An ERP That Finally Lets AI Get to Work: An AI Foundation Playbook From a Longtime Customer's Story

Your employees may already be using AI, even if your company's systems aren't. In an online session for Digiwin's longtime Vietnam customers, a Digiwin VP laid out a clear path: how four tech eras (ERP, Internet, Cloud, AI) each changed something different, the "low disruption, high value" starting principle, the "four highs" method for finding a first use case, and five real Agent Space scenarios. A practical AI-foundation guide for any company that has run ERP for years. #AiGP #ERPUpgrade #AgentSpace #EnterpriseAI #SoutheastAsianManufacturers #DigitalTransformation

Your employees may already be using AI, even if your company's systems aren't. In an online session for Digiwin's longtime Vietnam customers, a Digiwin VP laid out a clear path: how four tech eras (ERP, Internet, Cloud, AI) each changed something different, the "low disruption, high value" starting principle, the "four highs" method for finding a first use case, and five real Agent Space scenarios. A practical AI-foundation guide for any company that has run ERP for years. #AiGP #ERPUpgrade #AgentSpace #EnterpriseAI #SoutheastAsianManufacturers #DigitalTransformation


Your Employees Are Already Using AI. Is Your Company?

Here is a simple question: if asked right now, could you say how many of your employees are actually using AI at work?


In an online session for its longtime customers in Vietnam, Digiwin ran a quick poll on exactly that question. More than a third of the companies present could only answer "1 to 5 people," and a few simply chose "not sure." It sounds like a harmless survey question, but the session's keynote speaker, a Digiwin vice president, turned it around: not being able to say how many people in your own company are using AI, or where, means AI is still just an individual habit at that company, not an organizational capability.


That is where most businesses stand today. Employees have already been teaching themselves to use ChatGPT and various AI tools, and they can feel the efficiency gains. But the company's own systems, especially the ERP running every day, are still standing still, never truly connected to AI. Closing that gap was the whole point of the session. In his opening remarks, a company vice president set the tone for everything that followed with one line: AI is not built on thin air. It has to sit on top of rigorous business logic and real operating data.


Who Is Behind This Framework: A Longtime AI Field Operator in Southeast Asia

None of this was reasoned out in the abstract. It comes from years of hands-on observation by the session's keynote speaker, who has spent years advising Taiwanese-owned businesses across Southeast Asia. He currently leads long-time-customer service for Digiwin's operations, and in his own words at the session: "I currently lead our long-time-customer service division here in Vietnam."


Spending that much time inside customers' production lines, offices, and back-end systems means what he sees is not a pile of abstract AI concepts, but the problems that Taiwanese-owned factories actually run into every day: can't hire, can't retain, decisions that depend on a veteran's gut feel, systems that don't talk to each other. That is exactly why the frameworks he has distilled, the "four highs," the "two mismatches," "low disruption, high value," don't sound like laboratory theory. They read more like patterns filtered out of hundreds of real company cases, which is what the rest of this article unpacks.


From ERP to AI: Four Decades, Each Change Going Deeper Than the Last

To understand why this wave of AI is different, it helps to lay the last several decades of technology change out on a timeline.


Twenty years ago, ERP arrived and changed enterprise process, pushing management toward standardization and efficiency. Ten years ago, the Internet became ubiquitous and changed enterprise information, breaking data flow free of geography. Five years ago, cloud computing took off and changed enterprise deployment, making storage and services elastic and available on demand. All three waves changed how businesses got things done.


Today's AI is different. It does not change process, information, or deployment. It changes something more fundamental: who, or what, actually runs the business. The entity that thinks and decides is no longer only human.


That distinction matters, because it explains why AI feels so hard to adopt. In each of the earlier three waves, companies simply plugged a new tool into an existing process and kept running. This time, companies have to face a more fundamental question: who makes the decision, and who is accountable for the outcome?


📊 Slide reference | Timeline of four eras of technology change: ERP changed process, the Internet changed information, the Cloud changed deployment, AI changes who runs the business



Unpredictable Outside Pressure, Unsolved Problems Inside: Why Businesses Are Struggling

If only the technology were changing, businesses could probably adapt at their own pace. The problem is that internal and external pressures are converging at the same time.


