Skip to content

Keep AI out of half of it.

Shared by Derek from PeopleSoft Insider · November 22, 2026

Read the original on peoplesoftinsider.substack.com

Keep AI out of half of it.

View this post on the web at https://peoplesoftinsider.substack.com/p/keep-ai-out-of-half-of-it

And it happens to be the test for every PeopleSoft / AI decision you are going to make in the next two years.
A PeopleSoft developer versed in Integration Broker wrote to me this week. Twelve years on PeopleSoft, Integration Broker is his thing, and he’s helping his admin team automate patching. Java and WebLogic, PeopleTools, PUM.
He wanted ideas on where AI could eventually assist, or take it over entirely.
Then he wrote this:
“I understand it’s out of the PS env since it’s on the framework and may be more Change Assistant and patch downloading.”
He meant it as a disclaimer. An apology for asking a question he thought might be off-target.
It was the best sentence in the email, and he walked right past it.
Because what he’d done, without flagging it, was draw the line in exactly the right place. He had already separated the part of patching that’s framework and tooling from the part that’s actual decision-making. He just didn’t realize that line is the whole answer.
Here’s what I told him.
Split patching into two halves, and keep AI out of the first one.
The first half is the deterministic half.
Download the patch.
Stage it.
Run it.
Bounce the domains.
Verify the services came back.
That work has exactly one correct outcome every single time, and you already have excellent tools for it.
Change Assistant runs from a command line.
The DPK is Puppet underneath, which means your environment build is already code.
If any of it sits on OCI, Cloud Manager handles a good chunk of the provisioning and PUM cycle natively.
None of that needs a language model. Worse, a language model makes it strictly less reliable.
Think about what you’re actually asking for when you put a model in that path. You are taking a process that is repeatable by design and introducing a component that produces slightly different output on identical input. That is a feature when you’re drafting. It is a defect when you’re applying a patch to production at 2 AM.
You don’t want something that can be creative deciding whether to apply a patch.
The boring half should stay boring.
“But the whole point is to get the human out of it.”
I understand the goal, and I don’t disagree with it. But look at where the human time actually goes.
Nobody on your team is burning their week typing patch commands. That part already takes an afternoon.
They’re burning it afterward.
Which brings me to the second half, and this is where AI earns every dollar you spend on it.
After a PUM apply, somebody sits down with compare report output and goes object by object deciding keep-customization or take-Oracle.
Hundreds of objects.
Days of it.
It is high volume, it is genuinely analytical, it requires knowing why a customization exists, and it is the single largest consumer of human hours in the entire upgrade and patching cycle.
That is the judgment half. That’s your target. Not because it’s glamorous. Because it’s where the weeks go.
And notice what a wrong answer costs in each half.
A bad call on the mechanical side takes down an environment.
A bad call on compare triage gets caught by the person reviewing it, costs a minute, and teaches the system something.
You want AI where mistakes are cheap and recoverable, not where they’re expensive and public.
So here’s the test, and it’s the thing I actually want you to take from this. It works far beyond patching.
Ask whether the step has one right answer or a judgment call. Automate the first with code. Assist the second with AI.
Run anything through that. Payroll calc? One right answer.
Code.
Deciding which of four hundred deferred customizations still matter?
Judgment. Assist.
Nightly interface file transfer?
Code.
Reading a failed Application Engine log and telling you in plain language what broke?
Assist.
The deterministic half has been automatable for twenty years and most shops still haven’t done it.
That’s has nothing to do with an AI problem. That’s a scripting problem wearing an AI costume, and buying a model will not fix it.
If You Own the ERP / PeopleSoft Budget
Two things follow from this.
If someone brings you a proposal to put AI into a deployment or patching pipeline, the question to ask is not whether the technology works. It’s which half of the work it’s touching. If the answer is the mechanical half, you are paying a premium to make a reliable process less reliable.
And if your team has never automated the deterministic half with ordinary scripting, buying an AI tool will not skip that step. It will sit on top of the same manual process and you will have spent money to wrap it.
The order matters. Script the boring half first. Then put AI on the judgment half. Reversed, you get a demo instead of a result.
If You Place PeopleSoft People
This one’s worth repeating to clients, because it’s not obvious yet.
The developer who wrote to me leads with Integration Broker. That is not a legacy skill, and anyone treating it that way is about to be wrong in an expensive direction.
Oracle has been clear that their investment is in AI infrastructure rather than embedding AI into the applications. Which means the intelligence lives outside PeopleSoft and talks to it through integrations.
Service operations,
security,
throughput,
error handling.
The people who can design and secure that layer are the people who will build this. Not the people with the newest prompt library.
If you have IB-heavy candidates sitting on the bench, they are closer to the AI work than almost anyone else in your pipeline.
He thought he was apologizing for asking the wrong question.
He’d actually already solved it. He knew which part of the job was framework and tooling, and which part was decision-making. He’d just been told so often that AI goes everywhere that he didn’t trust his own line.
Trust the line.
So here’s what I’ll leave you with.
Look at whatever AI pilot is on your roadmap right now, and ask one question about it.
Is the step you’re pointing it at a step with one right answer, or a step that needs somebody’s judgment?
If it’s the first one, you don’t have an AI project. You have a script you never wrote.
Reply and tell me which half yours is on. I read every response.
THE JOB MARKET THIS WEEK
Architect and analyst titles are carrying the week, and the published rates are holding. The City of Seattle posting is the one to watch, because a public-sector hybrid role listing $54 to $82 an hour tells you what the midpoint actually looks like right now.
I’m working on the PeopleSoft Annual Market Report: for November 2026 and that should be posted in the community next week.
PeopleSoft Technical Analyst (ITP-B, BU) - Seattle, WA - City of Seattle - Hybrid - $54 to $82/hour
PeopleSoft HCM Solution Architect - McLean, VA - Guidehouse - Full-time
Product Owner, PeopleSoft Accounts Payable and eProcurement - Birmingham, AL - Genuine Parts Company - Full-time
ES OnPrem PeopleSoft Consultant - Australia - Hexaware - Contract
Campus Solutions, Applications Programmer V - Orlando, FL - University of Central Florida - Full-time - $89k to $140k
PeopleSoft Functional GL/KK Lead - Metaformers, Inc. - Contract
P.S. If you want to try the judgment half on something small this week, the Community AI Hub has the prompt library members are using for exactly this kind of work: reading logs, working through compare decisions, working through PUM updates, and getting unstuck faster on real PeopleSoft problems.
See you next week.
Derek Tomei Founder, PeopleSoftCareer.com Editor, The PeopleSoft Insider
Automate what has one right answer. Assist what needs a judgment call. Everything else is a demo.

