AI agents change BPM because they introduce a layer that decides, reads documents and reasons inside the flow. They solve what RPA could not: processes with judgement, unstructured data, conversational steps and difficult exceptions.
Gartner positions agentic orchestration as one of the dominant trends in its Hype Cycle for Process Automation 2026, projecting broad adoption across mid-market and large enterprises in the next 18 to 24 months. Industry analyst commentary points to roughly 80% of organisations evaluating or piloting agentic capabilities by year-end 2026.
The question is no longer "will we use AI agents?", but "how do we govern them without losing traceability or human control?". That is exactly where classical BPM, with BPMN 2.0 and human oversight, becomes strategic again.
This guide explains what an AI agent is inside a process, how it differs from RPA, five real use cases, the technology stack that supports it and how to start with a measurable pilot in weeks.
An AI agent is a software unit that receives an input (a document, a message, an event, a dataset), reasons using a large language model (LLM) and produces a structured output the process can consume: a classification, an extracted field, a justified decision or a conversational answer. Unlike a traditional microservice, the agent does not follow a closed algorithm; it applies a trained skill and adapts to the context it is given.
Inside a BPM platform, that agent lives as just another step of the process. An inbound invoice enters the flow, an automatic BPMN task calls the "invoice-classifier" agent, the agent returns {type: "energy", vendor: "Iberdrola", confidence: 0.94}, and the next task decides whether to approve automatically or escalate to a human. Everything is recorded in the process log.
The difference compared to a "bot" or RPA is a difference of nature: a bot executes a fixed sequence, RPA imitates a human's interaction with a UI, and an AI agent decides with judgement over unstructured data. This finally opens up processes that were previously intractable without manual intervention: reading a contract, classifying an email, interpreting a payslip or deciding whether an exception is legitimate.
They are not rival technologies; they cover different intents. The quick rule: RPA for deterministic tasks against stable UIs; AI agent for tasks requiring judgement or unstructured data.
| Characteristic | RPA | AI agent |
|---|---|---|
| Nature | Deterministic. Executes fixed clicks and sequences. | Reasoning. Decides based on context. |
| Data | Structured (forms, sheets, fields). | Mixed: structured + unstructured (PDFs, email, voice). |
| Adaptation | None. Breaks when the UI changes. | Adapts to context. Tolerates input variability. |
| Traceability | Execution logs (which click, when). | Logs + reasoning trace (what it decided and why). |
| Maintenance | High when the source UI or system changes. | Low as the LLM improves; only prompts and rules need tuning. |
| When to use | Rigid repetitive processes with stable UIs. | Processes that need decision, reading or context. |
In practice they coexist: a real-world process typically combines AI agents for reading and decision-making, RPA or API connectors to move data between systems, and a BPM engine orchestrating everything.
Not all agents do the same thing. A productive BPM platform usually combines four typical categories, each with its own specialised skill, success metrics and integration pattern inside the process.
Classifies inputs against business taxonomies: invoices (energy, telecom, supplies), contracts (NDA, employment, commercial), inbound mail (incident, sales, HR), support tickets. Its output routes work to the correct flow, eliminating manual triage.
Reads PDFs, images or emails and extracts key fields: amounts, IBAN, tax ID, due dates, clauses. It is template-free IDP (Intelligent Document Processing): it learns to read any vendor format without manual setup.
Receives the process data and proposes a recommendation with justification: approve an invoice, reject a leave request, escalate an incident, accept an account opening. Always with explainable reasoning; never as a black box. A human validates or overrides with one click.
Converses with a customer or employee inside the flow: requests missing data, answers questions from the knowledge base, validates identity or collects a signature. It is the front-end of the process for cases where a rigid form drives abandonment.
These five scenarios concentrate most of the value of AI agents in BPM. They share a pattern: high volume, repetitive decisions with judgement, unstructured data and the need for traceability.
Agents in series: classifier (invoice type) → extractor (amount, VAT, IBAN, vendor tax ID, due date) → decision (auto-approve if amount < €1,000 and recurring vendor; escalate to manager otherwise).
Typical outcome: between 65% and 80% of invoices are approved without human touch; the rest reach the human reviewer with the full context already extracted. The bottleneck is gone.
Agents in parallel: extractor over ID document and passport + KYC agent (matching against sanctions, PEP and compliance lists) + conversational agent to gather missing data + decision "auto-onboard" or "analyst review".
Account opening moves from 3 to 7 days down to hours. Compliance is recorded step by step for the regulator.
Agents in cascade: classifier (incident, sales, HR, other) → routing to the correct team → conversational agent backed by the knowledge base (FAQ + past tickets) → automatic answer for resolved cases → escalation to a human for sensitive cases or premium customers.
First response time drops to minutes; level-one load decreases sharply; human agents focus on cases where they truly add value.
Agents: triage of leave type → document verification (extractor over medical or family certificates) → validation against the applicable collective bargaining agreement (convenio) → decision "auto-approve" or "escalate to line manager".
The people team is freed from administrative workflow and refocuses on relationships, talent and sensitive cases.
Agents: extractor over the contract → detector for critical clauses (liability cap, governing law, non-compete, personal data) → comparison against the approved template → deviation report for the legal team.
The lawyer receives a pre-processed report instead of a blank PDF. Review time drops from hours per contract to minutes, without losing rigour.
