AI Orchestration: Tacit Knowledge Extraction for Multi-Agent Workflows

The Enterprise Orchestration Layer: Beyond Co-Pilots to Coordinated Digital Teams

This conversation reveals a profound shift in how work will be done within large organizations, moving beyond isolated AI tools to a coordinated system of agents that plan, analyze, and execute across departments. The non-obvious implication is that the true value of AI in the enterprise won't be in automating individual tasks or providing simple information retrieval, but in orchestrating complex, interdependent workflows. This is crucial for leaders in finance, operations, and technology who need to understand how to architect these new systems, manage the flow of critical context, and build commercially defensible platforms. Those who grasp this concept gain a significant advantage in navigating the accelerating pace of digital transformation and unlocking new levels of efficiency and revenue.

The Unseen Complexity of Coordinated AI: From Silos to Synchronized Action

The prevailing narrative around AI in the enterprise often focuses on individual tools or "co-pilots" designed to assist with specific tasks. However, this discussion highlights a more sophisticated evolution: the emergence of an "enterprise orchestration layer" built on multi-agent systems. Seema Amble articulates this shift, emphasizing that in 2026, enterprises will move from isolated AI experiments to coordinated digital teams. This isn't just about automating tasks; it's about agents managing complex, interdependent workflows like planning, analyzing, and executing across departments. The immediate benefit--increased efficiency--is clear, but the hidden consequence lies in the profound need for context extraction.

Organizations, particularly the Fortune 500, sit on vast reservoirs of siloed data and institutional knowledge, much of which is tacit, residing in people's heads or scattered across disparate systems. Amble stresses that turning this into "usable operational context" is the gating factor. This involves collecting documentation, observing human actions (literally watching how people click through software or conduct calls), and then piecing it together as shared context for these agents. Without this, agents operate autonomously, potentially optimizing for narrow metrics like efficiency, leading to misaligned outcomes for the business.

"The fortune 500 will feel this shift most acutely. They sit on the deepest reservoirs of siloed data, institutional knowledge, and operational complexity, much of which sits in people's brains."

-- Seema Amble

The potential for cascading failures when multiple agents work autonomously is a valid concern. However, the argument presented is that this mirrors existing human organizational failures. The key to mitigating this risk lies in establishing clear "eval functions" and KPIs for each agent, much like humans are measured. Feedback loops, where an agent's performance is measured against overall business objectives (e.g., sales agents closing deals versus customer support agents dealing with churn), allow for the adjustment of objective functions. This creates a system where agents are not just performing tasks but are synchronized towards overarching organizational goals. The immediate discomfort of re-architecting workflows and extracting tacit knowledge yields a significant long-term advantage: smoother operations and faster implementation of new systems, as the context layer makes changes much quicker than traditional, multi-year IT projects.

The Legacy Reckoning: Unlocking Speed and Margin Through Unified Data

Angela Strange pinpoints financial services and insurance as an industry where this orchestration layer becomes unavoidable, driven by the imperative to replace legacy systems. The risk of not replacing these aging mainframes and siloed data structures is finally outweighing the risk of change. This isn't just about adding AI; it's about unifying data from legacy cores, external systems, and unstructured sources into a new "system of record." This unification enables organizations to scale and fully leverage AI.

The downstream effects of this shift are threefold. First, workflows become parallelized. Instead of bouncing between screens and manually cutting-and-pasting data, tasks like loan underwriting can be performed concurrently, with agents handling mundane aspects. This immediate acceleration bypasses the bottlenecks of traditional, sequential processes. Second, data categories expand dramatically. Customer data from onboarding, KYC, transaction monitoring, and even customer service interactions can be integrated into a single risk platform, improving fraud detection and compliance. This creates a richer, more holistic view than previously possible.

"The next generation of infrastructure doesn't just add AI; they unify the data from legacy cores from external systems from unstructured data into a new system of record enabling FIs not only to scale but to take full advantage of AI."

-- Angela Strange

Third, and most exciting for builders, is the potential for 10x growth. Software can consume labor that humans don't want to do or that companies couldn't hire for fast enough. The competitive advantage here is stark: companies that fix their "plumbing" by unifying data and enabling parallelized, AI-driven workflows will be far more competitive. While cost reduction is a benefit, the true payoff is in revenue generation and margin expansion. Banks and insurance companies that have successfully transitioned have seen areas of their business move from 5% to 50% margins. This delayed payoff, requiring significant upfront investment in system replacement, creates a durable competitive moat that competitors will take years to replicate. The conventional wisdom of incremental upgrades fails here; only a fundamental re-architecture unlocks this level of advantage.

Multiplayer AI: The Command Center for Human-Agent Collaboration

Alex Immerman shifts the focus to the product implications of the enterprise orchestration layer, describing how vertical AI is moving into "multiplayer mode." This means collaboration not just between humans, but between multiple humans and multiple AI agents within a workflow. This evolution moves beyond simple information retrieval and reasoning to a more interactive, collaborative experience. Vertical software--applications built for specific industries like property management or legal services--is the ideal environment for this, due to deep integrations, proprietary data, and specialized interfaces that horizontal AI struggles to replicate.

