E-Commerce System Development
and Enterprise AI Adoption Consulting
MAQ Inc. has built software since 2002, delivering e-commerce, cloud systems, multilingual websites, ERP, and CRM for businesses. Our core team includes a former visual design director at an international magazine, so UI/UX gets the same attention to detail as macOS — when we bring AI in, interface design, software development, and compute hardware planning are handled by the same team.
MAQ applies this same capability to its own operations: our inventory and sales management, purchasing, order processing, e-invoicing, bookkeeping, cost reporting, and staff scheduling systems are all built in-house and have been running in production for more than 15 years. AI has since been integrated into them, letting us query inventory, purchase history, annual revenue, and best-sellers in natural language. Every method described on this page comes from a system actually running today.
Last updated: July 27, 2026
Four Core Development Services
E-Commerce Systems
Full-stack development spanning product configurators, shopping carts, payment integration, and order management. MAQ's own online workstation configurator is a case in point — it prices and schedules hundreds of component combinations in real time and carries our day-to-day order flow.
- Shopping cart, membership, order, and shipping management
- Payment/logistics integration (credit card, installments, convenience-store pickup)
- Product configurators with real-time price calculation
- Back-office reporting and permission controls
Corporate & Brand Websites
We don't just build the layout — we build a site that search engines and AI can find and understand: semantic structure, structured data (Schema.org), and performance and mobile optimization.
- Responsive (RWD) design and performance optimization
- SEO and structured data (JSON-LD) implementation
- Content management and multilingual support
- Site migration, redesign, and long-term maintenance
ERP Systems
We offer custom ERP development, and can also build integration layers and extension modules for an existing ERP. MAQ's own inventory, purchasing, order, e-invoicing, and cost-reporting systems are self-built and in daily use, which gives us first-hand insight into what a system needs after it goes live.
- Purchasing/sales/inventory management and safety stock
- Purchasing, supplier, and incoming invoice management
- Order processing, shipping, and e-invoice issuance
- Cost analysis, revenue reporting, and quotation systems
- API integration and data interfacing with existing ERP systems
- Tiered permissions, audit trails, and SMS two-factor login
CRM Systems
We bring customer interactions scattered across LINE, email, phone calls, and spreadsheets together into a single timeline that's searchable, analyzable, and easy to hand off. It can also connect to your e-commerce and ERP systems, so sales can see orders and inventory in one place.
- Customer profiles, interaction history, and opportunity management
- LINE official account, email, and customer-service integration
- Data interoperability with e-commerce/ERP
- Sales forecasting and customer analytics reports
How Do You Bring AI Into an ERP System?
Bringing AI into an ERP doesn't require replacing the existing system — the real decision is where the computation happens. The common approach today is sending ERP data to a cloud LLM through an API. It's the fastest path to adopt, but it also means quotes, cost structures, customer lists, and payroll data leave the company network. This trade-off is rarely spelled out in most discussions.
MAQ's position is to assess data sensitivity first, then decide the architecture. General queries and summaries suit a cloud API; anything touching finances, cost structures, customer lists, or payroll should run on a model deployed on your own AI server. The two things that requires — a program that safely pulls ERP data, and a machine powerful enough to run the model — are exactly what MAQ can provide together.
Four Practical Steps
Map the data and pain points
Identify the parts your ERP already covers but still eat up staff time — duplicate order entry, inventory judgment calls made by hand, month-end reports that take too long to produce. At the same time, flag the data that must never leave the company.
Decide where it runs
Route deployment by data sensitivity: general data goes through a cloud API, confidential data stays with an on-premises model. Most companies end up with a hybrid architecture — everyday queries in the cloud, anything sensitive kept in-house.
Build the integration layer
Pull ERP data safely through a read-only API or database views, index it for the AI to query, and carry over the original system's permission tiers so each role can only access information within its own scope.
Launch small, then expand
Start with a single department and a single workflow, measure the actual time saved, and only then consider expanding. Rolling it out company-wide all at once makes it harder to validate results and to maintain afterward.
We Put AI Into Our Own ERP First
There's no shortage of talk about bringing AI into ERP, but far fewer companies actually run it on their own operational data. MAQ's internal systems — inventory, purchasing, orders, e-invoicing, cost and revenue reporting, and staff scheduling — are all self-built and running in production. With AI added, the system can now answer natural-language queries like these.
This implementation produced two concrete conclusions. First, the model should never generate SQL queries on its own. MAQ's approach is to expose 11 clearly defined, read-only query tools, with the SQL fixed in code — the model can only pick a tool name and structured parameters, and any call outside that whitelist is rejected. The design trades away some flexibility for stable, auditable results, so it never invents a part number or an amount that doesn't exist.
Second, permissions should be enforced at the entry point. MAQ currently uses the existing back-office permission levels to decide who can open the AI interface — anyone with the right permission already had access to that data anyway. Finer-grained, row-level permission control — for example, letting each salesperson see only their own customers within the same report — is a harder problem, and MAQ is still working on it. Organizations that need this should build it into the data model from day one of the project; retrofitting it later comes at considerable cost.
These conclusions all came out of an actual build. So when discussing AI adoption for ERP, what MAQ brings to the table is concrete: how to structure permissions, how to design query tools, and how to decide where data should live — not just an architecture diagram.
