A production line grinds to a halt. A machine fails without warning. The instinct and hard-won judgment of a veteran maintenance technician walk out the door the day they retire. Across Japanese-owned factories in Thailand and the wider ASEAN region, these “invisible costs” quietly keep eating into profit, year after year. In 2026, one technology has emerged as a genuinely practical answer to this structural problem: the AI agent.
Unlike the chatbots and image-recognition tools that defined the previous wave, AI agents can assess a situation on their own, operate multiple tools, and see a task through from start to finish. That kind of autonomous AI has finally begun to reach the factory floor — in production management, equipment maintenance, and back-office operations alike. This article breaks it all down from the vantage point of a factory IT integrator: the latest movement in the foundational technologies such as Anthropic’s Claude Agent SDK and OpenAI’s Codex, real deployments at leading manufacturers like Daikin and Panasonic, and a concrete roadmap that manufacturers in Thailand can use to make AI agents stick on the shop floor.
Why AI Agents, and Why Now, for Manufacturing
The Tectonic Shift from Generative AI to Agentic AI
The generative-AI boom of 2023 and 2024 revolved around a single mode of use: answering questions. Draft this text, summarize that document, generate an image. Useful, but fundamentally reactive. Entering 2026, the center of gravity in industry has moved decisively to the next phase — agentic AI.
An AI agent is software that, given a goal, breaks it down into sub-tasks on its own, calls whatever tools it needs (internal systems, databases, sensors, external APIs), checks its own results as it goes, and completes an entire sequence of work without a human directing every step. The shift is from AI that returns an answer to AI that gets the job done. For a factory, that distinction is anything but academic.
Here is why. Most work in a plant is not a single question with a single answer — it is a sequence of steps. When a machine throws a fault, a maintenance technician (1) checks the alarm, (2) cross-references past failure history, (3) confirms the mechanical structure against drawings, (4) reasons out the probable cause, (5) executes a corrective procedure, and (6) logs what happened. The ability to support that entire workflow — not just one isolated step — is the essential value of an AI agent.
Three Structural Pressures Bearing Down on Manufacturers
Japanese manufacturers operating in Thailand and ASEAN have no real choice but to pay attention to agentic AI, and three structural pressures explain why.
The first is a shortage of skilled people and a looming crisis in knowledge transfer. Thai plants must keep developing maintenance and production-engineering capability as equipment grows more complex, while standardizing work across languages. Capture why experienced people make a decision and turn that reasoning into validated organizational knowledge.
The second is rising labor costs colliding with ever-higher quality demands. Thai production bases that once competed on cheap labor are now squeezed between climbing minimum wages and tightening quality requirements from Japanese headquarters and global customers. They are being pushed to reduce headcount and stabilize quality at the same time — two goals that used to pull in opposite directions.
The third is the reality of having data but not using it. As IoT and production-management systems have spread, factories now sit on mountains of data. Yet only a small fraction of companies have connected that data to actual decisions and shop-floor improvement. An AI agent can act as the translation layer that turns this “sleeping data” into concrete action on the line.
What Actually Separates “Automation” from “Agents”
It is worth drawing a clear line between conventional automation — RPA and fixed-rule programs — and AI agents. Traditional automation is excellent at one thing: repeating a defined procedure exactly as defined. The moment reality deviates even slightly from that script, however — an unexpected alarm, a form with a broken layout, a failure pattern never seen before — a human has to step in. In practice, a great deal of shop-floor labor is spent precisely on this exception handling.
An AI agent, by contrast, can interpret a situation, adjust its judgment to context, and choose the best action from several options. “Which past failure does this alarm most closely resemble?” “Which line item in the specification does this entry on the form correspond to?” These are exactly the questions that used to depend on human judgment — and stepping into that territory is the decisive difference. That is why agentic AI delivers value well beyond streamlining routine tasks: it standardizes and accelerates work that requires judgment, which is a far larger prize.
Put differently, conventional automation scales the parts of a job that were already predictable, while an AI agent begins to scale the parts that were previously bottlenecked on a scarce human expert. On a shop floor where that expert may be one or two people nearing retirement, the economic difference is not incremental — it is the difference between a capability that lives in a person and a capability that lives in the organization. This is also why “agent readiness” is becoming a question of process design as much as technology selection: the clearer and better-documented a workflow already is, the faster an agent can absorb it.
The State of Agentic Platforms in 2026
Before thinking about deployment on the floor, you have to understand how mature the underlying platforms have actually become. In the first half of 2026, this layer advanced at a remarkable pace.
