We Built a Working ERP System in Three Days: Agentic AI Training at Kacific Satellite Broadband

Short version: Over three days (2–4 September 2026), Dr. Alfred Ang ran an in-house agentic AI programme for the engineering and operations team at Kacific Satellite Broadband. Day 1 was Claude Code, Day 2 was Microsoft Copilot Studio, and on Day 3 the class built a real, deployable ERP system for Kacific’s own procurement workflow. The result is live: a working procure-to-pay application with AI agents and human approval gates, running at kacific-erp-tertiaryinfotech.vercel.app with the source at github.com/alfredang/kacificerpr. If you want to build the same class of agent-driven workflow on the Microsoft stack, our WSQ Business Process Automation with Power Automate and Copilot Studio Agents course covers it hands-on, funded up to 70%.

Who Kacific is, and why their workflow is hard

Kacific Satellite Broadband is a next-generation satellite broadband operator headquartered in Singapore. It runs Kacific1, a Boeing 702-platform satellite launched by SpaceX in December 2019, delivering high-speed Ka-band internet across the Asia-Pacific. The numbers give a sense of the scale: 56 spot beams covering 25 countries and a potential population of around 600 million people, from Bangladesh and Pakistan across to Fiji, Samoa and Vanuatu.

Kacific’s products range from PointCast self-pointing residential VSAT terminals through enterprise and government links, ISP bandwidth wholesale, portable SatPack units, and CommsBox disaster-response kits. In practice, that means connecting the places terrestrial fibre does not reach: island clinics, rural schools, remote mine sites, and emergency response teams after a cyclone.

Here is the operational problem that makes this an interesting AI training case. When your customers are spread over a quarter of the planet, so is your hardware. Terminal kits, modems, LNBs, cabling, and power systems sit in depots scattered across the Pacific and South-East Asia. Someone in Singapore has to know that a depot in Vanuatu is three modems away from being unable to fulfil an install, raise a purchase order against the right vendor before the lead time bites, get it approved by the right manager, receive the goods, and match the vendor’s invoice against what was actually delivered — at the agreed price.

That is a classic procure-to-pay problem, and it is exactly the shape of work that agentic AI is good at: high-volume, rule-governed, judgement-light in the middle, but judgement-critical at the approval points. It is also the shape of work that goes badly wrong if you let an AI act without a human gate.

The three-day structure

The programme was deliberately sequenced so that each day made the next one possible. Day 1 and Day 2 taught two different agentic platforms; Day 3 spent the accumulated skill on a real deliverable.

Day 1 — Claude Code: agentic coding in the terminal

Day 1 started with the mental shift that everything else depends on. Most people arrive at an AI coding tool expecting autocomplete. Agentic coding is a different activity: you describe an outcome, the agent explores the codebase, proposes a plan, edits many files, runs the tests, reads the failures, and iterates. Your job moves from typing to specifying and reviewing.

The practical skills we drilled:

• Write the specification, not the code. A vague instruction produces a plausible-looking wrong answer. A specification that names the files, the constraints, and the definition of done produces something you can actually merge.
• Give the agent a feedback loop. An agent that can run the test suite fixes its own mistakes. An agent that cannot is guessing.
• Keep changes reviewable. Small, scoped commits beat one giant rewrite you cannot audit — a rule that matters more, not less, when the code was machine-written.
• Persist project context. Conventions written down once (in a project instructions file) stop you re-explaining the same architecture every session.

Day 2 — Copilot Studio: agents inside the Microsoft stack

Day 2 moved to Microsoft Copilot Studio, because Kacific — like most enterprises — already runs on Microsoft 365. An agent that lives where the business already works beats a better agent that lives somewhere nobody visits.

We covered the pieces that make a Copilot Studio agent useful rather than decorative:

• Topics and instructions — shaping what the agent does and, more importantly, what it refuses to do.
• Knowledge grounding — pointing the agent at SharePoint documents and business data so answers cite reality instead of inventing it.
• Power Automate flows as tools — giving the agent the ability to act: look up a record, write a row, send a message, call an HTTP endpoint.
• Human-in-the-loop approvals — routing a proposed action to a person for a yes/no before anything irreversible happens.
• Governance — environments, connection permissions, and the boring-but-essential question of which identity the agent runs as.

This is the day that maps directly onto our WSQ Business Process Automation with Power Automate and Copilot Studio Agents course, which teaches the same stack over two days with SkillsFuture funding.

Day 3 — building the ERP for real

On Day 3 the class stopped doing exercises and built the thing. The brief was Kacific’s own problem: consolidate purchase orders, vendor invoices, per-depot SKU stock and low-stock alerts into one procurement workflow that works across a distributed network.

