Application intelligence vs service intelligence

Simon Leyland

Every enterprise software vendor is shipping AI agents. Salesforce has Agentforce. ServiceNow has AI Agents. Microsoft has Copilot Studio. AWS has Bedrock Agents. If you only watched the keynotes, you'd think this was sorted.

It isn't. And the reason tells you something about what comes next.

The bit nobody mentions in the keynotes

When a customer contacts you with a problem, their issue doesn't sit inside one application. It might start in your contact centre, need a logistics check in your ERP, trigger a refund in billing, and require a case in your CRM. The customer doesn't know about your tech stack. They want an answer.

Right now, the thing connecting those systems is a person. Someone alt-tabbing between six applications, working out what needs to happen, and pushing a process through to the end across system boundaries. That costs money, takes time, and doesn't scale. It's also where most of the friction in enterprise operations actually sits.

Agentic AI should fix this. The way it's being deployed right now mostly won't.

Three ways this could go

I'd rather map the range of outcomes than pretend I know which one wins. There are three credible scenarios, each with real logic and real problems.

Scenario 1: The SaaS vendors own it

This is the most visible right now. Every major SaaS platform is building agents that automate work inside their application. If your customer service runs entirely in Salesforce, Agentforce can handle cases within that ecosystem. For work that stays in one platform, it's useful.

The structural problem is that SaaS vendors have the wrong economic incentive. Their business model depends on being the system of record. They want you deeper inside their platform, not orchestrating across platforms. Every agent they build is designed to keep work inside their boundary.

Salesforce Agentforce is never going to hand off to ServiceNow mid-workflow. That's not a technical limitation. It's a commercial one.

So scenario 1 gives you what I'd call application intelligence. Each tool gets smarter on its own. But it can't give you service intelligence, which is the ability to reason and act across your full stack to deliver a complete outcome. You get islands of automation, with people still bridging between them. Better than today, but not what anyone's promising in the sales deck.

This is the most common approach right now because it's the lowest friction path. But common and durable aren't the same thing. I think it's a transitional state.

Scenario 2: The cloud providers become the orchestration layer

This has the strongest technical argument. AWS, Azure, and Google already provide identity, networking, security, and compute for most enterprise workloads. Adding agentic orchestration on top of that is architecturally logical. The AI models are there. The permissions frameworks exist. The logging and compliance tooling is mature.

An agent built on AWS using Bedrock can call APIs across multiple SaaS platforms, apply judgment using foundation models, and operate within security and governance boundaries that enterprises already trust. It isn't constrained to one vendor's ecosystem because it sits underneath all of them.

The problem is commercial, not technical. A CIO running multi-cloud isn't going to let AWS orchestrate workloads on Azure. And the SaaS vendors will resist being reduced to API endpoints that a hyperscaler's agent calls into.

But for enterprises committed to a primary cloud, this works. The infrastructure layer already handles security, compliance, and audit. Building the agentic layer there means no new vendor relationship and no new attack surface. You're extending something you've already bought.

Scenario 3: An independent agentic middleware category forms

History supports this one. Every time a cross-platform integration problem has appeared in enterprise software, an independent category has formed around it. EAI, ESB, iPaaS, RPA. The pattern repeats because no incumbent has the right incentive to solve the problem neutrally. A new type of company appears, sits between the platforms, and manages the complexity.

The difference this time is that the layer doesn't just move data between systems. It makes decisions. That's harder than what Mulesoft or Workato do, and it needs a much deeper relationship with the enterprise's business logic, compliance rules, and risk appetite.

Anyone trying to build this needs enterprise governance, security, audit trails, and broad connectivity from the start. High bar. But it's the same bar that previous generations of middleware cleared.

One nuance worth noting: this independent layer doesn't have to be built from scratch. It could be built on hyperscaler infrastructure, using cloud-native AI models and security primitives, assembled by specialists who understand specific enterprise domains. That puts scenario 3 closer to scenario 2 than it first looks. The question is who owns the customer relationship and who brings the domain knowledge.

What decides the outcome

It comes down to one question: how much of the value in agentic AI can be captured inside a single application?

If most useful agent work happens inside one system, routing tickets in ServiceNow or qualifying leads in Salesforce, then the SaaS vendors win and the independent layer never appears. Application intelligence is enough.

But if the value is in cross-platform work, resolving a customer issue that spans five systems, onboarding someone across HR, IT and finance, processing a claim that touches policy, medical and payments, then you need service intelligence. That's where scenario 2 or 3 takes over.

I think the cross-platform work is where most of the cost sits in enterprises today. Single-application automation is already happening through conventional means. The hard problem is the work that crosses boundaries. Which means the agentic layer that can operate across the full stack, with proper security and governance, is where the real market forms.

If you're making decisions now

The practical question isn't which vendor has the best demo. It's whether the work you're trying to automate lives in one application or crosses boundaries.

For single-platform work, the vendor-native agents are fine. Use them. They ship now and they'll deliver value quickly.

For cross-platform work, think about where the orchestration layer sits and who controls it. The choices you make now about cloud infrastructure and integration architecture will determine how easily you can deploy agentic AI across your full stack later.

The vendors selling you application intelligence will tell you it's enough. For the work that matters to your customers, it probably isn't.