How to Build an ERP Using AI (2026 Guide): A Step-by-Step Walkthrough with Google Antigravity
How to Build an ERP Using with Google Antigravity – Six months ago, if a business owner told me they wanted to build their own ERP system without hiring a development team, I would have talked them out of it. ERPs are notoriously expensive to build — most custom implementations run into tens of thousands of dollars, and off-the-shelf options like SAP or Odoo often force you to pay for modules you’ll never touch.
That calculation has changed. Google released Antigravity, an agent-first development platform, in late 2025, and by mid-2026 it had matured into something a non-programmer could actually use to build working software — including a basic ERP — by describing what they need in plain English instead of writing code line by line.
This guide walks through the actual process: setting up Antigravity, planning your ERP before you touch a keyboard, giving the AI agents instructions that produce usable results, and getting from a blank folder to a working system you can run your business on. I’m writing this for someone who owns or runs a business, not someone who already knows how to code — so we’ll skip the jargon wherever we can.
What “Building an ERP with AI” Actually Means
Let’s be clear about what you’re signing up for, because the marketing around AI coding tools tends to oversell it.
An ERP (Enterprise Resource Planning) system ties together the different functions a business runs on — inventory, invoicing, HR records, purchase orders, sometimes accounting — into one database instead of five spreadsheets that never match up. Building one from scratch, even with AI doing the heavy lifting, means you’re still making real decisions: what data you track, how modules connect to each other, who gets access to what.
What changes with a tool like Antigravity is who writes the code. You describe a feature — “when stock for an item drops below 10 units, flag it on the dashboard and email the purchasing manager” — and an AI agent writes the underlying code, tests it, and shows you the result. You’re acting more like a product manager than a developer. You still need to know your own business well enough to specify what you want, and you still need someone to review what gets built, even if that someone is you learning as you go.
If you want a full multi-warehouse, multi-currency ERP running for a 500-person company on day one, this guide isn’t for you — hire a systems integrator. If you run a business with, say, 5 to 60 employees and need inventory tracking, basic accounting, and staff records without paying for Oracle NetSuite, this is a realistic path.
Why Google Antigravity Specifically
There are other AI coding tools — Cursor, Windsurf, GitHub Copilot’s agent mode — and any of them can get you partway there. I’m using Antigravity for this guide for three practical reasons.
First, it’s free during its public preview, and it runs on Windows, macOS, and Linux, so you’re not locked into buying a Mac. Second, it gives you a choice of underlying AI models — Gemini 3 Pro, Claude Sonnet, or the open-source GPT-OSS — so if one model gets stuck on a task, you can switch and try another without switching your entire toolchain. Third, and this matters more than it sounds, Antigravity has what Google calls an “Artifact System”: every time an agent finishes a task, it produces a walkthrough, screenshots, and test results you can actually check, rather than just dumping code and hoping it works. For someone who can’t read code fluently, that verification layer is the difference between trusting the system and gambling with it.
Antigravity is built on a modified version of VS Code, the same editor a huge share of professional developers already use, so tutorials and troubleshooting for VS Code mostly apply here too.
Before You Open Antigravity: Plan the ERP on Paper First
This is the step people skip, and it’s the one that determines whether the next two weeks go smoothly or badly. AI agents are good at building what you ask for. They are not good at guessing what you actually needed but forgot to mention.
Sit down — literally with a notebook or a spreadsheet — and answer these questions before writing a single prompt:
What modules do you actually need? Most small businesses only need three: inventory, sales/invoicing, and basic accounting (or a bookkeeping export). HR and payroll add real complexity — tax rules, compliance — so consider whether a dedicated payroll tool integrated later makes more sense than building payroll from scratch.
What does your current data look like? If you’re moving off spreadsheets or an old system, export a sample. You’ll need this to show the AI what fields your inventory items or customer records actually have — SKU, unit cost, supplier name, reorder threshold, whatever’s specific to your business.