Anthropic Claude Agent SDK — Tooling Built to Finish the Job
Anthropic’s Claude Agent SDK was renamed from the “Claude Code SDK” in September 2025 and has developed into a foundation for production AI agents. It supports Python and TypeScript and provides tools such as file editing and command execution, along with permission controls and hooks, allowing implementers to design human approval checkpoints (human-in-the-loop) appropriate to the workflow’s risk.
Two capabilities matter especially for manufacturing: subagents and MCP (Model Context Protocol). Subagents let a large task be delegated to child agents that each carry their own context, which makes it possible to design a system where specialized jobs — production planning, equipment diagnosis, form processing — are handled by dedicated workers. MCP is a common standard for connecting an AI safely to external systems such as an in-house production-management platform or a sensor-data backbone. It is the bridge between the factory systems a company already runs and the AI agents it wants to add. For any plant with existing infrastructure, that bridge is the difference between a demo and a deployment.
Claude Cowork — From Desktop to Cloud, and Toward the Floor
In an official May 2026 announcement, Anthropic said Claude could work directly in Excel, PowerPoint, Word, and Outlook through Microsoft 365 add-ins. Dispatch in Claude Cowork lets users assign tasks by text or voice and allows Claude to continue working on permitted local files while they are away, with results ready for review on return. Consequential factory actions still require explicit approvals and separated permissions.
In a manufacturing context, this points to something specific: automation that spans the gap between “administrative work at the head office” and “operations on the factory floor.” Picture a workflow that starts with an equipment alert on the line, where an AI agent then drafts a maintenance work order, notifies the relevant people, and checks parts inventory — a genuinely cross-departmental sequence that this class of tooling now brings within reach.
OpenAI Codex — Rising to Enterprise Standard
OpenAI has been consolidating ground just as quickly. Codex was named a Leader in the 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents. In OpenAI’s June 2026 announcement, more than five million people used Codex each week, with non-engineers accounting for about 20%. Its ability to understand codebases, make changes, run tests, and prepare work for human review is increasingly relevant beyond engineering.
In June 2026, OpenAI added “Sites,” which lets teams build business applications without going through front-end development, along with “Annotations” for in-place editing — accelerating adoption into non-engineering departments such as finance, legal, and recruiting. For a manufacturer’s IT department, this means the environment for extending internal systems and building shop-floor tools in-house is coming together at a speed never seen before.
These advances expand the options for embedding AI agents in real workflows. Production use on a factory floor still requires permission controls, audit logs, human approval, fallback procedures, and adequate data quality, beginning with a bounded pilot and expanding only after the evidence supports it.
What Is Actually Happening on the Factory Floor: Leading Deployments
Theory alone does not move a shop floor. So let us look at real cases where Japanese manufacturers have begun to produce concrete results.
Daikin x Hitachi: An Equipment Fault-Diagnosis Agent Behind the “Factory That Never Stops”
Daikin Industries, the comprehensive air-conditioning manufacturer, partnered with Hitachi to co-create an AI agent that supports equipment fault diagnosis in the factory, and began a trial deployment in April 2025 aimed at putting it into practical use. This effort is a landmark illustration of how an AI agent can function on a real manufacturing floor.
The mechanism is instructive. Drawings for the production equipment and utility equipment that Daikin has accumulated over the years were converted into a “knowledge graph” that a generative AI can read. To this were added maintenance records and other “OT data,” along with Hitachi’s proprietary fault root-cause analysis process — its “OT skill,” grounded in methods such as STAMP. By training on this combination, the system achieved diagnoses on par with, or better than, those of a typical maintenance technician.
In a preliminary demonstration, the equipment fault-diagnosis AI agent identified causes and recommended corrective actions within 10 seconds, at over 90% accuracy. In April 2025, Daikin and Hitachi said they planned to complete the trial by September 2025, then pursue practical use and deployment to production sites in Japan and overseas. That was a forward plan, not evidence here that every later rollout was completed.
The lesson this case holds for manufacturers in Thailand is plain. The value of an AI agent lies exactly here: in converting assets that used to sit dormant — in a warehouse or inside one person’s head, such as drawings and maintenance records — into a form the shop floor can actually use. The problem was never a lack of data; it was the lack of a mechanism to put that data to work. This is a direct answer to a challenge that a great many factories share.
