Every large enterprise software vendor tells the same story this year. AI will unify your workflows, rationalise your process estate and bring order to fragmented operations. The adoption numbers point the other way.
A 2025 Zapier survey of more than 500 enterprise leaders found the average enterprise running 4.2 AI models in production, up from 1.9 a year earlier. Twenty-eight percent run more than 10 AI applications. Sixty-six percent plan to add more over the next 12 months. Three percent expect to run fewer.
Fragmentation is growing, and the rate is rising.
AI joins the systems already running
AI arrives as a capability inside existing systems. It replaces none of them.
At a regulated insurer or bank today it looks like this:
- Copilots inside case management tools, putting suggestions in front of adjusters and underwriters beside the existing decision workflow.
- LLM decision layers above DMN rule engines, adding to or overriding deterministic rules that compliance teams relied on for auditability.
- Agents calling external systems over REST, MCP and Arazzo, on execution paths the existing logging infrastructure never sees.
- Reasoning loops that span several agent calls, some synchronous, some queued, some triggered by runtime events that were never in the process model.
- Runtime variation, where one nominal workflow takes materially different paths depending on what a model concludes about an input.
The BPM engine, the RPA estate, the APIs and the people making decisions all stay. Each AI capability added beside them brings new execution paths, new failure modes, new governance requirements and new places nobody can observe.
The multiplier
Before AI, the coordination surface in a mature enterprise was BPM, RPA, APIs and human touchpoints. It was large, bounded and mostly deterministic.
AI adds to every layer of it: agent frameworks, LLM tool calling, reasoning loops, copilot integrations, several model providers, and the protocols that connect them. MCP, A2A, ACP and ANP compete as agent communication protocols, each covering a different tier. Every protocol is another surface to coordinate.
Reliability compounds across those surfaces. Three components at 90% reliability each give a system at about 73%. Ten sequential steps at 95% each give about 60% end to end. The MAST study of multi-agent systems documented failure rates between 41% and 86.7% across seven popular frameworks.
Projects fail on this. In June 2025 Gartner predicted that more than 40% of enterprise agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. S&P Global data showed that 42% of companies had abandoned most of their AI initiatives in 2025, up from 17% in 2024. MIT researchers, working from 52 executive interviews and 300 public deployments, found that 95% of enterprise AI pilots delivered no measurable P&L impact. They traced the failures to tools that could not integrate deeply with existing workflows or adapt to them.
Why the existing platforms fall short
BPM engines, agent frameworks and domain SaaS platforms are each sold as the answer. Each covers part of the problem.
BPM engines were designed for deterministic execution: the same input gives the same output every time. LLMs are probabilistic by design. Camunda acknowledged the mismatch when it added AI agents, noting that conventional BPMN models “cannot adapt in real-time based on AI reasoning”. Its response was a set of constraints on ad-hoc sub-processes, chain-of-thought capture, hard limits on model-call cost and durable event tracking. Those are workarounds inside one engine.
Agent frameworks such as LangGraph, CrewAI, AutoGen and Semantic Kernel optimise for reasoning. None of the major ones ships a complete production governance layer. None addresses data residency, encryption standards or compliance certification as part of the framework. ISACA notes that agents escalate their own permissions and spawn ephemeral identities with no formal provisioning or deprovisioning records. That breaks the separation-of-duties controls every regulated industry requires. The Composio AI Agent Report 2025 found that 12% of enterprise AI agent initiatives reach production.
Domain SaaS platforms such as ServiceNow, SAP and Salesforce embed AI product by product. AI workflows multiply inside and between them with no common governance layer underneath. ServiceNow built an AI Control Tower to govern the AI spreading across its own products. Appian named “agent sprawl” across fragmented SaaS apps and homegrown platforms as the defining enterprise risk of 2026, citing prompt injection, PII leakage and autonomous behaviour with no central management.
From design time to runtime
AI moves orchestration complexity from design time to runtime, and governance has to move with it.
