RICE Scoring an AI Enabled NetSuite CSV Importer

Last night's post on queue-based processing for AI enabled apps ended with a throwaway line: "I think I might have to build this app."

The app in question was an AI enabled CSV importer for NetSuite — upload a pile of files, drop the records on a queue, let workers chew through them in the background, and have AI do the field mapping and clean up the failures. It started life as nothing more than an example of why queues matter. By the end of writing it, the example was more interesting to me than the point it was supposed to be illustrating.

So rather than let a decent idea die at the bottom of a blog post, I had Claude Cowork read the post and write it up as a real idea in Application Manager — our AI Application Platform. Description, why it matters, nine features underneath it, dependencies between them. Then I asked for a RICE score.

A quick refresher on RICE

RICE = (Reach × Impact × Confidence) ÷ Effort.

Four numbers in, one number out. The score itself is not the valuable part. The valuable part is that you have to say out loud why you picked each number, and then somebody gets to disagree with you.

Here is what came back for the CSV importer:

• Reach — 100. The argument being that this is not a migration feature aimed at migration customers. Anybody who loads data into NetSuite at volume has hit the single concurrent import, the 25K line cap, or both.
• Impact — 4 out of 5. A serious time saver, not a business-changing event. Nobody buys NetSuite for the import screen.
• Confidence — 70%. More on this below, because it's the interesting one.
• Effort — 8. Greenfield app. Queues, workers, a write layer, AI mapping, error correction, and analytics on top.

Score: 93.

That's a pretty good RICE score. Good enough that it isn't going to sit at the bottom of a list of app ideas (I already have a bunch of dead or low priority ideas).

The number worth arguing about

Confidence at 70% is the honest one, and it's the one I'd push on if somebody handed me this.

The queue half of this app is not speculative anymore. Queues ship by default in every Application Manager app now — that was the whole point of writing about them in the first place. Multi-file upload, workers, retries, per-record status: known quantity.

The AI half is not. How well does a model map a stranger's CSV columns to NetSuite fields when the headers are named whatever the customer's spreadsheet guy named them in 2014? Don't know yet. How much write throughput does a real account give up before governance limits start pushing back? Also don't know — and that one matters, because the entire pitch is beating NetSuite's own import ceiling. Blow past the account's actual concurrency limit and you haven't built a faster importer, you've built a pain in the neck.

70% says: the risky parts are real, they're identified, and they're measurable before anybody writes much code. Not a rubber stamp. Not a shrug either.

Why bother scoring a blog post?

Because ideas are cheap and the pipeline is where they go to get expensive.

Everybody has the "I should build that" moment. What usually happens next is nothing, because the distance between a good sentence and a scoped, scored, prioritized piece of work is a couple of hours of unglamorous writing that nobody wants to do at 11pm. Handing that gap to Cowork took about five minutes, and what came out the other side is something I can actually argue with — features I can reorder, numbers I can dispute, dependencies I can call wrong.

Is 93 the right score? Maybe not. Reach might be optimistic. Effort at 8 assumes the NetSuite write layer behaves, and that assumption has burned better developers than me.

But an idea I can argue about beats an idea I liked once. More on this app as it moves — and if it turns out that governance limits kill it, you'll get that post too.

BTW, Claude Cowork understands how to write up application ideas and feature records and put them into Application Manager because I trained it to do this. Also, every Application Manager app and Application Manager itself has an MCP service built-in. This makes the process fluid and complete.

AI is amazing.