Who needs access, and to what? A warehouse staffer probably shouldn’t see payroll data. Write down three or four user roles now; retrofitting permissions later is one of the more error-prone changes to make in any system.
What’s your one non-negotiable feature? For a distribution business it might be automatic reorder alerts. For a services business it might be time tracking tied to invoicing. Name it, because it should be one of the first things you build, not an afterthought in week three.
I’d budget half a day for this planning stage. It feels slow compared to jumping straight into the tool, but every hour spent here saves several hours of rebuilding later.
Step 1: Install and Set Up Antigravity
- Go to antigravity.google/download and grab the installer for your operating system.
- Sign in with a Google account. The public preview doesn’t require a paid plan.
- On first launch, Antigravity will ask which model you want as your default — Gemini 3 Pro is a reasonable starting point since Google tunes the platform around it, but you can switch per-task later.
- Create a new project folder. Name it something plain, like
my-business-erp— you’ll be looking at this name in file paths for weeks, so skip the cute names.
Take ten minutes here to click through the interface. You’ll see two main views: the Editor View, which looks like a normal code editor with an AI chat panel, and the Agent Manager, where you assign tasks to agents and watch them work asynchronously — meaning you can give an agent a task and go do something else while it runs. For building an ERP, you’ll live mostly in the Agent Manager once the basic project structure exists.
Step 2: Describe the Foundation — Don’t Ask for Everything at Once
The single biggest mistake people make with agentic coding tools is asking for the whole system in one prompt: “build me a complete ERP with inventory, sales, HR, and accounting.” You’ll get something that looks impressive in a screenshot and falls apart the moment you try to actually use it, because the agent had to guess at hundreds of details you never specified.
Start narrow. Your first prompt should set up the skeleton:
“Set up a web application with a login page and a dashboard. Use a PostgreSQL database. I need three user roles to start: Admin, Manager, and Staff. The dashboard should be empty for now except for a welcome message showing the logged-in user’s name and role.”
This is intentionally small. It gives the agent a clear, checkable target, and it gives you something to click through and confirm works before adding complexity on top of it. Antigravity will show you its plan before executing — read it. If it proposes a database structure or a login method that doesn’t match what you had in mind, correct it before it starts writing code, not after.
Step 3: Build the Inventory Module
Once login and the dashboard work, move to the module you identified as most important during planning. I’ll use inventory as the example since it’s the most common starting point.
“Add an inventory module. Each item needs: name, SKU, category, unit cost, quantity on hand, and a reorder threshold. Managers and Admins can add or edit items; Staff can only view and update quantity on hand. Show a table of all items on a new /inventory page, with items below their reorder threshold highlighted in red.”
Here’s where Antigravity’s verification features earn their keep. After the agent finishes, it generates a walkthrough — screenshots of the inventory page, a description of what it tested, and often a short recording of a browser subagent actually clicking through the feature to confirm it works, not just that the code compiles. Go through this walkthrough item by item against your original spec. Don’t just glance at it and approve — this is the closest thing you have to a code review, and it’s the step that catches problems while they’re still cheap to fix.
If something’s wrong — say, Staff can edit prices when they shouldn’t be able to — don’t try to fix it yourself in the code. Tell the agent: “Staff should not be able to edit unit cost, only quantity on hand. Fix the permissions on the inventory edit form.” Let it iterate.
Step 4: Add Sales and Invoicing
With inventory working, connect it to sales:
“Add a sales module. When a Manager or Admin creates a new sale, they select items from inventory, and the quantity sold should automatically subtract from quantity on hand. Generate a simple invoice as a PDF with the business name, items, quantities, unit prices, and total. Store a record of every sale with a timestamp and the user who created it.”
This is a good moment to test something specific to your business rather than trusting a generic build. Create three or four test sales using real product names and prices from your actual inventory. Check that the math is right, that stock decreases correctly, and that the PDF looks like something you’d actually hand to a customer. If your invoices need a tax line, a discount field, or your logo, say so now — these are much easier to add while the invoicing feature is fresh than to bolt on six months later.
