A recent discussion with a colleague about AI-generated BPMN models got me thinking about a broader question for organisations adopting AI within Business Analysis and process modelling.
Much of the current conversation around AI understandably focuses on productivity. We see demonstrations of tools generating requirements, documentation and process models in a fraction of the time these activities may previously have taken, and there is genuine value in that capability. For organisations under pressure to deliver transformation more quickly and make better use of scarce capability, the opportunity is significant.
However, when process models are being used to inform requirements, automation, operating models, controls or investment decisions, speed cannot be the only measure that matters.
The more important question may be:
If AI can generate process models faster, what gives us confidence that those models are accurate enough to support the decisions we make from them?
AI does not create the modelling risk. It can accelerate it
At first glance, BPMN diagrams can appear deceptively straightforward. Activities are connected through sequence flows, gateways represent points of decision or divergence, and events describe how processes begin, change or end. Experienced Business Analysts and Process Modellers know, however, that the meaning of a BPMN model sits in the detail.
The choice between an Exclusive Gateway and an Inclusive Gateway changes the behaviour being represented. Message Flows and Sequence Flows serve different purposes. Boundary events, subprocesses, exception handling and escalation paths can fundamentally alter the meaning of the model. These are not simply presentation choices; they affect how the organisation understands the process and how that understanding may later be translated into requirements, controls or automation.
Importantly, this is not a problem created by AI. Human modellers can also misunderstand the domain, use an inappropriate BPMN construct or make assumptions that result in an inaccurate representation of the process.
Research into BPMN models used in industry reinforces that point. In a study of 585 BPMN 2.0 models from six organisations, around 99% were free from syntactical errors, yet significant structural and quality issues remained. Twenty-two per cent contained deadlocks, 42% contained multi-merges, and among models containing subprocesses, 86% showed inconsistencies between roles in the subprocess and those in the corresponding main process.
The important distinction is that a model can be technically valid and still contain issues that affect how reliably it represents the process.
AI introduces a different dimension to that existing risk: speed and scale. A plausible-looking model can now be produced very quickly, which means incomplete assumptions, incorrect source information or weak modelling choices can also be propagated very quickly if appropriate review is not in place.
Professional-looking does not necessarily mean decision-ready
One of the more interesting characteristics of generative AI is how convincing its outputs can appear. A generated process model may look structured, complete and professional, particularly to stakeholders who are not experienced in BPMN.
That creates a subtle governance risk because visual credibility can easily be mistaken for modelling quality.
The same industry research found that BPMN quality problems frequently sat beyond basic syntax. Almost half of the models examined exceeded recommended diagram size, while between 40% and 47% of labels followed patterns the researchers considered potentially ambiguous. The authors also identified issues around message flow, model decomposition and consistent terminology.
For organisations, the significance is not simply whether a diagram contains a modelling error. Process models often become inputs into other activities. They may influence solution requirements, automation logic, control design, operating procedures and transformation decisions. If the underlying model is incomplete or misleading, the consequences can move downstream into rework, repeated analysis, implementation delay and decisions made from an inaccurate understanding of the process.
The commercial impact is not easy to isolate because organisations rarely measure the cost of process modelling separately from the broader transformation activity. Research into process modelling effort has itself noted the lack of an established, validated approach for estimating these costs, while identifying factors such as process complexity, documentation quality, required correctness, modeller capability and BPM experience as important drivers of effort.
That research also reinforces a principle familiar to anyone involved in delivery: errors identified later in implementation or execution are more expensive to correct than errors avoided earlier in the modelling process.
From a transformation perspective, that is where the productivity conversation becomes more interesting. Generating an initial process model faster may save time, but some of that value can quickly disappear if the organisation then needs to correct downstream decisions based on an inaccurate or incomplete model.
The value of process modelling is not the diagram
This brings me back to something I have observed throughout my career.
The value of process modelling has never really been the diagram itself. The value comes from the work required to understand what the diagram should represent.
It comes from bringing the right stakeholders together, uncovering exceptions and workarounds, challenging assumptions and resolving different interpretations of how work is actually performed. It comes from asking why something happens, not simply what happens next, and from building enough shared understanding that the organisation can make decisions with greater confidence.
