Enterprise Agents Are Growing Teeth

Yesterday you could feel the enterprise AI story trying to become less mystical and more mechanical.

Anthropic’s latest push for enterprise agents comes bundled with a very specific promise: don’t just chat with a model—let it actually do work inside the tools where money moves and mistakes hurt. The details matter here. “Plug-ins for finance, engineering, and design” is not a vibe. It’s a claim about integration, permissioning, and workflows that have to survive contact with real organizations.

If you’ve spent the last year watching agent demos, you’ve probably noticed the pattern. The model is impressive for thirty seconds, then the demo quietly assumes the hard parts away. The agent has perfect access. The data is clean. The tool always responds. Nothing breaks. No one asks, “What happens when it’s wrong?” That’s why this new wave of enterprise positioning is interesting: it’s trying to make the demo live in the same rooms as procurement, compliance, and your CFO’s blood pressure.

In the same 24 hours, OpenAI’s COO said the quiet part out loud: we haven’t yet seen AI truly penetrate enterprise business processes. That’s not a knock on capability. It’s an indictment of adoption. Most companies don’t fail to use AI because the model can’t write a paragraph. They fail because the paragraph is never the job.

The job is: open the spreadsheet, reconcile the numbers, update the CRM, attach the supporting doc, notify the right person, and leave an audit trail you can explain in six months. The job is: draft the contract clause, cite the precedent, route it to review, and record who approved what. The job is: make a decision inside a system that was designed for humans to be slow and careful.

This is where “plug-ins” stop being a cute idea and start being the entire game. A plug-in is a bridge between language and action. It’s how you turn “please pull last month’s churn by segment” into a concrete query that runs against the right dataset, in the right environment, with the right access. It’s also where everything can go wrong—because the moment a model can do something, it can also do the wrong thing.

Google adding automated workflows to Opal is the same story from a different angle. Workflows are the grown-up version of prompts. A workflow says: here’s the trigger, here are the steps, here’s the output, and here’s what we do when reality refuses to cooperate. It’s not about one smart answer. It’s about repeatable behavior.

Here’s the turn that’s worth sitting with: enterprise AI isn’t stuck because the models are too dumb. It’s stuck because business processes are not made of text. They’re made of permissions, state, and accountability. When an AI system updates a record, that update becomes part of the company’s memory. When it changes a number, someone will eventually have to defend it. When it sends an email, it becomes your brand voice in someone else’s inbox.

So if you’re wondering what to do with this as an everyday operator, the move is not “go buy an agent platform.” The move is smaller and more honest. Pick one workflow you run every week that has three properties: it is repetitive, it is annoying, and it has a clear definition of “done.” Then design the system like you don’t trust yourself on a tired Tuesday.

Make AI responsible for the first draft and the first pull of context. Make a human responsible for the final commit. And make the system responsible for leaving receipts: what it touched, what it produced, what it couldn’t access, and where it asked for help.

You’ll know the enterprise agent era is real when the question stops being “How smart is the model?” and becomes “How safe is the handoff?” When your team can let an AI operator do the boring parts—without losing the ability to undo, explain, and improve—then you don’t just have AI. You have throughput.

And that’s the boring secret: the next big AI advantage won’t come from a clever prompt. It’ll come from building the smallest workflow you can actually run, every day, with guardrails you can live with.