Digital Process Automation Software

Forrester's guidance in the current evaluation of this market contains a warning aimed squarely at how these products are now being sold.

AI agents will drive the next wave of innovation here, and buyers should look for vendors that integrate those agents with existing process automation strengths rather than vendors that simply demonstrate agent-building capability.

The distinction is doing real work. Building an agent is not difficult and every vendor can show one. Running a business process that takes six weeks, involves fourteen people, touches four systems of record, handles exceptions nobody anticipated, and survives infrastructure failures is a different engineering problem, and it is the one this category exists to solve.

Long-running is the defining property

Forrester's inclusion criteria for the evaluation state the requirement plainly: the vendor must handle complex long-running processes and support automation fabric principles including scalability, process complexity, and variability, with a track record of deploying large numbers of applications at scale, support for data and event-driven processes, governance, and native end-to-end orchestration and visibility.

Long-running is the phrase that separates this market from everything adjacent to it.

Consider what a genuinely long-running process requires. A mortgage application, an insurance claim, a supplier onboarding, a clinical trial enrolment, a regulatory submission. These run for weeks or months. They pause waiting for a human decision, a document, an external response, or a calendar date. They must survive server restarts, deployments, and version changes without losing their place. They accumulate state that has to be durable and auditable. And when something goes wrong at step forty, someone has to be able to see what happened at step three.

That is a state management problem before it is anything else, and it is where most alternatives fall down.

A workflow tool handles the happy path and struggles with exceptions. A robotic process automation script executes a sequence and has no concept of a process waiting three weeks for a response. An agent framework maintains context for the duration of a task and was not designed to persist a process across a quarterly release cycle.

Camunda's positioning in the current evaluation reflects this directly, with Forrester placing it for enterprises with professional developers automating complex processes that require best-in-class orchestration, and awarding maximum scores in orchestration and process standards.

Process standards matter here for an unglamorous reason. A process modelled in an open standard is portable, inspectable, and survives the vendor. One modelled in a proprietary canvas does not.

Inside The Forrester Wave: Digital Process Automation Software, Q3 2025

Authored by Craig Le Clair with four contributors, the evaluation scores vendors against thirteen current offering criteria and six strategy criteria, drawing on questionnaires, demonstrations, briefings, and customer interviews.

Pegasystems placed as a Leader with the highest scores among all evaluated vendors in both the current offering and strategy categories. Forrester credited its centre-out approach to journey orchestration, its architecture and AI decisioning for superior vision and roadmap scores, and noted its plan to differentiate through agent experience tooling and APIs supported by its blueprint design capability.

Appian also placed as a Leader. Forrester recorded customers praising rapid prototyping and flexible interface development tools that streamline application changes, alongside its data fabric and orchestration capabilities for integration challenges such as extracting data from core enterprise systems.

Newgen placed as a Strong Performer, evaluated across native process automation, robotic process automation, agentic document processing, and content management.

Camunda and Pipefy were also evaluated.

A companion Landscape published in Q2 2025 covers additional vendors Forrester considers notable but which fall outside the Wave's inclusion criteria.

The inclusion criteria are the real definition

Two of the qualifying requirements are worth reading as a description of what this category actually is.

The first is geographic presence, requiring the vendor to operate in at least two major geographies with each contributing at least fifteen percent of annual process automation revenue. That is a filter for enterprise viability rather than capability, and it excludes strong regional players.

The second is more interesting: the vendor must handle a complex ad hoc environment.

Ad hoc is the opposite of what process automation traditionally meant. A structured process is one where the path is known: apply, verify, approve, disburse. An ad hoc process is one where the path is discovered as the work proceeds, because the situation is unusual, the exception is novel, or the case requires judgement.

Real enterprise work is a mixture. Most cases follow the standard path and a meaningful minority do not, and the ones that do not consume most of the effort and produce most of the risk.

Traditional process automation handled the structured majority well and dumped everything else into a queue for a human. The requirement that a platform handle ad hoc complexity is the acknowledgement that the exception path is where the remaining value sits, and it is also exactly where agents become genuinely useful.

Where agents fit, and where they do not

Put the long-running requirement and the ad hoc requirement together and the appropriate role for agents becomes clearer than most vendor positioning makes it.

The process backbone should remain deterministic. It is auditable, it is testable, it survives restarts, and in regulated contexts it produces the evidence a supervisor expects. Nothing about AI improves those properties.

The exception branches are where agents earn their place. Reading an unstructured document that does not match any template. Deciding which of six paths a novel case should take. Gathering information from systems that were never integrated. Drafting a response to a situation nobody wrote a rule for.

Forrester's warning about integration rather than demonstration maps onto exactly this. A vendor showing an agent that can complete a task in isolation has shown you the easy part. The hard part is that agent operating inside a process that has been running for five weeks, with access to the case state, respecting the governance model, writing back in a way the audit trail can represent, and handing control cleanly to the deterministic path when it is done.

That handoff is the engineering problem, and it is where the difference between vendors will show up over the next two years.

The related question for a buyer is what happens when the agent gets it wrong at step twenty of a forty-step process. Reversibility, escalation, and the ability to reconstruct why the agent chose what it chose are properties of the platform rather than of the model.

The category next door

Forrester also maintains adaptive process orchestration as a separate category, mapped in a Q2 2026 Landscape covering thirty five vendors, and the vendor overlap between the two is substantial.

The distinction Forrester draws is about control flow. Adaptive process orchestration is defined around platforms combining deterministic control flows with nondeterministic ones, where an agent decides the sequence at runtime rather than following a modelled path.

Digital process automation is the established market, with a scored evaluation, mature vendors, and criteria built around long-running structured work.

For a buyer the practical implication is that vendors will cite whichever category flatters them, and the two evaluations are answering different questions. If your requirement is running complex regulated processes at scale with agents handling exceptions, the DPA evaluation is the relevant one. If your requirement is genuinely agent-led processes where the path is not modelled in advance, the newer category describes what you are looking for and does not yet score anyone.

What decides these implementations

The consistent failure mode in process automation has little to do with the platform.

Automating a process requires knowing what the process is, and organisations reliably discover that the documented process and the actual process differ. The documented version describes what should happen. The actual version includes the workaround for the system that has been broken since 2019, the exception someone approves informally, and the step that exists because a customer complained once.

Process mining addresses part of this by reconstructing the actual process from system logs, which is why several vendors in this market have acquired or built that capability.

The rest is a decision nobody wants to make: whether to automate the process as it exists or to fix it first. Automating as-is is faster, delivers value sooner, and encodes accumulated dysfunction into software where it becomes harder to change. Fixing first is correct and requires the political capital to tell several departments their local optimisations are going away.

Most organisations choose the first and regret it around month eighteen, which is the point at which the automated process needs its first significant change and everyone discovers what was encoded.

Forrester's emphasis on process standards, orchestration visibility, and governance across the criteria is partly a response to this. A process you can see, model in a portable standard, and change without a rebuild is one you can fix later. A process buried in proprietary configuration is one you will live with.

That is the durable buying criterion in this category, and it matters more than the agent capability that currently dominates the conversation.

Forrester does not endorse any vendor named here, and tier placement should not be read as a recommendation to buy.