The speaker pointed to something that resonates with many Taiwanese-owned factories: revenue growth no longer requires a proportional increase in headcount, yet hiring and retention pressure has been the most common complaint heard from customers over the past year or two. Externally, there is currency volatility, tariffs, carbon taxes, and sharp market swings. Internally, there is a labor shortage, a widening talent gap, rising turnover, and escalating cybersecurity threats. Squeezed from both sides, the result is slower decision-making, knowledge that is hard to pass on, and rising operational load across the board.


Under this pressure, the way businesses traditionally operate starts to show five signs of fatigue: decisions that depend on manual experience and process; information that is always behind reality, driven by long planning cycles and manual reporting; a poor response chain caused by siloed, vertical structures; systems that each go their own way and never create synergy, with IT reduced to a toolset and data trapped in silos; and, worst of all, an organization that can only react and clean up after the fact, never getting ahead of what's coming.


The warning here is direct: without changing the underlying logic and technology foundation, a company's operating resilience will only keep getting more fragile.


📊 Slide reference | Five signs of fatigue in today's businesses: experience-driven decisions, lagging information, weak response chains, systems with no synergy, and no way to act ahead of problems



What AI Can Add: From "Iron Man" to "Transformers"

Facing these five weaknesses, what can AI actually do? The session distilled four kinds of breakthroughs: making up for a shortage of people, by letting AI handle multiple tasks in parallel for parallel productivity; making up for a shortage of capability, by packaging expertise so it can be called on and passed down endlessly for replicable productivity; finding new answers to problems that used to have none, for simulation-driven productivity; and creating value in places that were previously out of reach, for sustained productivity.


📊 Slide reference | Four breakthroughs AI brings: parallel productivity, replicable productivity, inferential productivity, and sustained productivity


But there is a point that is easy to misread here: using more AI tools does not automatically mean a company has gotten stronger. A comparison to two well-known movie archetypes makes the distinction clear.


Treating AI as an "Iron Man"-style add-on can genuinely amplify one person's efficiency and ability, but that ability still stays tied to that one person, bounded by their own capacity, way of thinking, and time. It is hard for that kind of capability to flow or accumulate across an organization. A "Transformers"-style approach to development is completely different: people only need to define the goal, and AI carries out the execution. People shift from being the executor to being the knowledge contributor and key decision-maker, and their capability can be extracted, packaged, and handed to others across the organization.


In other words, the real competitive gap between companies was never about how much AI they use. It is about whether the way work actually gets done has been genuinely restructured.


📊 Slide reference | Comparing an "Iron Man"-style personal add-on with a "Transformers"-style organizational rebuild



From Employees Using AI to a Company Becoming AI-Native

Following that logic, when AI agents move into a company, they typically go through three stages. In the first stage, every employee is issued an AI assistant to handle repetitive work. In the second stage, AI agents start joining teams like digital coworkers, able to complete tasks across departments. In the third stage, employees only need to set the direction, while actual operations are handed to digital counterparts, with these "digital workers" filling the labor gap.


The speaker was candid about what he sees today: most companies in Taiwan are still at the early stage where employees explore AI tools on their own, AI lit up here and there like scattered fires, without yet becoming an organization-wide capability. The companies actually moving ahead are the ones that have started building a data foundation, governance mechanisms, access controls, and an operating environment. Digiwin currently has more than 400 longtime customers who have reached this stage, gradually moving toward having digital counterparts embedded in organizational collaboration, with the ultimate goal of becoming an AI-native enterprise built on data-driven self-determination and generative intelligence.


📊 Slide reference | Four stages toward an AI-native company: employees using AI tools -> AI built into core infrastructure -> digital counterparts joining organizational collaboration -> a fully AI-native company


That road looks long, but broken down step by step, each stage is traceable, and that is exactly what the rest of this article unpacks: how to actually get there.


The Real Blocker Isn't Technology. It's Two Kinds of Mismatch

When companies get stuck adopting AI, the first instinct is to assume their technology just isn't advanced enough. But what actually holds companies back, according to the session, is usually one of two mismatches.