Step 5: Basic Accounting or Bookkeeping Export
Full double-entry accounting is a deep rabbit hole, and I’d steer most small businesses away from building it from scratch — the tax and compliance risk of a bug in your own accounting logic is real. A more practical middle ground:
“Add a page that summarizes total sales and total cost of goods sold by month, and lets me export all sales records as a CSV file formatted for import into QuickBooks.”
This gets you visibility into your numbers without you (or an AI agent) reinventing accounting standards. If you later want deeper integration, connecting to an existing accounting API is a safer path than building the ledger logic yourself.
Step 6: HR Records (If You Need Them)
If staff records are on your list, keep the scope tight at first — names, roles, start dates, and contact information, not payroll:
“Add an HR page, visible only to Admins, listing all staff with name, role, start date, and phone number. Add a simple form to add a new staff member.”
Payroll, tax withholding, and benefits administration involve legal requirements that vary by country and even by region within a country. I’d treat this as a separate project — likely using a dedicated payroll service — rather than something to build inside your ERP.
Step 7: Testing With People Who Aren’t You
Once the core modules exist, hand the system to two or three actual staff members — ideally the least technical people on your team — and watch them use it without helping. You’ll learn more from watching someone get confused by a button label than from a week of your own testing, because you already know how the system is supposed to work; they don’t.
Take notes on every point of confusion, then feed them back to Antigravity as specific fixes: “On the sales page, staff don’t understand that they need to click ‘Add Item’ before selecting a product. Make this clearer — maybe combine it into one step.”
Step 8: Deployment
Antigravity can help here too — ask it to prepare your application for deployment to a hosting provider (Google Cloud is a natural fit given the platform, and there’s an official codelab walking through exactly this). For a small business, a modest cloud instance is enough; you don’t need enterprise infrastructure for a system serving under a hundred users.
Before going live, ask the agent explicitly to review authentication and permissions one more time:
“Review the login and permission system for security issues. Check that Staff accounts genuinely cannot access Admin-only pages by changing the URL directly.”
This last check matters more than it might seem. A system that looks locked down in the interface can still leave a page accessible if you type its address directly, and that’s exactly the kind of gap a rushed AI-generated build can leave open.
What This Approach Gets Wrong, Honestly
It’s worth being upfront about the limits, since most guides on this topic skip them.
You will hit points where the agent misunderstands a request and builds the wrong thing twice in a row. Rephrasing helps more than repeating yourself. You are also responsible for data backups — an AI agent building your database doesn’t automatically mean someone is backing it up; set that up yourself or ask explicitly for it to be configured. And if your business has genuine regulatory requirements — inventory tracking for controlled goods, payroll tax compliance, healthcare data rules — get a professional to review the relevant module before you rely on it, the same way you would with any software vendor.
None of this means the approach doesn’t work. It means it works the way hiring a junior developer works: you get real output, but you stay involved, you check the work, and you don’t hand over decisions you don’t understand yet.
A Realistic Timeline
For a business with three core modules — inventory, sales, and basic staff records — expect two to four weeks of part-time work: a few hours a day reviewing agent output, testing, and refining requests. That’s slower than the marketing around “build an app in an afternoon” suggests, but it’s also considerably faster and cheaper than a traditional custom ERP build, which typically runs three to six months and a five-figure budget minimum.
Final Thought
The interesting shift with tools like Antigravity isn’t that AI writes code — that’s been true for a couple of years now. It’s that the verification layer — the walkthroughs, the browser subagents actually clicking through your app to confirm it works — has gotten good enough that someone without a programming background can supervise the process instead of just hoping it worked. That’s what makes building your own ERP a realistic project in 2026 instead of a pitch you’d only hear at a startup demo day.
Start small, plan your data before you prompt anything, and review every walkthrough like you’re signing off on an invoice — because in a sense, you are.