The Quiet Takeover — Embedded AI vs. the Agent Hype

Last week I had a conversation with a managing partner at one of the big analyst firms about the enterprise AI landscape. We were about thirty minutes in when he said something that stuck with me: "The companies getting the most value from AI right now aren't the ones building agents. They're the ones whose vendors already shipped them."

That one line changed how I think about enterprise AI strategy.

Everyone's talking about building AI agents. The conference slides, the LinkedIn posts, the vendor keynotes — it's all agents, agents, agents. But while everyone's debating which LLM to use for their custom agent framework, something quieter is happening. The platforms you already pay for — SAP, Microsoft, Oracle, ServiceNow — are embedding AI directly into the workflows you run every day. And most enterprise teams don't even know it's there yet.

This issue is about that gap. What's actually shipping, what it can do right now, and how to figure out whether the embedded stuff is enough — or whether you still need to build.

What shipped while you weren't looking

Let's get specific, because the specifics are what matter.

SAP now has dozens of specialized AI agents and more than 2,100 Joule skills. Joule Studio — their agent builder — went generally available in Q1 2026. You can build custom agents that connect directly into your S/4HANA environment through the AI Agent Hub. The catch? Most SAP customers haven't activated Business AI in production. The tools are there. The adoption isn't.

Microsoft has Copilot embedded across the entire Dynamics 365 suite. In Finance, it reads incoming invoices, suggests GL coding based on vendor history, routes approvals based on thresholds, and predicts which customers will delay payment. In Supply Chain Management, it flags external disruptions (weather, supplier financials, logistics issues), surfaces impacted orders, and drafts alerts to affected partners. Over 90% of Fortune 500 companies now use Microsoft 365 Copilot. But for most, “adopted” still means pilots, not production.

Oracle launched Fusion Agentic Applications in March 2026 — 22 AI-driven apps embedded in Fusion Cloud for finance and supply chain. These aren't chatbots. The Claims Settlement Workspace helps finance teams settle claims faster. The Collectors Workspace improves promise-to-pay conversion rates. And Oracle AI Agent Studio lets you modify pre-built agents or create new ones that operate inside the transactional system — not on top of it.

ServiceNow is perhaps the furthest along on actual production adoption. Their Now Assist customers spending over $1 million annually grew 130% year-over-year. CEO McDermott says their 2026 AI revenue target of $1 billion will "blow past a billion and a half." Robinhood deflects 70% of employee requests with ServiceNow AI before they reach a human — eliminating roughly 2,200 hours of manual work per month. An online travel company delivered 11 million autonomous AI resolutions annually for HR and IT alone, with over 230% ROI.

Those aren't pilot numbers. Those are production numbers.

Why embedded AI is winning on adoption

The answer is almost boringly simple: friction.

A standalone AI agent requires a business case, a procurement process, a security review, an integration project, governance decisions, change management, and training. That's 6-12 months before anyone sees value.

An embedded AI feature requires... the next software update. It shows up in the interface your team already uses. No separate login. No new vendor relationship. No integration. The adoption barrier drops to near zero.

This is what the major analyst firms are seeing when they predict 40% of enterprise applications will include task-specific AI agents by end of 2026 — up from less than 5% in 2025. The growth isn't coming from companies building custom agents. It's coming from vendors embedding agents into existing applications.

And there's a second-order effect that most people miss: embedded AI doesn't need to be "approved" by IT in the same way. It arrives as part of a platform update you've already licensed. It lives inside existing security boundaries. It uses data that's already governed. The compliance team barely needs to get involved.

Compare that to the standalone route: OutSystems surveyed nearly 1,900 global IT leaders and found that 94% are concerned that agent sprawl is increasing complexity, technical debt, and security risk. Only 12% have a centralized platform to manage it. The rest are winging it — different agents for different teams, no standard governance, no unified view of what AI is doing across the organization.

Where standalone still wins (and it's not where you'd expect)

So should you just wait for your vendors to ship everything? No. And here's why.

Embedded AI is excellent at optimizing existing workflows — the processes your vendor already understands. Invoice matching in SAP. Collections in Oracle. Ticket routing in ServiceNow. These are well-defined, high-volume, single-system tasks. Perfect for embedded automation.

But enterprise reality is messier than any single vendor's product.

Cross-platform orchestration. Your actual end-to-end process probably touches SAP for finance, ServiceNow for approvals, Salesforce for customer data, and three SharePoint sites nobody maintains. No single vendor's embedded AI can orchestrate across all of that. A standalone agent can.

Novel use cases your vendor hasn't imagined. The embedded agents handle standard processes. But what about the weird, company-specific workflow your operations team built in 2019 that somehow became mission-critical? No vendor is building embedded AI for that. You need a custom agent.

Multi-step reasoning across data sources. Embedded AI is getting smarter, but it's still mostly single-task. "Summarize this invoice." "Flag this risk." A standalone agent can chain: pull last quarter's supplier performance data from SAP, cross-reference it with contract terms in your CLM tool, check current market pricing, and draft a renegotiation recommendation with specific numbers. That kind of multi-hop reasoning across systems is where standalone agents still dominate.

Model flexibility. Embedded AI locks you into your vendor's model choices. SAP uses its own models plus select partners. Microsoft uses Azure OpenAI. If a better model ships next quarter — or if you need specialized reasoning for a specific domain — you can't swap. Standalone gives you that flexibility.

The practical framework: when to embed, when to build

Here's how I'd think about it, and this is the part worth sharing with your team.