The first is an operating-environment mismatch: data scattered across systems with inconsistent standards, so AI cannot read the full picture; frequent on-the-ground changes and exceptions that make models hard to run stably; aging systems and limited equipment connectivity; and security and compliance pressure that slows AI down.


The second is an operating-relationship mismatch: departmental silos and vertical management that break up process flow; a talent and technology gap; decision-making habits that rely on personal experience, making AI's recommendations hard to trust and adopt; and a lack of internal governance and accountability mechanisms.


These two mismatches map neatly onto two corresponding solutions.


📊 Slide reference | The two mismatches that stall adoption: operating-environment mismatch vs. operating-relationship mismatch



Solution One: Replace the Foundation First, So AI Can Actually Read Your Business

To address the operating-environment mismatch, Digiwin's approach is not to rush AI applications on top of an old system. Instead, it first replaces the underlying foundation of existing systems like ERP and MES for longtime customers.


The foundation upgrade has four layers. The data layer lets AI understand the data structures and process logic inside the ERP. The application layer breaks apart and configures AI's application flexibility, so it can connect to ChatGPT, Gemini, or other large models as needed, without being locked into a single framework. The integration layer establishes the API channel through which enterprise systems and AI communicate. The compute layer brings in a configurable, governed AI model, where data is organized and filtered on the ERP side first and de-identified before it is sent to the cloud for processing, rather than handing an entire report over to AI untouched. This directly addresses the concern many companies have about sensitive data leaking out.


📊 Slide reference | The four-layer AiGP foundation: data layer, application layer, integration layer, compute layer


If you want to see how this "foundation upgrade" connects with the "four highs" method for finding use cases, a more complete breakdown is available in Southeast Asia's Veteran Taiwanese Businesses' AI Upgrade Path: How to Let AI Get to Work After You Already Have an ERP.


Solution Two: Agent Space, Where People and AI Divide Labor and Collaborate

For the operating-relationship mismatch, the answer is an enterprise digital operating environment called "Agent Space," made up of two parts: a personal space and an enterprise space.


The personal space is centered on each individual worker, understanding their operating habits and accumulating their decision history, while keeping their privacy boundaries intact. The enterprise space is organized around dedicated roles within the organization, supporting centralized management and providing an environment for agents to collaborate, with governance and traceability built in. The two spaces flow into each other naturally through a unified control mechanism, letting capability build up gradually at the individual level while also settling into a real organizational asset, one that does not disappear just because a particular employee leaves.


📊 Slide reference | Agent Space architecture: personal space and enterprise space connected through a shared governance mechanism



Five Real Scenarios: How AI Counterparts Work Together

With the theory covered, the most vivid part of the session was five live scenario demonstrations, showing what Agent Space actually looks like running inside a company.


The first scenario is order-taking decisions. When a salesperson receives a customer's price inquiry, confirming whether to accept the order and whether it will be profitable used to require cross-department back-and-forth and multiple phone calls. Now, once AI detects the inquiry email, it automatically parses it and activates finance, production-control, and sales counterparts to collaborate: finance calculates margin and cost, production control assesses capacity and lead time, and sales pulls it all together, completing an order recommendation and profit analysis within seconds (actual results vary by company size and industry and are for reference only).


📊 Slide reference | Scenario 1: order-taking decisions, with finance, production-control, and sales counterparts collaborating on profit analysis


The second scenario is procurement decisions. Using a key raw material as the example: a procurement counterpart continuously scans the market, detects that the price is near a recent low, and simultaneously spots potential risk from shipping congestion and a possible supplier break in the chain. It automatically compiles this intelligence, generates an early-stocking recommendation, and calls on production-control, materials, and finance counterparts to hold a virtual production-and-sales meeting, each producing demand forecasts, inventory-gap analysis, and cost trade-off reports, leaving only a one-click sign-off for the manager.


