I actually really like what NetSuite is doing with AI prompts.
That may surprise a few people who have read some of the things I've written about AI and ERP lately. I'm not particularly interested in putting a chatbot on top of an ERP and pretending we've solved something. But Prompt Studio is different. I think there is a genuinely good idea in there.
NetSuite is giving administrators a way to create and manage prompts, use variables, connect them to business context and call them from SuiteScript. Text Enhance can already use information from the record you're sitting on as context. Ask Oracle is starting to reach further into NetSuite data, including saved searches.
That's cool.
More importantly, it starts putting AI somewhere I think it belongs. Inside the application, close to the transaction, with some control over what context it gets and what we're asking it to do.
But I think it's unfinished.
Not because the AI isn't clever enough. The LLM is probably the least interesting part of the problem.
The problem is that giving an AI access to ERP data doesn't necessarily give it access to what the business means by that data.
Take what sounds like a ridiculously simple question:
Why is gross margin down this week?
There's probably no shortage of data in NetSuite that could help answer it. There are transactions, accounts, customers, items, locations, classes and probably several saved searches somebody has built over the years.
But what exactly did I just ask?
What does Gross Margin mean in this business? Are sales claims included? What about rebates? Are we talking manufactured margin, third-party margin or both? Which week is "this week"? Is "down" compared with last week, budget or the equivalent week last year? Which customer hierarchy are we using? Are we talking accounting gross profit, invoice margin or whatever management has decided to call Achieved Gross Margin?
These aren't AI questions.
They're business questions.
And every reasonably complicated ERP I've ever worked with has accumulated answers to them somewhere. Some are in configuration. Some are in saved searches. Some are buried in workbook formulas. Some live in spreadsheets. Some are in scripts written six years ago by somebody who has since left the company.
And some exist only because Bazza in Finance knows that account 40100 doesn't really belong where the chart of accounts says it belongs.
Now give an LLM access to all of that.
Congratulations. You've given AI faster access to the ambiguity.
That's the bit I think gets missed in a lot of the discussion about AI and ERP. We keep concentrating on how easily the AI can reach the data, or how cleverly we can write the prompt, when the difficult part is deciding what the data means once it gets there.
This is also where the CSR Semantic Map starts to get interesting.
I originally built the Semantic Map because I had a reporting problem to solve. I wanted to take messy accounting data and turn it into stable business reporting without burying another mountain of business logic inside scripts and saved searches.
So instead of teaching every report separately that this account belongs to Net Sales, that account is a discount, these classes mean Manufactured and Third Party, this sign needs to be reversed, and this particular rule takes precedence over the default rule, we describe those things once.
The scripts can then be relatively dumb. - They read the rules.
That has turned out to be useful for reporting. But it's potentially much more useful for AI. Because Prompt Studio and the Semantic Map solve two different problems. The prompt tells the AI what I'm asking it to do. The Semantic Map tells it what the business means.
That distinction matters.
If I ask why Gross Margin is down this week, I don't want to write a magnificent 1,500-word prompt explaining our chart of accounts, fiscal calendar, reporting hierarchy, sign conventions, customer structure and management definition of Gross Margin.
I don't want every user doing that either. And I definitely don't want the AI figuring it out differently every time somebody asks.
CSR already has that context. It knows which accounts contribute to the measure. It knows the classifications. It knows the reporting signs. It knows the fiscal calendar. It knows the hierarchy. It knows which rules override other rules.
We can give it the governed facts too.
Because CSR has persisted the resulting facts through SpeedLedger or Weekly Facts, we can go another step further. We aren't asking the AI to reconstruct the answer from hundreds of thousands of transaction lines every time somebody asks the question.
So the architecture starts becoming surprisingly simple.
The prompt provides the instruction.
The CSR Semantic Map provides the meaning.
The curated facts provide the evidence.
The LLM provides the reasoning and the language.
That's much more interesting to me than another chatbot attached to an ERP. It also means I don't think NetSuite needs to become an AI company and somehow teach a language model every strange piece of business logic accumulated inside every NetSuite account on earth.
They've already started building the pieces that let customers bring AI into the application. Prompt Studio is one of those pieces. SuiteScript is another. Saved searches are another.
The ERP itself already contains an enormous amount of useful context.
CSR can complete the other side of that architecture by giving those tools something they don't naturally have: a governed description of what this particular business means when it talks about its numbers.
And once you've done that, the AI question becomes much easier.
Not:
Can AI read my NetSuite data?
Of course it can.
The much more useful question is:
Can AI understand my NetSuite data the same way my business does?
That's the bit worth solving.