Panasonic Connect: A 97% Cut in Drawing Verification Signals a Back-Office Revolution
A second case worth close attention is the AI agent Panasonic Connect introduced in February 2026, purpose-built for matching drawings against design specifications. In this effort, the company reported cutting the verification time by up to 97%.
This case demonstrates that the impact of AI agents reaches beyond “floor operations” like equipment maintenance and into “indirect work” such as design, procurement, and quality control. Drawing verification, specification checking, voucher processing, inspection-record reconciliation — manufacturing is full of manual, visual cross-checking work, much of which is prime territory for dramatic gains from the combination of AI agents and AI-OCR technology.
The Success Pattern You Can Extract from These Cases
Three things are common to these leading examples. First, they take aim at a specific business process. Rather than “do everything with AI,” they start from a clear use case — fault diagnosis, drawing verification. Second, they run on the data assets already sleeping inside the company — drawings, maintenance records, specifications — as their fuel. Third, they do not replace people; they accelerate and standardize human judgment.
The Reality — and the Opportunity — for Manufacturers in Thailand and ASEAN
The leading cases are compelling, but what does the ground truth look like for manufacturers operating in Thailand right now? The data reveals a large gap and, at the same time, a large opportunity.
The “Eager to Adopt, Slow to Implement” Gap
Thailand’s Board of Investment (BOI) promotes investment in automation, robotics, and digital technology to improve manufacturing efficiency and competitiveness. Frameworks such as the Thailand i4.0 Index assess equipment, data, and organizational capability by maturity stage, so the useful question is where each process in a specific factory stands today.
Recurring issues include missing or inconsistent data, legacy-equipment connectivity, departmental permissions, cybersecurity, and operating capability. Assess data quality, connectivity, blast radius, approvers, and recovery procedures for each target process.
The Gap Itself Is the Source of Competitive Advantage
The need for staged improvement creates room for companies that start learning early with a bounded process. Check each investment against current BOI requirements and begin with a small use case that can safely fall back to manual operation.
The essential move is not to “wait until a perfect data foundation is in place” but to “start small with the data you have today.” As the Daikin case showed, existing assets like drawings and maintenance records are exactly what fuel an AI agent. Data-quality issues are something to resolve incrementally, in parallel with adoption — not a reason to delay getting started.
There is another tailwind unique to Japanese manufacturers in Thailand: the know-how their headquarters have accumulated ahead of them can be transplanted to the Thai site. Just as Daikin explicitly stated it would “roll out to production sites at home and abroad,” for a Japanese maker with global operations, turning the tacit knowledge gained through AI agents into explicit, documented knowledge becomes an asset that directly raises the speed and quality of standing up overseas bases. The Thai site need not be a follower of headquarters. Precisely because the labor-shortage challenge is so acute there, it can instead become a leading site for applying AI agents.
TOMAS TECH’s Practical Know-How: Connecting the Floor with AI Agents
TOMAS TECH is a factory IT integrator that has delivered PEGASUS production and energy management systems, IoT equipment data collection, AI-OCR, and FA and robot control solutions to Japanese manufacturers in Thailand. What that hands-on experience reveals is the set of practical pressure points that keep an AI agent from ending up as a nice idea that never runs.
Why an “AI Agent on Its Own” Fails to Move the Floor
Where most companies stumble in AI adoption is not the performance of the AI model itself — it is the absence of the “plumbing” that connects the AI to the systems on the floor. However clever an AI agent may be, it cannot build a production plan if it cannot access production-results data, and it cannot catch the early signs of a failure if it cannot read the values coming off the equipment sensors.
The strength TOMAS TECH brings is precisely this plumbing. Production-results, process, and inventory data held in the PEGASUS production-management system; equipment operating, power-consumption, and vibration data collected via IoT; information from paper forms digitized by AI-OCR — the experience of organizing all of this into a single, coherent whole is a decisive foundation for connecting an AI agent to the floor. A connection standard like the MCP described earlier only delivers value once this bridge to existing systems is in place.
It is worth being concrete about what “plumbing” means in practice. It is the data model that reconciles how the same machine is named in the production-management system, the maintenance log, and the drawing archive. It is the access layer that lets an agent read a sensor value without opening a security hole into the control network. It is the audit trail that records what the agent looked at and why it recommended what it did, so that a human reviewer can trust — or challenge — the output. None of this is glamorous, and none of it appears in a vendor demo. But it is exactly the layer that determines whether an AI agent survives contact with a real factory or quietly gets switched off after the pilot.