📊 Slide reference | Scenario 2: procurement decisions, with automated market intelligence and a virtual sales-and-production meeting


The third scenario is rush-order rescue. When a customer suddenly adds an order, whether the production line can keep up traditionally turns into chaos on the floor and finger-pointing across departments. In Agent Space, a governance-and-compliance counterpart first checks trade rules, customer credit, and compliance terms to confirm the deal is sound and legal, then automatically rallies sales, warehouse, and procurement counterparts to work in parallel, completing inventory checks and lead-time decisions, and generating the formal order with a single click while syncing it into the ERP (a process that used to take multiple rounds of phone confirmation can, based on experience, be shortened substantially; actual results are for reference only).


📊 Slide reference | Scenario 3: rush-order rescue, with a governance-and-compliance counterpart checking the deal before cross-department counterparts generate the order


The fourth scenario is an executive think tank. Facing a spike in raw material prices or a tariff shock, decision-makers can summon virtual experts in finance, procurement, and supply chain to run multiple rounds of in-depth debate: reviewing safety stock, calculating capital risk, and simulating how different strategies would affect margin, ultimately producing a well-organized risk assessment report. The system even scores each counterpart's performance, adding a layer of traceable, professional backing to the decision.


📊 Slide reference | Scenario 4: executive think tank, with virtual experts debating across multiple rounds to produce a risk-assessment report


The fifth and most far-reaching scenario is knowledge transfer. Decision-making experience from handling a tricky situation in the past can, with a single command, be automatically packaged, capturing tacit knowledge and the logic behind the response as a standardized, dedicated skill. The next time a similar situation comes up, a digital counterpart can call on it directly, sharply speeding up thinking and response time, and turning every past success into a digital asset that belongs to the company rather than disappearing the moment a veteran employee leaves.


📊 Slide reference | Scenario 5: knowledge transfer, packaging tacit experience into standardized skills that counterparts can call on


Behind this scenario sits a more fundamental issue: a company's inability to retain a veteran's knowledge and its inability to retain the veteran are really two sides of the same coin. For more on the hiring and retention pressures facing Southeast Asia's Taiwanese-owned factories, and how to respond to them, see Competing With Japanese and Korean Factories for the Same Workers: Hiring, Retention, and Operational Health at Taiwanese-Owned Factories in Southeast Asia.


Not Sure Where to Start? Find Your "Four Highs" First

After seeing all five scenarios, many people react the same way: the logic makes sense, but where should my company actually start?


The method here is to find your "four highs": high load, meaning tedious, repetitive paperwork that can't be skipped; high frequency, meaning reconciliation and dispatch reporting that happens every day and can't afford errors or delays; high knowledge, meaning scheduling compliance or after-sales Q&A where the data and process exist but someone still has to be asked every day, because the knowledge is hard to pass on; and high experience, meaning design changes that depend entirely on a veteran's gut judgment and are hard to hand off. From order intake, procurement, and R&D design, to production scheduling, customer service, HR, and financial audit, almost every function has its own "four highs" hotspot, and that hotspot is the first use case worth prioritizing.


📊 Slide reference | The "Four Highs" method: high load, high frequency, high knowledge, high experience


The starting principle can be summed up in four words: low disruption, high value. Don't change a company's existing processes, to keep organizational resistance and switching costs down. Focus on high-frequency, high-impact scenarios so employees and the organization both feel the results quickly. Start from natural-language interaction and Q&A-style operation, so AI blends naturally into daily workflow. And don't rebuild the system architecture; adopt through layering and integration instead, to keep technical risk low.


📊 Slide reference | Four starting principles for low disruption, high value


The recommended rollout path has three stages: start with operations and accounting, upgrading the ERP's AI foundation so AI can actually read the existing system's data; then connect operating, plant, security, and sustainability data to form digital counterparts that support cross-functional collaboration; and finally restructure data semantics, permissions, and process context to move toward a fully digital-native state. No stage has to be perfected before the next one begins; each can expand while the last is still being built out.


Years of Running ERP Is an Advantage, Not Baggage

Many companies worry that a system they have run for over a decade will actually slow down AI adoption. The session's answer was the opposite: companies that have run ERP for years actually have an advantage over those starting from zero, for three reasons.


