The arrival of capable large language models in the past year has generated a particular category of confusion among enterprise software buyers and investors alike: the assumption that LLMs and workflow automation are addressing the same problem through different means, and that one is about to make the other obsolete. This framing is wrong, but it is wrong in a specific way that is worth unpacking rather than dismissing. LLMs and workflow automation are solving different problems, operating at different levels of the enterprise software stack, and interact in ways that create opportunity rather than competition. Understanding the distinction matters practically — both for building products that can survive the LLM wave and for evaluating investments in this space.
A workflow automation system encodes process logic: the sequence of steps that constitute a business process, the conditions that determine which branch to take at a decision point, the integrations that connect the automation to the systems of record that feed and receive its outputs. The intelligence in a workflow automation system is primarily deterministic: if document type equals purchase order and supplier code matches validated list then route to accounts payable queue. Where ML models appear in workflow automation, they are typically doing classification or extraction tasks — taking an unstructured input and producing a structured output that the workflow logic can then act on. The workflow automation layer handles process orchestration; the ML components handle perception tasks at the edges of the workflow.
A large language model, by contrast, is fundamentally a next-token prediction system that has generalised remarkably well across language understanding and generation tasks. What LLMs are genuinely good at, in an enterprise context: synthesising unstructured text into structured summaries, generating first drafts of documents from structured inputs, answering questions about a body of text, and — with appropriate prompting — reasoning through ambiguous situations in natural language. What they are not good at, and are unlikely to become good at without significant additional engineering: reliably executing multi-step processes with defined error handling, maintaining state across complex workflows with branching logic and external system integrations, or providing the kind of deterministic audit trail that regulated enterprise processes require. These are architectural limitations, not capability gaps that will close with more training data or a larger context window.
The practical implication for enterprise software architecture: LLMs belong at the perception and reasoning edges of workflow systems — they are the components that extract meaning from unstructured inputs and generate structured outputs for downstream processing, or that handle the conversational interface layer through which users interact with an automation. The workflow orchestration layer, with its process definitions, decision logic, exception handling, and audit trail, remains the domain of deterministic systems. Products that are trying to replace the entire workflow layer with an LLM — "just describe the process in natural language and the LLM will handle it" — are building on an architectural premise that does not hold for the kinds of business-critical, regulated, multi-step processes that enterprise workflow automation actually addresses.
For our portfolio, this distinction has shaped how we think about LLM integration specifically. Companies like Workist, processing document flows in procurement, are already integrating LLM-based extraction capabilities where they improve accuracy on complex document layouts — but the workflow orchestration layer, the three-way match logic, and the ERP integration remain conventional software. Bryter, with no-code decision automation, is exploring LLM-assisted workflow design, where the LLM helps a non-technical user translate a process description into a workflow definition — but the resulting workflow definition is then executed deterministically. The intelligence at the edge, the determinism in the process: that is the architecture that actually works for enterprise workflow automation today, and we expect it to remain the right architecture even as LLM capabilities continue to improve significantly.