Linking Four Shop-Floor Assets Through AI Agents
Concretely, the four domains TOMAS TECH has worked in become, as-is, the implementation base for AI agents.
Production management (PEGASUS): Supporting production planning, answering delivery-date questions based on actual results data, and automatically detecting abnormal progress. An AI agent continuously monitors production results and raises an early warning when reality diverges from the plan.
IoT equipment data: The foundation for equipment fault diagnosis and predictive maintenance, as in the Daikin case. An AI agent interprets the time series of sensor data and presents both the early signs of an anomaly and the recommended action.
AI-OCR x back office: The automation of drawing and form verification demonstrated by the Panasonic case. Paper inspection records, delivery notes, and specifications are digitized and fed into automated matching and reconciliation by an AI agent.
FA and robot control: The connection to floor operations. This is the final stage that bridges an AI agent’s judgment into actual equipment control and line adjustment.
Rather than optimizing each of these four domains in isolation, linking them organically through AI agents is what makes a cross-departmental, end-to-end workflow real: detect an early sign of equipment anomaly, automatically draft a maintenance work order, check parts inventory, and re-adjust the production plan. This is manufacturing DX in the age of AI agents — clearly distinct from merely installing a tool.
The Value of a Partner Rooted in the Local Reality
One thing that is easy to overlook is the handling of language and shop-floor culture. A Thai factory is a mix of Japanese managers, Thai staff, and a multinational workforce. If the instructions and records an AI agent generates do not reach every one of those people correctly, its value is cut in half. Operating designs that span Japanese, Thai, and English, and implementations grounded in an understanding of the business customs and equipment realities of the Thai shop floor — these are gaps that installing a generic global tool, as-is, will never close. It is precisely because an integrator has walked the floor with Japanese manufacturers in Thailand that it can close that gap and turn an AI agent into a mechanism that actually gets used.
An Implementation Roadmap: From a Small Start to Autonomy
So how should a manufacturer in Thailand proceed, concretely? Drawing on TOMAS TECH’s field experience, here is a four-stage roadmap for building up results without overreaching.
Step 1: Inventory Your Data Assets and Build the Connection Base
First, take stock of what data assets your company actually holds. Drawings, maintenance records, production results, inspection records, paper forms — as the Daikin and Panasonic cases show, these are the fuel for an AI agent. At the same time, get your production-management system and IoT backbone into a state where the AI can access them safely. There is no need to aim for perfection. Start with “the data needed for a single business process.”
Step 2: Pilot with a Single Use Case
Next, select a single use case with high impact and high feasibility, and run a pilot. The recommended targets are areas with fast payoff — supporting equipment fault diagnosis, form verification with AI-OCR, anomaly detection in production progress. At this stage, the iron rule is to always build in a human approval checkpoint (human-in-the-loop) and to begin with an operating model where a person confirms and corrects the AI’s suggestions.
When defining a pilot budget, timeline, and acceptance criteria, see our guide to AI PoC costs and success criteria for manufacturers in Thailand. For governance and role-sharing between headquarters and overseas sites, read implementing generative AI in an overseas subsidiary.
Step 3: Link Workflows and Scale Across
Once the pilot has confirmed reliability, link multiple business processes together. Following a chain like “anomaly detection to maintenance instruction to inventory check,” connect the AI agents you introduced as isolated points into a line, growing them into a cross-departmental workflow. In parallel, scale the know-how from a successful site out to other sites and other lines.
Step 4: Gradual Autonomy and Knowledge Transfer
In the final stage, raise autonomy incrementally where trust and evidence have accumulated. Document how experienced people handle exceptions and maintain it as validated knowledge the agent can consult. Position the AI agent as a collaborator that raises shop-floor capability, not as a presumed replacement for people.
Common Failures in AI Agent Adoption — and How to Avoid Them
Behind the glamour of the leading cases, there is no shortage of AI-agent projects that do not go well. Surveys, too, point out that the main causes of failed adoption concentrate not in technical limits but in problems of governance, data preparation, and shop-floor engagement. Here are the typical pitfalls TOMAS TECH has seen on the floor, along with how to avoid them.
Failure Pattern 1: Chasing “AI for Everything” Without Narrowing the Goal
The most common failure is spreading the tablecloth too wide — “let’s transform the whole company with AI” — without ever narrowing down the target work. Costs pile up while results stay invisible, and internal expectations deflate. The remedy is clear: concentrate first on one specific, measurable use case. It was no accident that Daikin narrowed to “equipment fault diagnosis” and Panasonic to “drawing verification.”
