Forrester opened a new category in 2025 and mapped it in its first landscape in 2026: adaptive process orchestration, orchestration that combines deterministic flows with AI agents to meet business objectives. The category is young, but the word worth reading slowly is already there. Not agents. Adaptive.
Because “adaptive” quietly changes what a process specification is.
The sequence was the specification
For twenty years, the contract of process automation was simple: the diagram is the spec. You modelled the route — activities, gateways, exceptions — and the engine’s job was to follow it. Conformance meant the path was walked as drawn. Audit meant proving it.
That contract works exactly as long as the route is knowable in advance. For most critical processes, most of the time, it is — which is why executable BPMN remains the backbone of anything that must survive an audit.
Agents break the contract
An AI agent, by definition, decides its next step within limits. You cannot draw its route in advance; if you could, you would not need an agent. So the moment an agent enters a process, the old contract fails: if the sequence is the specification, an agent is unspecifiable.
Organizations facing this tend to reach for one of two bad answers. Ban agents from anything that matters — which concedes the efficiency to whoever doesn’t. Or let them run and hope — which no regulated organization can say out loud.
The objective is the specification
The third answer is the one the new category’s name is pointing at: define the process by the outcome it must reach and the constraints it must respect, so the route can vary while the accountability cannot.
For that to be more than a slogan, the objective has to be a formal object, not a paragraph in the documentation. In our methodology this discipline is called objective-driven process design, and the objective of a process is specified as a complete structure: the stakeholders it serves; its outcomes expressed as business states; one criterion per outcome with a target, a tolerance, a direction, and a declared source — the entity and field, or the indicator, that will feed the measurement; its constraints, typed; and its risks. The methodology is public and downloadable, so none of this has to be taken on faith.
An objective specified that way is machine-checkable. Which yields the property that matters:
- Conformance is consulted, not opined. The engine can verify, on demand, whether a process satisfies its declared constraints — the verifier is a query, not a workshop.
- The route can vary; the objective cannot. A fixed sub-flow here, a person there, an agent elsewhere: all admissible ways of meeting the same declared outcome, under the same declared constraints.
- Every variation leaves evidence. Who or what took the step, on what data, under which rule — the same audit trail whether the executor was a human or a machine.
On agents specifically, the methodology already defines the agent as an executor class — with bounded scope, escalation criteria, and evidence per intervention. In Dexon BPM, governed AI agents already execute process activities, built into exactly this framework: the agent takes a step, the objective still decides whether the process met its outcome.
Three questions for any vendor
If you are evaluating platforms against this emerging vocabulary — Forrester’s adaptive process orchestration, Gartner’s BOAT, agentic process automation — the names matter less than three questions:
- Where does the objective live? As a formal object the engine checks — with targets, tolerances and declared sources — or as a paragraph in the project documentation?
- Can conformance be consulted? If the answer involves scheduling a meeting, the answer is no.
- What does a variation leave behind? When the route differs from the happy path — because a person escalated or an agent decided — what exactly can you show an auditor?
A platform that answers all three has the thing the category’s name promises: adaptation with accountability. That is what process orchestration on one engine means on this site — and the objective, not the sequence, is what makes it safe to say.