The model is the artefact that captures that understanding.
That distinction becomes increasingly important as AI begins to assist with process modelling. AI may help structure, document and visualise information more efficiently, and that should be welcomed where it creates genuine value. What it does not remove is the need to determine whether the resulting model reflects operational reality closely enough for the purpose for which it will be used.
For a transformation leader, that purpose matters. A draft model used to facilitate an exploratory workshop carries a very different risk profile from one being used to define automation logic, inform control design or support a significant investment decision.
Faster modelling needs stronger validation, not less
If AI reduces the effort required to produce an initial BPMN model, the opportunity should not simply be to produce more diagrams in less time. Some of that saved effort can instead be redirected towards validation, stakeholder engagement and testing whether the model genuinely improves the organisation’s understanding of the process.
Before relying on an AI-generated BPMN model, organisations should be able to establish that:
- The model has been validated against operational reality, including with people who perform or own the process.
- Gateway logic, joins and splits are explicit and unambiguous, particularly where incorrect logic could create unintended paths or deadlocks.
- Subprocesses and roles remain consistent across the model, rather than introducing contradictions between different levels of the process.
- Message flows and other BPMN constructs are being used appropriately, particularly where interactions cross organisational or system boundaries.
- The model is readable and sufficiently well decomposed, rather than technically valid but too large or complex to support shared understanding.
- Terminology and labels are consistent, particularly where the model will be reused across teams or become part of a broader process architecture.
These checks are not intended to introduce another layer of bureaucracy. They recognise that syntactical correctness is only one dimension of process quality, and that the more important question is whether the model is clear, consistent and accurate enough to support the decision that follows.
The recommendations align closely with findings from the industry BPMN study, which highlighted implicit splits and joins, model decomposition, message flow usage, terminology and linguistic consistency as recurring quality concerns.
Governance needs to keep pace with generation
As AI reduces the effort required to generate process models, organisations need to be equally deliberate about who remains accountable for determining whether those models are fit for use.
The appropriate level of review should be proportionate to the consequence of the decision the model is supporting. A model created as a starting point for discussion may require relatively light validation, while one informing automation, controls, solution design or significant transformation investment warrants a much higher level of assurance.
For those higher-consequence uses, organisations should retain sufficient BPMN and domain capability to challenge the output, validate assumptions and confirm that the model accurately represents operational reality. This does not mean every AI-generated diagram requires a new approval gate. It means accountability for model quality needs to remain clear even as generation becomes faster and easier.
That may ultimately be the more important governance question.
AI changes the economics of process modelling by reducing the effort required to generate an initial artefact, but it does not remove the effort required for understanding, validation or governance. If organisations reinvest some of the time saved through AI into stronger validation and better process understanding, speed can contribute to quality. If they do not, AI may simply allow existing modelling weaknesses to scale more quickly.
Perhaps the measure of success should therefore be less about how quickly AI can create a BPMN diagram and more about whether the resulting model gives the organisation greater confidence in the decisions it is being used to support.
About the Author
Delene is an experienced Business Analyst with over ten years’ experience supporting our clients in transformation, delivery, and operational improvement initiatives. She is passionate about helping organisations navigate change in practical, people-centred ways, combining strong business analysis with empathy, clarity, and a focus on outcomes.
Sources Used
Leopold, Mendling and Günther’s study of 585 industry BPMN models provided the evidence on syntactical correctness, deadlocks, multi-merges, subprocess inconsistencies, model size, labelling and their five recommendations for improved BPMN modelling.
Baumann, Milutinovic and Roller’s work on process-modelling cost estimation informed the discussion of modelling effort, cost drivers, model complexity, correctness, modeller capability and the higher cost of errors found later in execution. The authors also note that their proposed cost-estimation approach was not yet validated on real-world project data, so I have deliberately avoided presenting it as evidence of a precise financial impact.
Working with Business Aspect
Business Aspect has worked with hundreds of clients, helping them enrich their business processes. Our team of highly experienced business analysts can help you create new value and reduce rework by bringing experience and thought leadership to your specific business challenges. If you need to optimise your processes and improve business efficiency, contact Business Aspect to discuss your needs.