You’re in the office. Someone’s talking about AI again.

They say it worked. “We ran a proof-of-concept,” they say. “It nailed the summary, cited all the right things. Scary good.” Everyone nods like something profound just happened. Someone else jokes about being replaced. You open another tab.

Here’s what no one mentions: the AI doesn’t remember what it said. It doesn’t care. It’s not logging the output. It’s not versioning the result. There’s no audit trail. No owner. It generates. You copy. You paste. Then what?

Nothing. It disappears.

You’ve made a ghost. Again.

⸻

RAG Is a Buzzword with a Brain

RAG—Retrieval-Augmented Generation—isn’t the problem. It’s fine. It works. It does things.

You point it at a curated corpus—journal scopes, editorial notes, XML metadata, whatever—and suddenly the AI isn’t hallucinating. It’s sourcing. Citing. Playing nice. Saying things that sound real. Sometimes are real.

Everyone claps. Great demo.

Then they forget to ask: where does this fit?

⸻

The Human Isn’t Optional

Here’s the thing: RAG doesn’t know what it doesn’t know. It fetches, it rewrites. But it can’t judge.

It can’t tell you if a generated classification actually maps to your taxonomy. It can’t flag tone. It can’t prioritize accuracy over fluency. It doesn’t care if it’s wrong.

That’s your job. Or hers. Or someone else’s. A human. In the loop. Always.

If there’s no human in the loop, you’re not building a workflow. You’re building a liability.

⸻

What Process Looks Like (Spoiler: It’s Boring Until It’s Not)

Process is where the work hides.

You diagram the steps. You define the trigger points. You build the API calls. You log the output. You insert a review task. You add an approval button. You don’t call it AI anymore. You call it “step seven.”

That’s orchestration. BPMN. Camunda. SiteFusion. Whatever.

AI happens inside the process. Not beside it.

⸻

Alt-Text, or: The Things We Pretend We’ll Remember to Do

Case study: alt-text generation.

Without process:

• Image uploaded
• Someone forgets
• Someone remembers
• AI runs
• Text shows up in an email
• Someone pastes it in, maybe
• No one reviews
• It ships

With process:

• Image uploaded
• Workflow triggers
• RAG fetches context, generates
• Human reviews
• Output logged, versioned, embedded
• Pipeline continues

You don’t hope it gets done. You know it did.

⸻

Process Makes AI Real

Everyone wants the AI to be magic. No one wants to talk about governance. But governance is what makes the magic repeatable.

You don’t scale by hiring more prompt engineers. You scale by building systems where AI is just another service node—like translation, or metadata enhancement, or a CMS plugin.

You wrap it in BPMN. You drop it in a workflow. You surround it with rules, checkpoints, users, versions, logs.

Now it’s real.

⸻

This Only Matters If You Publish Things That Matter

If you’re in scholarly publishing, you don’t just push content. You push evidence. Citations. Provenance. Structured meaning.

And if that’s what you traffic in, then the process isn’t overhead. The process is the product.

⸻

Next

We’re going to talk about structured content.

The stuff everyone forgot about in the GenAI hype cycle. The reason RAG works. The real asset.

XML. Metadata. Controlled vocabularies. The bones of meaning. You’ve already built it. You just don’t know how powerful it is—yet.

#ssp2025