has advantages in talent and large-scale IT services. Singapore plays the role of a governance and regional experimentation hub. Japan and South Korea are bringing AI into industry, robotics and manufacturing automation. What these markets share is fast adoption, a strong government role and high demand for AI deployment in real enterprise environments.
China has created global impact through Qwen and DeepSeek. Alibaba’s Qwen has surpassed 400 million downloads on Hugging Face, while DeepSeek R1 helped push the competition around model-running costs further in 2025. For many APAC countries, these open-weight models offer a capable and low-cost foundation for building domestic AI applications. But the cost benefit always comes with questions about technology origin, dependency and long-term control.
Japan and South Korea create a different type of demand. These are economies with deep industrial bases, aging populations, shortages of technology talent and complex enterprise systems. In these markets, AI is tied directly to productivity, automation, legacy modernization and the ability to maintain competitiveness in manufacturing. Deployment demand therefore does not only go to model providers, but also to partners that can bring AI into existing workflows, data and systems.
Vietnam Enters AX From Two Directions
Vietnam is entering the regional AI map through two parallel roles: an adoption market and an engineering delivery hub. At the policy level, Resolution 57-NQ/TW, issued at the end of 2024, made science, technology, innovation and national digital transformation a top strategic breakthrough, with the goal of placing Vietnam in the top 3 in ASEAN and the top 50 globally for AI readiness by 2030. The AI Law passed in late 2025, together with more open policies for data infrastructure, sends a signal that AI has become part of the country’s development strategy.
The domestic market remains small but is growing quickly. Estimates in the report put Vietnam’s AI market at around $0.75 billion in 2024 and project that it could reach $2.81 billion by 2033, with long-term growth of nearly 15% per year. Some more optimistic scenarios project growth of more than 28% through 2030. Google estimates that AI could contribute about 12% of Vietnam’s GDP by 2030, while NIC, BCG and JICA have outlined a scenario in which AI could add $120–130 billion to the economy by 2040.
Banking is where adoption is most visible. ACB has deployed AI and RPA across nearly 400 processes, automating about 60 million task operations each year; Techcombank has invested heavily in digital transformation, built an AI agent platform for the software development lifecycle and used AI to personalize offers for more than 13 million retail customers. VietinBank, TPBank and several other commercial banks have also brought chatbots, RPA, fraud detection and credit approval into operation.
Manufacturing, healthcare, agriculture and the public sector are moving more slowly and unevenly. VinBrain, DrAid, NIC programs and smart-city experiments show that application capability is expanding, but many domestic manufacturers remain at the stage of IoT, basic automation or partial digitization. The gap between banking and manufacturing clearly reflects that AI moves faster where data is cleaner, processes are more standardized and business metrics are clearer.
The Japan and South Korea Window
As a deployment hub, Vietnam has a stronger foundation than many other Southeast Asian markets. Vietnam has about 560,000 IT professionals, adds 55,000–60,000 IT graduates each year, and aims to train 5,000 AI graduates annually by 2028 and 100,000 ICT professionals annually by 2030. Its ecosystem includes 765 AI/ML startups according to PitchBook, ranking second in Southeast Asia after Singapore. Investments such as FPT’s $200 million AI Factory, the GreenNode GPU hub and CMC’s $1 billion data center deal with Samsung C&T also add infrastructure for AI services with higher data and compute requirements.
Vietnam’s short-term advantage lies in its cost-to-quality ratio. Engineer rates of about $20–50 per hour are far below those in the United States and 15–30% lower than some Indian benchmarks for comparable capability, making Vietnam suitable for mid-sized AI delivery contracts. Still, the scale gap with India remains very large, especially for contracts that require hundreds of engineers, many senior AI specialists and enterprise consulting capability across global markets.
That is why Vietnam’s nearest opportunity lies in Japan and South Korea. These two markets lack AI talent, have strong demand for legacy modernization, are familiar with outsourcing to Vietnam and place high value on trust, time-zone alignment, communication and understanding of work culture. FPT has positioned its AI Factory toward Vietnam and Japan; mid-sized Vietnamese technology companies can enter more specific niches such as assessing legacy systems before modernization, building back-office agents, deploying RAG for enterprise knowledge, automating parts of old software interfaces, and testing and operating AI.
Vietnam’s practical strategy should move from its existing outsourcing advantage toward the role of AX delivery. Instead of only receiving software development requirements by engineering hours, Vietnamese providers can join earlier in the process: assessing workflows, standardizing data, designing agents, operating models and measuring post-deployment impact. The higher-value portion will belong to companies that can speak the language of business problems, understand customers’ legacy systems and prove impact after AI enters operations.
AX Will Reclassify Enterprise Capability
AI Transformation is creating a new line of separation inside enterprises. One group stops at demos, pilots and scattered use cases; another begins to build controlled data, choose models based on cost and risk, design agents with clear permissions, measure impact through P&L and assign responsibility to process owners. As model costs continue to fall, access to technology will become less scarce, while the ability to bring AI into the organization will become the harder part to copy.
