Business Architecture as the control plane for enterprise AI

Moving from AI pilots to enterprise-wide impact requires more than technology alone – it demands clear structure, governance and design at scale. Paul Lalovich and Tesha Teshanovich of Agile Dynamics outline why enterprise design must sit at the centre of any organization’s ambition to scale AI successfully.

Enterprise AI programs are no longer constrained primarily by model quality or tooling availability. The binding constraint is organizational: fragmented use cases, duplicated data work, inconsistent controls, and operating models that cannot absorb agentic workflows. These are not “AI problems” in the narrow sense. They are enterprise design problems – and they require an architecture answer.

A useful way to state the value reality is the 10/20/70 pattern: the minority of effort sits in algorithms and platforms, while the majority sits in changes to people, processes, and the operating model.

The ‘Business Architecture Playbook’ from Agile Dynamics operationalizes that logic for AI-era transformation and advances a sharper internal diagnosis: even where the technology is available, most enterprises remain structurally organized in ways that suppress compounding AI value – a pattern the Playbook summarizes using an illustrative distribution of operating-model maturity and an explicit emphasis on the people/process component of value.

The practical implication is the same: promising pilots frequently fail to compound because the enterprise lacks a repeatable mechanism to redesign work, govern it at runtime, and scale patterns across value streams rather than reinvent them team by team.

Closing the organizational gap

Business Architecture is the discipline that closes this gap. Business Architecture provides holistic, multidimensional views of capabilities, end-to-end value delivery, information, and organizational structure – and the relationships among these views and strategy, products, policies, initiatives, and stakeholders.

That relationship-model is the point. AI does not fail at scale because enterprises lack intelligence; it fails because enterprises cannot integrate intelligence into execution without fragmenting accountability.

Our Business Architecture Playbook formalizes this approach into a practitioner methodology spanning enterprise blueprinting, AI-era operating design, and governance. The framework described below draws on that body of work, while aligning with widely used Business Architecture reference models and operating-model research.

From hierarchies to agentic networks

When AI moves from assist to execute, the unit of design shifts from isolated tasks to end-to-end value creation. The emerging operating pattern is the agentic network: humans and AI agents collaborating in outcome-oriented teams with explicit decision rights and embedded controls. This is an operating requirement, not a cultural preference. Agents change throughput and decision cadence; they stress approval gates, handoffs, and ownership boundaries.

Organizations that cannot model where decisions sit, what information they require, and how exceptions are handled will either over-control – losing speed – or under-control – creating operational and compliance exposure.

Business Architecture provides the modeling discipline to redesign these mechanisms without losing traceability. The goal is not to “document the business.” It is to make work legible enough that redesign can be executed repeatedly – across domains, portfolios, and value streams – without turning every AI deployment into a bespoke governance negotiation.

Domains and minimum blueprints

For AI implementation, Business Architecture is most useful as an integrated view across the domains that determine execution coherence: vision and strategy, capabilities, value streams, information, organization, policies and rules, stakeholders, products and services, initiatives, metrics, and decisions/events. The objective is not exhaustive documentation. It is to make the enterprise coherent to itself – a shared language for prioritization and governance that prevents AI from amplifying interpretations into duplicated work, inconsistent controls, and pilots that cannot compound.

In practice, foundational views carry most of the load because they translate coherence into clarity. Capability maps and scorecards anchor investment in what the business must be able to do. Value stream views force redesign of end-to-end outcomes rather than automation of local steps. Information views define concepts, sources, ownership, access boundaries, and lineage.

Operating model views specify roles, RACI, decision rights, and escalation so accountability survives hybrid execution. Strategic transparency follows: dependencies surface, overlaps are exposed, and portfolio choices become architecture-based trade-offs rather than executive opinion. Over time, a stable capability and value-delivery language turns AI from use-case accumulation into an operating design program.

A 4-layer design model

AI programs accelerate when execution focuses on four architectural layers that directly shape delivery: capabilities, value streams, information, and organization. Together, they define what to change, where to change it, what it needs to run, and who owns outcomes.

Capabilities: invest in stable outcomes, not shifting projects
Capability-based planning corrects the common bias toward tools (“we need copilots”) or functions (“automate finance”) as the unit of prioritization. Capabilities are relatively stable; processes and systems are not. Mapping AI opportunities to capabilities makes portfolios resilient to reorganizations and platform cycles.

Operationally, the method is simple: assess capability maturity and constraints – decision latency, exception volume, data quality, cost-to-serve, control intensity – then prioritize where AI changes the constraint system: prediction for decision support, generation for knowledge work throughput, automation for workflow execution, and orchestration for cross-step agentic coordination.

Value streams: redesign performance, not individual activities
AI value compounds when it reduces friction across the value stream, not when it optimizes one activity in isolation. Order-to-cash, issue-to-resolution, claim-to-settlement, recruit-to-onboard – end-to-end flows provide the outside-in view needed to redesign for outcomes. The critical design move is to treat the stream as a sequence of decisions and exceptions: define where agentic execution is appropriate, specify where human oversight is mandatory, and make escalation thresholds explicit.

The future state becomes a hybrid workflow with measurable targets for cycle time, rework, error rates, and experience outcomes – paired with clear control boundaries that preserve accountability under faster execution.

Information: build defensible proprietary data assets
AI is data-intensive, but not all data creates advantage. Durable performance increasingly comes from proprietary information assets captured through operations and relationships, combined with governance strong enough to reuse them safely. Information mapping turns this into concrete decisions: core information concepts, authoritative sources, ownership, access controls, and lineage. High-value datasets are treated as reusable data products with enforceable quality thresholds and change management.

Knowledge-intensive workflows benefit from retrieval architectures that keep models grounded in controlled, current information rather than ad hoc prompts. This is how “trustworthy AI” is built operationally: not as aspiration, but as an information contract embedded into execution.

Organization: scale through fusion teams and clear decision rights
AI delivery breaks when it is owned by a single function. The scalable pattern is a fusion team: cross-functional units combining domain expertise, product delivery, data engineering, model operations, and risk oversight. The architecture work here is precision, not rhetoric – role design, RACI, decision-rights maps, escalation paths, and operational ownership for what agents do in production.

As execution becomes hybrid, jobs shift from task execution to supervision, output curation, and exception handling, while accountability for outcomes must remain unambiguous.

Governance that keeps up

Traditional governance assumes stable processes and infrequent change. AI introduces continuous change: models drift, data shifts, prompts evolve, workflows are reconfigured. Governance must therefore move from periodic review to embedded controls.

Business Architecture contributes by defining policies, decision boundaries, and accountability in artifacts that can be operationalized directly in platforms and workflows – policy-to-control mapping, monitoring tied to business outcomes and risk indicators, explicit human accountability for automated decisions, and change processes proportionate to risk.

In practice, “control plane” stops being metaphor: it becomes the explicit layer where policy, permissions, and decision rights are designed so that runtime enforcement is possible.

Conclusion

AI implementation is, at its core, an operating model and architecture problem. Business Architecture makes the enterprise legible enough to redesign work safely, govern it continuously, and replicate design patterns across value streams. Organizations that treat Business Architecture as the control plane move from isolated wins to compounding improvements under real-world constraints.