Workflow design used to be an act of anticipation. You modelled the process, found the decision points, wired the integrations and deployed a static artifact. The model was the governance artifact. Reading the BPMN diagram told you how the process worked, which decision was made where, and who was accountable at each stage.
An agent breaks that. Given the same case on two different days, it can make different tool calls, reach different intermediate conclusions and produce different outputs. Kaptein et al. (2026) state it directly: “The same agent given the same task may follow different paths on different runs.” Design-time controls cannot hold that, because, in the same paper’s words, “Prompting merely reduces the probability of policy-violating paths; it provides no enforcement.”
Orchestration in this environment is a continuous control problem. Governance has to run at runtime and outside the agent’s execution loop. From there it can observe the agent, intervene, and record what happened as well as what was planned.
The exposure is already here
Zapier’s survey found 91% of AI tools in enterprise environments operating outside IT control, and 35% going through proper approval channels. CIO Magazine described a property and casualty insurer with more than a dozen separate AI proofs of concept across claims intake, underwriting and fraud detection. Six solved overlapping problems. None shared data infrastructure. Two abandoned systems were still running and still being billed.
Regulation reaches the same systems. Article 12 of the EU AI Act requires high-risk AI systems to record events automatically over their lifetime, so that what the system did can be traced. Under FINRA Rule 17a-4, an AI-generated customer communication is a broker-dealer record like any other. Deloitte’s 2026 State of AI report found that one company in five has a mature governance model for autonomous agents.
A decision you cannot explain is a decision you cannot defend. Most agent deployments today cannot explain their decisions in the terms a regulator asks for.
A new category
Forrester has defined two emerging market categories. The agent control plane is an enterprise layer that inventories, governs, orchestrates and assures heterogeneous AI agents across vendors and domains. Adaptive process orchestration coordinates non-deterministic AI control flows alongside deterministic ones. Forrester Principal Analyst Leslie Joseph states the core requirement: “A control plane can’t govern what it can’t observe.” In his framing, oversight lives outside the agent’s execution loop.
Gartner published its first Magic Quadrant for the category it abbreviates BOAT in October 2025. It projects that 70% of enterprises will consolidate onto a single orchestration platform by 2030, up from 5% today. Flowable co-founder Paul Holmes-Higgin wrote in early 2026 that “BPMN is still the strongest, most efficient way to orchestrate AI agents, maintain governance, and execute at enterprise scale.” DMN gives traceable, explainable decisions that an agent alone cannot guarantee to a compliance team. BPMN gives process owners an artifact a regulator can audit without reading code.
The analysts agree on the shape of the answer: a layer that governs every framework and engine in the estate. It is independent of any vendor, it observes what runs, and it enforces policy at runtime across a mixed execution environment.
Where Kirto fits
Kirto starts from the assumption that fragmentation is the permanent state of the enterprise. It has two products, and each holds one half of the answer above.
Flow runs the processes agents work inside. It executes BPMN, CMMN and DMN as the standards specify, and Arazzo for multi-step API sequences. An AI agent is a step in the process, run by its own engine beside the BPMN and CMMN engines. The process says what the agent is for, which tools it may reach, and where a person has to approve before the run continues. Tools are declared apart from the process logic, through MCP, OpenAPI and Arazzo connectors, so a change of tool or model provider leaves the process definition as it was. Every step goes into the case’s record: each branch condition with its expression and result, each command dispatched, each agent turn with the raw model response, each human approval. A run can be replayed from that record and checked against the process definition that authorised it.
Gate watches agents from outside their loop. It sits on the request path for model calls, MCP tool calls and A2A agent calls, whoever built the agent. A key decides who may call what. Detectors inspect each call before it leaves. Every decision goes into a hash-chained, signed audit trail that a third party can verify with kirto-gate audit verify.
Neither product makes an AI estate simpler. Together they make it governable. Models improve, frameworks change and regulation tightens, and the process definitions and the records stay yours through all of it. The platform overview shows how the two fit together, and Part 2 asks why the incumbents cannot build this layer themselves.