For decades, some of the most valuable enterprise businesses in the world sat at the coordination layer. They corralled humans, data, and technology to manage critical business functions. This coordination layer, which included software platforms/systems of record, BPOs, management consultants, and case management vendors, represented hundreds of billions in annual spend. Despite the scale, this sector was slow, labor-intensive, and expensive. Now, it is being displaced.

As agentic AI matures, a new orchestration layer is emerging out of necessity. Static legacy systems and professional services firms that charge by the hour cannot effectively coordinate across dynamic, automated AI stacks, fragmented corporate data, and nuanced human judgment. The new layer will be flexible, AI-native, and built to handle an infinite range of inputs.

We see evidence of this shift in the actions of legacy players. Accenture tripled its AI revenue in 2025 and a large portion of their playbook is building “agent-to-agent orchestration frameworks.” They are essentially building systems to replace project-based coordination. In October 2025, Capgemini acquired WNS for $3.3B, explicitly framing the deal as a bet on “Agentic AI-powered Intelligent Operations” replacing traditional business process services. A global consulting giant paying a premium to reposition itself around agentic infrastructure tells you everything about where the pressure is mounting. 

In reality, this shift will likely be a combination of outright displacement and value capture: the AI orchestration layer eats the margin of these categories while delivering outcomes that previously required their headcount.

This Is Different From What Came Before

We have seen waves of enterprise automation before, but this moment is structurally different.

Prior automation technology, like robotic process automation (“RPA”), could execute rules and automate tasks. It was powerful in its domain but fundamentally brittle, because RPA breaks when rules change, and it has no capacity to reason around exceptions. Because older technology was rules-based, the surrounding orchestration still made sense as a human-centered function. You still needed people to handle judgment and decision-making. The machine executed; the human decided.

AI agents change that equation. They plan, adapt, and reason, dramatically increasing their potential impact and fundamentally shifting the accountability question. When RPA executes a rule, accountability is clear: a human set the rule. When an agent makes a decision, accountability requires a new infrastructure layer: who decided, on what basis, with what information?

This is the current gap between the promise of agentic AI and its production reality. Bain’s 2025 report on “Building the Foundation for Agentic AI” found that early AI ambitions significant efficiency improvements haven’t materialized, a failure they attribute directly to orchestration gaps. Systems that automate individual tasks but not multi-step workflows produce modest gains. The bottleneck isn’t the model. It’s the coordination layer around it.

The Emerging Orchestration Layer

The agentic orchestration layer is the control plane. It makes multi-agent work accountable, streamlines the human-to-agent handoff, and drives ROI for internal AI strategies. Based on what we are seeing in the market, we believe this emerging layer has four key components:

Decision Management provides the parameters, guidance, and governance for each decision — decision-makers, authorization, necessary evidence, rules, and model selection.

Escalation Logic determines how and when issues must be elevated and when a human needs to be in the loop.

Context and Memory is the organizational brain that holds prior interactions, documents, signals, and preferences, and understands how and why each applies across scenarios.

Audit Trail is the legible, timestamped record of every action and decision that establishes accountability and feeds regulatory and compliance requirements.

AI-native and AI-fluent enterprises will eventually have an orchestration layer spanning all four. This layer coordinates across AI agents, legacy software, and humans. We believe it will enable enterprises to realize the ROI they have been promised but have not yet seen.

Why This Must Be a “New” Layer

The most common counterargument is that the large frontier labs will simply build the orchestration layer themselves and own it. This misunderstands how enterprises actually adopt AI.

Frontier labs optimize within their own ecosystems. But task ownership inside most enterprises is deeply fragmented. Despite the dominance of the largest players, people seek out the best AI solution for each task or functional area. The evidence is concrete: Harvey has reached $190M ARR and an $11B valuation by going deep on legal workflows no horizontal provider would build. Abridge nearly doubled its valuation in four months (from $2.75B to $5.3B) by embedding itself in clinical documentation tied to EHR systems and HIPAA compliance.7 Glean, at $200M ARR and a $7.2B valuation, exists precisely because the enterprise SaaS stack is so fragmented that no single vendor can solve the knowledge retrieval problem. 

These are not niche businesses. They are large, fast-growing companies thriving because horizontal AI cannot match their domain depth. That fragmentation creates a natural pull toward multi-agent stacks that no single lab can or will coordinate.

For financial services and other regulated industries, this becomes even more acute. The desire to avoid single-vendor lock-in is a security and compliance imperative. When OpenAI experienced a global outage in June 2025, business-critical workflows froze across thousands of enterprises. That event accelerated the multi-provider conversation. According to KPMG’s AI Quarterly Pulse Survey, Q3 2025, 57% of organizations now favor a blended approach: building some capabilities in-house while buying others externally.11 Businesses need coordination that spans custom, in-house, and third-party agents, as well as legacy software, internal documents, and humans. That is precisely why this layer cannot live inside the large model providers.

Our Investment Perspective

As AI adoption grows, forward-thinking leaders are encountering what some are calling “automation sprawl.” The more disparate AI tools a business adopts, the faster the complexity compounds. As one technology leader put it in CIO Magazine: “We have dozens of RPA bots, various API integrations, and manual processes that handle exceptions. An orchestrating AI agent could manage this complexity by understanding the full landscape of available capabilities and routing work appropriately.”

We believe the need for an intelligent orchestration layer will only intensify and this is where durable value will be built. Models are commoditizing rapidly, and the application layer—however sophisticated—will increasingly depend on the orchestration layer to execute with reliability, accountability, and scale. What sits between the agents and the enterprise is harder to replicate than the agents themselves.

At Anthemis, we are actively looking to invest in Pre-Seed and Seed stage businesses building this critical coordination layer, particularly in financial services, where privacy and compliance requirements raise the stakes even further. While early-stage businesses rarely own every piece of the orchestration layer from day one, we focus on companies that immediately own one component and have a clear line of sight to a second.

With two components (e.g. decision management and audit trail or escalation logic and context and memory) businesses begin accumulating data and intelligence that compounds their value. As the data housed within such systems becomes essential to how enterprises operate, switching costs will skyrocket.

We are already investing here. Last year, we invested in Reload, which helps create shared context and memory by driving coordination across function-specific AI agents. Reload ensures your Claude Code is talking to your OpenClaw, your OpenClaw is talking to your Hermes, and all of your agents are communicating with the agents your teammates are using. This year, we invested in a stealth company which acts as an authorization layer that allows enterprises to establish and manage approvals before AI executes. This ensures that all actions are taken with verifiable human authority and creates auditable execution receipts.

Looking Forward

Today, companies operating at the AI orchestration layer look like decently compelling, standalone software businesses. In the future, we believe their roles will be essential.

They will serve as the operational infrastructure for their respective industries. Strategy, accountability, and execution logic will all be derived from or housed in this layer of the stack. The comparison isn’t Salesforce — it’s SWIFT, or the credit bureau stack. These are businesses that, once embedded in core operations, become too important to replace.

If you are building in this space, we would love to hear from you.