We connect Claude or ChatGPT to your NetSuite account with its own read-only role, test it in your sandbox, and check its answers against your own reports. AI prepares it. Your team approves it.







Connecting an AI assistant to NetSuite takes an afternoon. Trusting what it says about your books is the part nobody shows you.
What was Q3 revenue for the whole group, in HKD?
The question and the report cover the same scope: every subsidiary, in HKD, July to September. The assistant doesn’t show what it used, so nobody can say which figure is wrong.
Only what someone pastes in
Self-checked · by whoever typed it inA snapshot from the day it was run
Unchecked · needs the source file and its dateEvery subsidiary its role can see
Unchecked · needs scope, currency and period shownCan create and change records
Unchecked · needs a named approverA safe integration needs scoped access, traceable answers and approval for write actions.
One fixed scope, agreed before we start. You see what the assistant gets right and wrong in your sandbox before it goes near production.
Start with a scoping callWe start with one thing that is painful this month: bills retyped from PDFs, a report rebuilt by hand every close, questions that wait a week for a saved search. The setup is scoped around that job.
An assistant repeats whatever is in the records. We look at the chart of accounts across entities, duplicate vendor names and intercompany mapping, and tell you what needs tidying before the answers can be trusted.
The connection gets its own role, never an administrator login. It sees the records the job needs and nothing else, and it cannot create, edit or delete. We handle the connector permissions and single sign-on along the way.
Your team writes ten questions it already knows the answers to, by period and by subsidiary. We run them in sandbox and put each AI answer next to the matching NetSuite report. You see what ties out and what doesn’t.
Most wrong answers come from the assistant not knowing your fiscal periods, your consolidation structure or your base currencies. We write those rules down for it, then run the test set again until the answers match.
Read-only questions go to production first, with your team trained on what to ask and what to double-check. Drafting, such as supplier bills prepared for approval, is a second scope. Every draft waits for a named person.
The first stage can look and answer. It cannot post, edit or delete, because the role it runs under doesn’t allow it.
When drafting is switched on, each draft sits next to its source document until someone on your team signs it off.
The connection has its own role and login, so its logins and anything it drafts are on the record, separate from what your people did.
You get a written note of what the assistant can see, what it was tested on and where it was wrong, with a consultant’s name on it.







Answers from your own recordsRead-only role · Sandbox tested · Checked by periodThe assistant reads NetSuite through a connection Oracle provides. It doesn’t keep a second copy of your books, so there is one set of numbers to check.
Ask in plain English and get the figure with the records behind it, so anyone can click through and check.
The assistant reads the supplier document and prepares the bill. Your AP team checks the exceptions and stops keying in the rest.
Describe what you need and get a first version to review. Your admin spends their time checking it, not building it.
Consolidated or by subsidiary, in the right period and currency, because the assistant has been told how your group is built.
Where the test set shows it can’t be trusted yet, we say so in writing. That list is as useful as the list of what works.
Already built something with Claude or ChatGPT? We’ll review it and keep what’s sound. We don’t start again for the sake of it.
Most AI demos run on one entity in one currency. Your ledger doesn’t look like that, so the test set doesn’t either.
Questions are tested per entity and consolidated, so a group total is never one company’s figure by mistake.
We check that amounts come back in the currency you asked for, at the rate NetSuite used, for HKD, RMB and USD books.
Fiscal year-ends, adjustment periods and locked periods are written into the instructions the assistant works from.
Day-month-year dates are a known source of wrong answers. They go in the test set from the start.
English and Chinese versions of one supplier are matched up during the data review, before the assistant double-counts them.
This setup uses NetSuite’s AI Connector, which doesn’t depend on NetSuite Next. Oracle hasn’t published a Hong Kong date for Next.
Finance wants to know what the assistant can’t do. Admins want to know exactly how it’s wired.
The assistant starts with read-only access, its answers are checked against reports you already trust, and nothing reaches the ledger without a person approving it.
Talk through the controlsThe aim is fewer keystrokes on routine bills and more time on the ones that don’t match, with the source document on screen beside every draft.
Scope an AP use caseNetSuite’s AI Connector Service (its MCP server), a custom non-administrator role, OAuth 2.0, and a sandbox run before production. Your AI subscription is your own.
Ask a technical questionWe have been really impressed by ONE Pacific’s responsiveness and your knowledge within the finance automation space.

Straight answers, including the ones where the answer is “not yet”.
Ask a consultantThree things are worth setting up now: answering plain-English questions from your own records, drafting saved searches and reports for someone to review, and preparing supplier bills from documents for a person to approve. Each one needs testing against your own reports before you rely on it. Letting a model post transactions on its own is not something we set up. Our review of what works and what doesn’t goes through the detail.
NetSuite’s AI Connector is built on an open standard, so it isn’t tied to one assistant. You bring your own subscription. On the scoping call we confirm which assistants your NetSuite account and your IT policy support before you commit to anything.
They probably could, and for a first look that’s fine. The connection is the easy part. The work is in what the assistant is allowed to see, whether its answers match your reports across subsidiaries, currencies and periods, and who is answerable when it’s wrong. That is what the role design and the test set are for. If your admin has already built something, we’ll review it instead of redoing it.
You check them, with us, before go-live. Your team writes ten questions it already knows the answers to. We run them in sandbox and put each AI answer next to the matching NetSuite report, by period and by subsidiary. Where it’s wrong we fix the cause or tell you not to use it for that question.
Not in the first stage. The connection runs under a read-only role, so it can’t. Drafting comes later as a separate scope, and every draft waits for a named person on your team to approve it.
Only what the dedicated role can see, and only when someone asks a question. We write a short data-handling note for your account: which records are in scope, which provider plan you’re on and what its retention terms are. It doubles as the written policy on what staff may put into AI tools that Hong Kong’s Privacy Commissioner recommends. For groups with Mainland entities, we flag the data that should stay out of scope.
We strongly prefer one, because it lets you see mistakes where they cost nothing. If you don’t have a sandbox, tell us on the scoping call and we’ll go through the options.
No. The AI Connector is separate from NetSuite Next. Oracle is releasing Next in North America first and hasn’t published a Hong Kong date; we track it in our note on timing. We’ll also tell you which of NetSuite’s built-in AI features are worth switching on in your account, and which aren’t yet.
There are two parts. Your AI assistant subscription is paid to the provider directly. Our setup is a fixed scope with a price agreed before we start. Oracle doesn’t charge for the AI Connector itself, and using it doesn’t draw on your AI Units. Some of NetSuite’s own built-in AI features are metered separately, and we’ll check what applies to your account instead of guessing.
Tell us the task that eats the most time. In 30 minutes we’ll say whether an assistant can take it on safely in your account, and what we’d test first.
OnePac’s team provided vision and perspective on how long it should take or what we should really think about the path forward.