Unsubscribe https://substack.com/redirect/2/eyJlIjoiaHR0cHM6Ly9wZW9wbGVzb2Z0aW5zaWRlci5zdWJzdGFjay5jb20vYWN0aW9uL2Rpc2FibGVfZW1haWw_dG9rZW49ZXlKMWMyVnlYMmxrSWpvMk9EY3lNall5T0N3aWNHOXpkRjlwWkNJNk1qSTFPRFl3TWpFekxDSnBZWFFpT2pFM09UVXlOekF3T1Rnc0ltVjRjQ0k2TVRneU5qZ3dOakE1T0N3aWFYTnpJam9pY0hWaUxUWTBOREk0TURZaUxDSnpkV0lpT2lKa2FYTmhZbXhsWDJWdFlXbHNJbjAuUnRNbGQtUFVrdTJBV3JjU01pbDV5Y3liXzluVnhSd080TWlpczhBZWp1VSIsInAiOjIyNTg2MDIxMywicyI6NjQ0MjgwNiwiZiI6dHJ1ZSwidSI6Njg3MjI2MjgsImlhdCI6MTc5NTI3MDA5OCwiZXhwIjoyMTEwODQ2MDk4LCJpc3MiOiJwdWItMCIsInN1YiI6ImxpbmstcmVkaXJlY3QifQ.Td_tF4OLHE2R0SASgqZ39MgilkdIhEhTJct9ScnCdyM?