The progression Immerman outlines is from information retrieval (summarizing documents) to reasoning (analyzing financial statements, reconciling trial balances) and now, to collaboration. In 2026, accomplishing a full job requires multi-human and multi-agent collaboration. This increases the value of these platforms and, crucially, raises switching costs, enhancing defensibility. The key to enabling this multiplayer mode is building trust through "AI operating agreements." These define when an agent can act autonomously and when it needs to flag an issue for human review.

"Software won't be just another chat interface, but you can think of it as a command center. There is a list of activities that are being negotiated on that agents have full ability to go and act and then there's a separate section, the flags, where humans need to engage and take action."

-- Alex Immerman

This creates a "command center" interface. Agents can execute tasks, while a separate section highlights issues requiring human engagement. For instance, in an M&A transaction, agents might negotiate a high-level agreement within set parameters, but outstanding questions like working capital arrangements would be flagged for human decision-making. This approach moves beyond simple automation to a model where humans review and approve, rather than perform, complex tasks. The immediate benefit is efficiency and accuracy, but the delayed payoff is the creation of deeply embedded platforms with high switching costs, built on trust and a clear division of labor between humans and AI, a structure that is inherently defensible.

Commercial Defensibility: Reinforcing Business Models for Lasting Advantage

David Haber brings the crucial commercial filter to the discussion, arguing that the AI systems that will win and persist are those where AI reinforces the business model by driving revenue and outcomes, not just cost reduction. While automating work and reducing costs are attractive, companies are willing to adopt AI without limit when it actively drives revenue. This creates a powerful market pull that vastly exceeds that of cost-saving-only solutions.

Haber uses the example of Eve, an AI workspace for plaintiff law firms. These attorneys operate on a contingency basis, meaning they only get paid if they win. Eve's AI doesn't erode their billable hours; instead, it reinforces their business model by enabling them to take on more clients and achieve better outcomes. This directly drives revenue, making the market pull for such a solution tremendous. Similarly, Salient applies voice agents in loan servicing not just for cost reduction but to improve collection rates, reinforcing the lender's business model.

The compounding competitive advantage in these AI applications resides in "workflow ownership" and "proprietary outcomes data." Companies like Eve embed themselves deeply within their customer's workflow, from intake to outcome. This daily engagement creates a strong source of defensibility. More importantly, by processing cases from intake to outcome, they generate a unique data asset--outcome data--that is not publicly available. This data is then used to improve intake, triage labor, and inform demand letters, making the platform smarter and more powerful with every case processed.

"I think in instances where AI is actually reinforcing the business model in driving revenue, there's really no limit to the amount that customers may want to adopt that technology."

-- David Haber

This creates a virtuous cycle: more data leads to a smarter platform, which leads to better outcomes for clients, which reinforces the business model and drives more adoption. The immediate benefit is improved client outcomes and revenue generation. The delayed payoff is a powerful, self-reinforcing moat built on proprietary data and deep workflow integration, a position that conventional AI tools, focused solely on cost reduction or isolated tasks, cannot replicate. This is where true, durable competitive advantage is built.

Key Action Items: Architecting the Orchestration Layer

• Immediate Action (Next Quarter):

  • Map Existing Tacit Knowledge: Identify critical institutional knowledge currently held by individuals within key departments (e.g., sales, support, operations). Document current workflows and observe human actions.
  • Audit Legacy Systems: For organizations in heavily regulated industries like finance, conduct a thorough audit of legacy systems to assess their readiness for data unification and AI integration.
  • Pilot Context Extraction Tools: Experiment with tools or methodologies that can extract and operationalize tacit knowledge into usable context for potential AI agents.
  • Define Agent Eval Functions: For any existing or planned AI initiatives, clearly define agent-specific evaluation functions and KPIs that align with broader business objectives, not just isolated task efficiency.

• Longer-Term Investments (6-18 Months):

  • Develop Multi-Agent Coordination Strategy: Design a roadmap for transitioning from isolated AI tools to coordinated multi-agent systems, focusing on interoperability and communication protocols between agents.
  • Invest in Data Unification Platforms: Prioritize investments in platforms that can unify siloed data from legacy systems and external sources, creating a robust system of record. This is crucial for enabling parallelized workflows.
  • Build "Command Center" Interfaces: Design and implement user interfaces that facilitate collaboration between humans and AI agents, clearly separating autonomous agent actions from human review and decision-making points.
  • Focus on Outcome-Driven AI: Shift AI investment strategy to prioritize applications that demonstrably reinforce business models by driving revenue and measurable outcomes, rather than solely focusing on cost reduction.
  • Cultivate Proprietary Data Assets: Develop strategies to generate and leverage unique, proprietary data from end-to-end workflows, which can inform smarter AI decision-making and create defensible competitive advantages.