Failure Pattern 2: Driving It from IT Alone, Without Involving the Floor
If the operators and maintenance technicians on the floor cannot trust what an AI agent proposes, it will not be used no matter how accurate it is. The floor’s sense of “why did it reach that judgment?” is the lifeline of adoption. The remedy is to design for use with a human approval checkpoint, so the floor can verify as it works, and to reflect the floor’s tacit knowledge into the AI’s learning — so people take ownership of it as “an AI packed with our own wisdom.”
Failure Pattern 3: Deferring the Connection to Existing Systems
Pour all your energy into selecting an AI model and validating accuracy while treating the connection design to your production-management system and equipment data as an afterthought, and the proof-of-concept may succeed only for the production rollout to stall. “The demo runs, but it cannot be used on the floor” — crossing this valley requires designing for integration with existing systems from the very earliest stage.
Frequently Asked Questions (FAQ)
Q1. Does adopting an AI agent require a large-scale data foundation?
No. There is no need to wait for a perfect data foundation to be completed. As the Daikin case shows, you can start from “the data you already have,” such as existing drawings and maintenance records. What matters more is narrowing to a single use case, starting small, and expanding while confirming results. Data-quality challenges can be improved incrementally, in parallel with adoption.
Q2. Will AI agents take away the jobs of workers on the floor?
The direction the leading cases point to is not “replacement” but “capability uplift.” Daikin’s fault-diagnosis agent accelerates and standardizes maintenance technicians’ judgment; Panasonic’s drawing verification automates the simple matching that people used to do, so those people can concentrate on higher-value judgment. It also becomes a vessel for passing a veteran’s tacit knowledge on as organizational knowledge — a countermeasure to the very crisis of skill transfer.
Q3. Can a factory in Thailand adopt this the same way Japan does?
Yes. A Thai plant can deploy the same class of technology when process data, equipment connections, language, permissions, and support are designed locally. Validate manual operation during outages, Thai-language standard work, approval accountability, cybersecurity, and integration with shop-floor systems.
Q4. Can Claude Agent SDK or OpenAI Codex be used on the factory floor as-is?
These are powerful foundational technologies, but on their own they do not connect to the production-management systems or equipment data on the floor. It is essential to bridge them to existing systems using a connection standard like MCP, and to design and implement around the floor’s actual business processes. Even more than the choice of foundational technology, the integration design with shop-floor systems is what decides success or failure.
Q5. Which use case should we tackle first?
The three areas with the best balance of fast payoff and feasibility are: (1) form and drawing verification combined with AI-OCR, (2) equipment fault-diagnosis support that leverages existing maintenance records, and (3) detection of progress anomalies based on production-management data. The standard approach is to choose the area where your company holds the richest data assets and where the floor’s pain point is clearest. When selecting, weigh not only “the size of the effect” but also “whether a failure has limited impact and lets you run a fast learning loop.” That first taste of success is the engine that draws out internal understanding and the next round of investment. Start small and learn fast — that iteration is the surest, shortest route to moving AI-agent adoption steadily forward.
Conclusion: 2026 Is Year One of AI Agents in Manufacturing
In 2026, the foundations and manufacturing examples for embedding AI agents in real work are expanding. Official announcements reported that the Daikin–Hitachi preliminary demonstration returned answers within 10 seconds at over 90% accuracy, while Panasonic Connect reduced drawing-verification work from 50–340 minutes to 10 minutes, an 80–97% reduction. These were bounded use cases; each factory still needs its own validation.
For manufacturers in Thailand and ASEAN, implementation depends on legacy-system connectivity, data quality, multilingual operations, and permission design. Inventory existing drawings, maintenance records, and production results, then start with one bounded use case. Measure value and risk with human approval and manual recovery before increasing autonomy.
And to keep an AI agent from ending up as a nice idea that never runs, the plumbing that connects the AI to shop-floor systems is indispensable. Only by organically linking the four shop-floor assets — production management, IoT equipment data, AI-OCR, and FA control — does cross-departmental, end-to-end manufacturing DX become real. TOMAS TECH stands close to the floor of Japanese manufacturers in Thailand, and as a partner who walks alongside you through this implementation, we support your work of building a factory that never stops. Let us take that firm first step into manufacturing DX in the age of AI agents — together.
If you are still deciding which process to pilot or how to connect existing data and equipment, TOMAS TECH can help from the planning stage. Visit our contact page to discuss a small, controlled starting point.