Default to embedded when:

• The process lives entirely within one vendor's system
• It's a high-volume, well-defined task (invoice processing, ticket routing, report generation)
• You need value in weeks, not months
• Your team isn't technical enough to maintain custom AI
• Governance and compliance are paramount

Build standalone when:

• The process crosses multiple systems
• You need the AI to reason across different data sources
• The use case is company-specific and no vendor has addressed it
• You need model flexibility or cutting-edge capabilities
• You have the engineering team to maintain it

And here's the move most enterprises miss: use embedded AI as your foundation, then build standalone agents only for the gaps. Don't build custom where your vendor has already shipped. That's expensive duplication. But don't assume embedded covers everything either — it won't.

Five things you can do this week

This is the part I want you to forward to your colleagues.

1. Audit what's already available. Log into your SAP BTP cockpit and search for Joule under Entitlements. Check if your Dynamics 365 tenant has Copilot features enabled (Admin center → Settings → Copilot). Open your ServiceNow instance and look for Now Assist in the application navigator. Odds are you're paying for AI capabilities you're not using.

2. Run the Joule booster. If you're on S/4HANA Cloud, navigate to your BTP Global Account → Boosters → search "Joule" → click Start. It automatically checks prerequisites and configures the integration. This takes about 30 minutes, not 30 days.

3. Try one embedded feature in production this week. Not a pilot. Not a sandbox. Pick one low-risk, high-volume task — invoice coding suggestions in Dynamics 365, ticket summarization in ServiceNow, or Joule-assisted data entry in S/4HANA — and turn it on for one team. Embedded AI is designed to be low-risk. The worst case is that your team ignores the suggestions.

4. Map your cross-system processes. Before you build any standalone agent, draw the actual end-to-end flow of your three most painful processes. Where does data move between systems? Where do people copy-paste between screens? Those handoff points are where standalone agents add value — and where embedded AI can't help.

5. Ask your vendor one question. Call your SAP, Microsoft, or Oracle account manager and ask: "What AI capabilities are included in my current license that we're not using?" I guarantee the answer will surprise you. Most enterprise licenses now include AI features that nobody has activated.

Deep dive: what Joule actually looks like inside SAP IBP

I want to make this concrete, because "embedded AI" is meaningless until you see it in a real module. I spent years working in SAP IBP — Integrated Business Planning — so let me show you what Joule does there right now.

The first thing that surprised me: you can ask Joule to run a master data health check in natural language. Just type "run a master data health check for products" and it kicks off the job, then shows you the results in a table inside the Manage Master Data app. If you've ever spent a Friday afternoon manually auditing location-product combinations before a planning run, you know why this matters.

But the more interesting capability is what SAP calls supply optimization analysis. After a planning run in IBP, Joule can explain why demand wasn't fulfilled. Not just "here's a shortage" — it tells you what caused the gap, which constraints were binding, and suggests mitigation options. It can compare two different planning runs side by side and highlight what changed. That analysis used to take a senior planner half a day of digging through logs. Now it takes a question.

You can also trigger forecasting runs, manage S&OP scenarios, and execute order-based planning through conversational commands. The new harmonized planning area lets you move between strategic, tactical, and operational planning levels without switching apps. And the what-if scenario simulations — asking Joule "what happens to my supply plan if this supplier is two weeks late?" — are genuinely useful for the kind of conversations that happen in monthly S&OP reviews.

Is all of this perfect? No. The natural language interface still struggles with complex multi-level planning queries, and you need your BTP integration set up properly or none of it works. But if you're running IBP and you haven't enabled Joule yet, you're doing manual work that the system can already handle. That's the whole point of this newsletter — the tools are there, the adoption isn't.

What I'm still figuring out

I'll be honest — I don't have a clean answer for the change management side of all this.

I've been through enough S/4HANA go-lives to know what happens when you change someone's workflow without warning. People don't resist change because they're difficult. They resist because the last three "improvements" made their job harder and nobody asked them first.

Embedded AI has that same energy. Your demand planner opens IBP on Monday and there's a Joule button they didn't ask for. Your AP clerk gets invoice coding suggestions from Copilot that are right 80% of the time — which sounds great until you realize they now have to figure out which 20% is wrong instead of just coding everything themselves. That's actually a harder cognitive task, not an easier one.

I was talking to a supply chain lead last month who said something that stuck with me: "The problem isn't that the AI is bad. The problem is that nobody told my team whether they're supposed to trust it." And she's right. There's no playbook for this. When we rolled out a new Fiori app, there was a training plan, a test phase, super users. When Joule shows up in the next update, who's responsible for figuring out if the suggestions are good enough to act on?

I don't think the answer is to slow down adoption — the capabilities are too useful for that. But I do think most organisations are going to learn this the hard way: someone will follow an AI suggestion without checking, something will go wrong, and then suddenly there's a governance conversation that should have happened six months earlier.

If you're running one of these platforms, maybe the move is just to start small and be intentional about it. Pick one team, turn on one feature, watch what happens. Not as a "pilot" with a steering committee and a PowerPoint — just... see how real people react when the AI shows up in their workflow. That'll teach you more about your readiness than any assessment framework will.

The bottom line

The embedded vs. standalone debate isn't really a technology choice. It's a strategy choice about where you want to invest your limited time, budget, and attention.

The vendors are moving fast. SAP has dozens of agents and counting. Microsoft has Copilot in every product. Oracle and ServiceNow are shipping production-grade agentic applications. This train has left the station.

Your job isn't to build everything from scratch. It's to know what's already in your stack, activate it, identify the gaps, and build targeted standalone agents only where no vendor can help.

That's the bridge between AI capability and enterprise reality. And honestly, it's more pragmatic than exciting. But pragmatic is what actually ships.

Dive deeper: platform-by-platform guides

Everything mentioned above — the embedded capabilities, the activation steps, the honest limitations — is covered in detail in my practitioner guides. Each one walks through what's actually available today, how to enable it, and where the gaps are.

Executive overview: the strategic decisions that matter →