Why Agentic AI Needs Company-Level Orchestration

McKinsey just published a paper arguing that AI is making ERP "headless" — that agents will mediate between users and the system of record, executing transactions and orchestrating workflows on top of the traditional stack. The direction is right. The layer is wrong.

The piece sketches a five-tier architecture where the ERP backbone stays put and an "agentic operating model" sits above it, running decisions across domains. It's a clean diagram. It also quietly assumes the orchestration belongs to the ERP vendor — that the agents live inside the application, governed by the platform, drawing on one stack.

That's where the model breaks. Real businesses don't run one-app decisions. They run whole-business decisions.

What the McKinsey view gets right

The high-level moves in the paper are the right ones.

The ERP front-end as we know it is collapsing into conversation. Measurement is becoming architectural — you can't run autonomous decisions without telemetry on whether the decisions were any good. Data models are evolving from static schemas to dynamic ontologies. End users shift from executing transactions to setting intent and validating outcomes.

All of that is correct. We'd add only one thing: it's a description of how ERP vendors want the future to play out, not how it actually works once you run a business with it.

The hidden assumption

The McKinsey diagram puts the orchestration layer inside the ERP perimeter. Every decision the agents make is bounded by what the platform knows.

That works fine for processes that live entirely inside one system. Almost no important decision works that way.

Take a food manufacturer deciding whether to accept a large last-minute order. To answer that one question honestly, the business has to weigh:

• Current stock and shelf life sitting in the WMS
• Production capacity and changeover cost in the MES
• Available labour and overtime exposure in rostering and payroll
• Cold chain compliance for the delivery window in the IoT and compliance systems
• Margin after trade promotion deductions across the ERP and CRM
• Customer payment history and credit risk in finance

No ERP sees all of that. An ERP-bound agent sees the order in ERP, makes a call inside its world, and hands it forward. The decision is locally optimal and globally wrong.

This is the gap the McKinsey paper doesn't address. Software-level agents are smarter than they used to be, but they still inherit the boundary of the application they live in. Smart inside the sandbox. Blind outside it.

Software-level agents are components, not operators

Software-level agents are genuinely useful. An agent inside Xero that drafts journal entries is faster than a person. An agent inside Salesforce that ranks pipeline saves real time. None of those agents is running the business — they're running a function inside an application.

The mistake is treating them as the operating model. They aren't. They're tools that an operating model uses.

The actual operating model has to sit above every system in the stack. It has to read across them, reason across them, and act across them. That's not an ERP feature. It's a different layer entirely — and it doesn't belong to any single vendor.

Where the real orchestration layer sits

This is the part the McKinsey piece doesn't address, because it's outside the frame the vendors think in.

The company-level orchestration layer:

• Reads from every system. ERP, WMS, payroll, CRM, IoT, compliance, finance, marketing — every operational source feeding into one unified data foundation.
• Reasons across them in plain English. A manager asks a question, the system answers from combined live data. No dashboards, no exports, no waiting for someone to compile a spreadsheet.
• Acts through digital workers. Entities that hold a function across the whole business — compliance, operations, HR, marketing — not just inside one app.
• Treats software-level agents as plug-in components. The Xero agent, the Salesforce agent, the ERP-embedded copilot — they're useful tools a digital worker can call on, not the actors making the decision.

A compliance digital worker doesn't just live inside the labelling system. It watches recipes in the ERP, batch data in the MES, regulatory updates in its knowledge base, contract obligations in the legal hub, and marketing copy in the asset library — because compliance isn't a feature of one app. It's a property of the company.

The same is true for operations, HR, and finance. The decisions that actually matter span systems. So the layer that makes them has to span systems too.

What this means

ERP vendors will keep building inward. They'll embed agents in their stack, add copilots to their workflows, and call it the future of work. For processes that genuinely live inside one application, they'll be right. That's a real and useful slice.

Everything else — the decisions that actually decide whether the business is well run — will get made somewhere else. At a company-owned orchestration layer that operates above the stack, model-agnostic, vendor-neutral, and built around how the business actually thinks rather than how the software is structured.

This is the layer we build at ALTEQ. The Company Brain unifies the data. Digital workers operate the functions. Industry-trained models give them domain expertise. Software-level agents plug in as components, not as the operators.

The McKinsey paper is a useful map of the inside of the ERP. The harder, more valuable question is what gets built on the outside — where cross-system decisions actually happen, and where the real operating model lives.

That's the layer that matters. It's also the layer ERP vendors can't own.