By the end of the day there was a deployed application. You can see it running at kacific-erp-tertiaryinfotech.vercel.app and read the source at github.com/alfredang/kacificerpr.

What the software actually does

The Kacific ERP is a procurement, inventory and vendor operations system for a Ka-band satellite broadband network. It is seeded with realistic (but fictional) data — terminal kits, modems, power systems — across 8 depots, 10 vendors and 34 SKUs, with 16 purchase orders and 12 invoices already in flight so the workflow can be demonstrated end to end.

Procure-to-pay, with gates

The core of the system is a purchase order that moves through an explicit lifecycle:

1. Draft — created by a requester, or drafted by an AI agent from a plain-language instruction.
2. Pending approval — a human gate. Approvals can arrive by signed email link, an Asana card, a Telegram message, or an API call.
3. Approved and ordered — sent to the vendor.
4. Received — goods booked in at the destination depot, updating stock.
5. Closed — after the invoice is matched and paid.

Invoices go through 3-way matching: the purchase order, the goods receipt, and the invoiced price are compared within a tolerance. A match inside tolerance is straightforward; anything outside it becomes a dispute that a finance user has to look at. This is the classic control that stops an organisation quietly paying for goods it never received.

Inventory across a distributed network

Stock is tracked per depot with full movement history and per-SKU reorder levels. The system scans the whole network for shortfalls and groups them by preferred vendor, so instead of eight separate small orders you get one consolidated purchase order per vendor, generated with one click. For a business whose depots are islands, consolidating shipments is not a nice-to-have.

The agentic layer

This is where the training showed up in the product. The system carries a set of AI agents that propose work but never unilaterally commit it:

• PO drafting — turn “we need 20 more modems in Suva before the end of the month” into a structured draft purchase order against the right vendor.
• Reorder recommendations — read stock levels, lead times and reorder points, and suggest what to replenish.
• Invoice-match assistant — explain why an invoice failed to match and what the discrepancy is.
• Vendor risk summary — summarise a vendor’s delivery and pricing behaviour.
• Co-pilot chat — a Hermes chat widget in the app, plus a Telegram bot, both answering questions against live data.

Every agent proposal is a proposal. Purchase order submission, invoice matches, and all agent output pass through a human approval gate before they become facts. That design decision is the single most important thing in the build, and it is the point we hammered hardest during the training.

Built to be integrated

The system exposes a versioned REST API at /api/v1 with OpenAPI 3.1 documentation, plus an MCP endpoint so external AI agents can query and act on it directly. Outbound events are delivered as HMAC-signed webhooks with retry logic, and scheduled jobs handle low-stock scanning and digest emails. Approvals integrate with Asana boards, notifications with Telegram, and email through Resend.

The technical stack

Security was not left as an exercise: Argon2id password hashing, httpOnly JWT sessions with server-side revocation, single-use hashed approval tokens, AES-256-GCM encryption for stored secrets, HMAC-signed webhooks, and a full audit log. Access is governed by seven roles — admin, manager, procurement, finance, sales/requester, operations, and read-only viewer — so a salesperson can raise a request but only a manager can approve it.

What the team actually learned

Stripping away the tooling, four lessons carried across all three days.

1. Agentic AI is a workflow design problem, not a prompting problem. The hard part was never getting the model to write code or answer a question. It was deciding where a human must stand in the loop, and making that gate impossible to bypass.
2. Give the agent tools and a feedback loop. An agent that can read the database, run the tests and see its own errors is useful. An agent that can only talk is a demo.
3. Ground the agent in your own data. Generic AI gives generic answers. The moment the agent could read actual depot stock and vendor lead times, the output became something operations staff would act on.
4. Ship something small and real. Three days produced a deployed, working system precisely because the scope was one workflow done properly, not an entire enterprise suite done badly.

Want to build this on the Microsoft stack?

Day 2 of the Kacific programme — Copilot Studio agents, Power Automate flows, connectors, knowledge grounding, and human-in-the-loop approvals — is available as a public WSQ course in Singapore.

WSQ Business Process Automation with Power Automate and Copilot Studio Agents (TGS-2022017524) runs over 2 days / 16 hours and covers business process automation with the Power Platform, workflow automation in Power Automate, building AI agents in Copilot Studio, integrating the two, enterprise automation scenarios, and deployment and governance.

Course fee is $750 before GST, with WSQ funding of up to 50% for Singapore Citizens and PRs aged 21 and above, and up to 70% under MCES for Singapore Citizens aged 40 and above, plus SME support. SkillsFuture Credit, PSEA and SFEC can be applied. Check the dates and register here.

We also run this and other agentic AI programmes in-house for corporate teams, built around your own systems and data — which is exactly what Kacific did. If you would like a version of this three-day programme scoped to your workflow, get in touch.

Frequently asked questions