What Hardware Do You Need to Run a Model on Your Own AI Server?
This is the part most ERP-and-AI discussions skip. The table below maps the practical correspondence, as a reference for feasibility.
| Use Case | Recommended Model Size | Key Hardware Requirements |
|---|---|---|
| ERP data Q&A, report summaries (small number of users) | 7B–14B quantized model | A single 24GB-class GPU |
| Cross-department knowledge base Q&A, document retrieval (RAG) | 20B–32B quantized model | 32GB VRAM, 64GB+ system memory |
| Multiple concurrent users, higher reasoning quality needed | 70B, 4-bit quantized | 48GB+ VRAM, 128GB ECC system memory |
| Enterprise knowledge hub, long context and high concurrency | 120B-class quantized model | 96GB VRAM, 256GB ECC system memory |
The table above is based on measured benchmarks for internal use by 5–10 people. Each doubling of concurrent users adds roughly 50% to system memory requirements. For a detailed breakdown, see the AI Hardware Buying Guide; for an on-premises enterprise knowledge server, see MAQ Alishan.
Why MAQ
- Software and hardware, one team
- Once a software vendor finishes a system, you're usually left to source the compute hardware yourself; once a hardware supplier delivers a machine, you still need someone else to handle the integration. MAQ does both, so hardware specs are sized to the system's actual workload — not picked off a catalog tier.
- Visual design as meticulous as macOS
- Our team includes former visual design directors from international magazines, and we build interfaces that are as good-looking as they are usable — a capability most system-development teams simply don't have. A system shouldn't just work; people should want to open it every day.
- Twenty-plus years, and we still use it ourselves
- We've been building websites and e-commerce systems since 2002. The site you're browsing right now, along with its online configurator and order system, was built and is still maintained by MAQ.
- We assess whether you need it before we build it
- Not every process is a good fit for AI. MAQ helps you estimate the time and cost it would actually save first — and if the numbers don't add up, we'll tell you so.
- Your data stays on your own AI server
- A fully on-premises deployment is available on request, so your data never has to pass through a third-party API. This is a capability already built into MAQ's product line — not something assembled ad hoc for a one-off project.
Start With Whatever Takes the Most Time Right Now
Whether you need a new system built, want to bring AI into an existing ERP, or just want an initial assessment of whether it's worth doing — reach out. The assessment stage is free.
Frequently Asked Questions
- How do you bring AI into an ERP system? Do we need to replace what we already have?
- No replacement needed. In practice, you keep the existing ERP, pull the data you need safely through a read-only API or database views, index it, hand queries to an LLM, and surface the results in your existing interface. The decision that actually matters isn't whether to swap systems — it's where the computation happens: whether data leaves the company network determines everything that follows, from architecture to cost to compliance.
- What are the risks of sending ERP data to a cloud AI?
- ERP data typically covers quotes and cost structures, customer lists, supplier terms, and payroll — all trade secrets or personal data. Before sending anything to a third-party API, check its data-retention policy, whether it's used for model training, whether data crosses borders, and whether audit trails are complete. General queries are fine on a cloud service; anything touching the data above should run on an on-premises model instead, so it never leaves the company network.
- What hardware do you need to run AI on-premises?
- It depends on model size and how many people use it at once. For small-scale ERP Q&A, a 7B–14B quantized model with a single 24GB-class GPU is enough. For a cross-department knowledge base, a 20B–32B model with 32GB VRAM is a better fit. If you need higher reasoning quality with multiple concurrent users, we'd recommend a 70B model at 4-bit quantization, with 48GB+ VRAM and 128GB of ECC system memory.
- Does MAQ have hands-on experience bringing AI into an ERP?
- Yes — on our own company's system. MAQ's inventory, purchasing, order, e-invoicing, cost/revenue reporting, and staff-scheduling systems are all self-built and in production, and we've already connected AI to that ERP: we can query inventory levels, the last purchase price and how it changed, a specific customer's purchase history, annual revenue, and best-sellers, all in natural language. Two principles turned out to matter most. First, never let the model generate its own SQL — MAQ exposes 11 read-only query tools with the SQL fixed in code, so the model can only choose a tool and structured parameters, and anything outside that whitelist is rejected, which keeps results stable and auditable. Second, enforce permissions at the entry point, using the existing back-office permission levels to decide who can open the AI interface. Finer, row-level control — say, letting each salesperson see only their own customers — needs to be designed into the data model from the start of the project; retrofitting it later is costly.
- Can MAQ just handle system development, without buying hardware?
- Yes. System development, integration with an existing ERP, and AI adoption consulting are all offered as standalone services, not bundled with hardware purchases. If the assessment shows a cloud architecture would actually serve you better, we'll say so.
- Can MAQ take over maintenance or a redesign of an existing system?
- Yes, including taking over maintenance, performance improvements, and feature additions for websites, e-commerce platforms, or internal systems originally built by another team. Before taking anything over, we review the code and architecture first, then quote once the risks and scope are clear.
- How long does a typical project take?
- It depends on scope. A data-integration and AI-query proof of concept for an existing ERP typically shows results within a few weeks; a full custom e-commerce or ERP build is measured in months. MAQ prefers to ship a small, usable version first, then expand once its real-world results are confirmed.