First, the learning curve for employees is extremely low, because the new way of working isn't far from the AI tools they already use; employees simply move from being an "operator" to being AI's dispatcher. Second, operating data already lives in their own systems and tends to be highly structured, which is exactly the real, grounded brain that AI needs most. Third, the existing ERP process framework means employees don't have to explore and make mistakes without any guardrails; AI can be folded directly into workflows they already know well.


By contrast, a company that chooses to build AI from scratch on its own has to relearn how its people work with the new system, take on the leak risk of public models and external APIs as its data source, absorb the cost of training and trial-and-error from a standing start, and solve the accuracy and structuring of its own data by itself. This is exactly why an original-vendor upgrade path is different: the AI's source is reliable and does not send sensitive data to the cloud untouched; technical integration is handled by the original vendor, which can also allocate usage quotas for employees; and most importantly, it does not change the company's existing processes and data structures.


📊 Slide reference | Comparing three pillars of a self-led rollout vs. an OEM/vendor-led upgrade path


Viewed as part of a full annual plan, this kind of investment looks more like a piece of infrastructure than a one-off purchase. For how to fit a "foundation upgrade" into your budgeting logic, see Should Taiwanese Businesses Put AI in Next Year's Budget? Three Budgeting Approaches and Four Hidden Costs.


A Few Frequently Asked Questions

In the Q&A session before the event wrapped up, a few questions came up repeatedly, the exact answers most companies that have run ERP for years and are considering AI adoption want to know.


The first question: what do we need to prepare before upgrading to an AI foundation, and will it disrupt how our current system runs? The answer is that the first step is usually confirming whether the existing host system meets the upgrade requirements, which may need some system-level updates and case-by-case adjustments. The operating workflow stays largely similar to what employees already use, so the impact on staff is actually minimal. The real key factor is whether the company itself has its environment in order and is mentally prepared.


The second question: there are so many "four highs" scenarios, which one should come first? Most companies are first concerned with the "order" scenario, because it touches on what leadership cares about most: whether an order can be accepted and whether it will be profitable. It is usually also the easiest place to see results quickly and the one employees feel the most.


The third question: why is upgrading the AI foundation often described as doing a cybersecurity upgrade at the same time? The reason is that the new foundation already constrains how AI communicates and collaborates, so it does not repeat the risk of an employee habitually uploading an entire dataset to the cloud untouched. At the same time, a foundation upgrade usually comes with a host system refresh, retiring older, unsupported operating systems that carry known vulnerabilities, which also happens to satisfy the cybersecurity requirements common to publicly listed companies. Doing all three together means process, data, and security all move forward at once.


The speaker closed with a single line that captures the whole session: one person getting stronger is a gain in efficiency; a whole group getting stronger together is what genuine competitive capability looks like for a company. That may also be the question every company still hesitating over where to start should ask itself first: do you want AI to stay a personal habit for a handful of employees, or do you want it to become a capability the whole organization owns together?


If your company has also been running ERP for years and wants to know the specific path and investment required to upgrade to an AI foundation, feel free to submit a request form, and a consultant familiar with your industry and country will assess an upgrade path suited to you.


FAQ


Q1: Does this approach work for companies of any size?

A1: The "four highs" method and the "low disruption, high value" principle described in this article are meant to help identify the first suitable AI use case inside a company, and are not limited to any particular company size. The actual upgrade conditions and investment required still need to be assessed individually based on a company's existing systems and data.


Q2: Do the order-taking, procurement, and rush-order scenarios described here need to be purchased separately?

A2: These scenarios are applications built on an upgraded ERP foundation, connected through Agent Space to existing system data. They are not standalone products. The actual scope and upgrade conditions should be confirmed with a consultant.


Q3: In practice, how do you tell whether you've actually achieved "low disruption, high value"?

A3: Two checkpoints: first, whether existing processes, documents, and report structures stay unchanged after adopting AI; second, whether the first use case happens every day and employees notice the impact within a few weeks. If both hold true, that is a low-disruption, high-value starting point.

知識主題
管理與擴張期講座
成長與量產期講座

We use our own and third-party cookies for analytics and to show you ads based on your browsing habits and profile. For more information, see our Privacy PolicyPrivacy Policy.