The BPMN 2.0 standard already had the right primitive: the Service Task. An AI agent is modelled as a Service Task with a specific configuration: skill (classify, extract, decide, converse), prompt or business instruction, context data received from the process, expected output with its schema (typed JSON), and fallback policy if the agent does not respond with sufficient confidence.
The BPM engine keeps three responsibilities the agent alone cannot cover. First, orchestration: which agent is invoked, in what order, with what data and what happens with its output. Second, traceability: every invocation is recorded with timestamp, input, output, model used, cost and confidence level. Third, human oversight: the flow decides when to escalate to a person based on rules (amount, risk, confidence, customer tier).
This architecture lets you version processes like code: when a prompt is improved or a model is swapped, the change is recorded in the new BPMN version. In-flight instances keep running with the previous version until they complete; new instances pick up the updated version. Without this, an AI agent "in production" is ungovernable at scale. BPMN 2.0 is published by the Object Management Group (OMG) as a global ISO/IEC 19510 standard.
In any productive process with business decisions, human oversight is not a nice-to-have: it is a governance obligation and, in many sectors (banking, healthcare, insurance, public sector), also a legal one. The EU AI Act (Regulation (EU) 2024/1689) explicitly reinforces the need for meaningful human oversight for high-risk systems, and Article 22 of the GDPR (Regulation (EU) 2016/679) protects individuals from purely automated decisions with legal or similarly significant effects.
The human-in-the-loop pattern in BPM with agents works like this. Each agent returns, alongside its output, a confidence score (0-1). If confidence exceeds the configured threshold (for example 0.85), the process continues automatically. Below that, the flow routes the task to a human inbox with full context: original input, agent proposal, justification and comparable data.
The full audit trail includes: timestamp of each invocation, LLM model used, prompt version, exact input sent (sanitised if it contains personal data), structured output, confidence score, final decision (automatic or human) and who took it if human. This record is what enables responding, months later, to an internal audit or a regulatory request.
Without explicit traceability of the agent's reasoning and without confidence-based human escalation, there is no productive AI project. There is only a demo.
Without going into code, this is how a productive AI agent is composed inside a BPM platform. Five layers chosen and combined based on the use case.
The reasoning engine: GPT (OpenAI/Azure), Claude (Anthropic), Gemini (Google) or open-source models deployable on-premise (Llama, Mistral). Multi-model lets you choose by cost, latency and data sovereignty.
A vector database stores company knowledge (manuals, contracts, FAQs, past cases) in a form the LLM can query. RAG (Retrieval-Augmented Generation) retrieves what is relevant before reasoning, avoiding hallucinations.
The agent can invoke system actions: query the ERP, read a customer in the CRM, call a REST API, check an IBAN, generate a document. Each tool has a clear input/output contract.
Decision Tables, DMN or coded rules that validate or constrain the agent's output. Example: "approve only if amount < €1,000 and vendor in allowlist". This is the deterministic safety net around the LLM.
The orchestration, governance and traceability layer. BPMN defines order, human escalation, SLAs, versions and logs. This is what turns the technology stack into an auditable business process. ISO/IEC 42001 (AI management system, AIMS) describes the management practices that formalise this layer.
The ROI of a BPM-with-AI-agents project is not a single number; it depends on the process, the volume and how much human effort was involved before. The ranges we observe in real projects are these (indicative, not a guarantee):
To measure it, you need a baseline before the project. If you do not measure the pre-agent process (time, errors, cost, satisfaction), you cannot prove ROI afterwards. That is why a well-governed pilot always includes 4 to 8 weeks of baseline measurement before the agent goes live in production.
Do not start with the most complex process, nor the one with the largest potential impact. Start with the most measurable and lowest-exception-risk process. Five steps:
High volume, repetitive decisions with judgement, unstructured data and clear before/after metrics. Invoices and ticket triage are the favourites.
Does it classify, extract, decide or converse? What input does it receive? What structured output does it return? What minimum confidence is required before it acts?
Model the flow in BPMN 2.0, add the agent Service Task, wire input/output to the neighbouring tasks, and define human escalation rules and SLAs.
For 4 to 8 weeks, the agent proposes but humans validate every decision. This lets you tune prompts, rules and thresholds on real data without risk.
When accuracy is stable, gradually raise the automation threshold. Measure cycle time, errors and cost before/after. Repeat the pattern with the next process.
AI agents in production carry real risks. The good news is that they are well understood and have standard mitigations.
The model "invents" a data point that was not in the input.
RAG against verifiable sources, business-rule validation before action, human escalation when confidence is low, and prompts that require citing the source.
Personal or confidential information sent to an external LLM.
EU data residency, a DPA with the LLM provider, encryption in transit and at rest, pre-prompt anonymisation where applicable, and an on-premise model option for critical workloads.
Pricing, policy or availability changes from the provider.
Multi-model architecture (GPT, Claude, Gemini, open source) with a provider-abstraction layer, portable prompts and automatic fallback to an alternative model.
The agent discriminates due to bias inherited from the model or the data.
Regular decision audits, fairness metrics per segment, mandatory human oversight for legally significant decisions (AI Act art. 14) and periodic review of the prompt and the data used in RAG.
RPA executes deterministic clicks against user interfaces; it breaks when the UI changes and it does not decide. An AI agent reasons over structured and unstructured data, makes decisions with justification and adapts to context. RPA is for rigid repetitive tasks; AI agents are for processes that need judgement, document reading or conversation.
Yes. In Dokuflex an AI agent is configured as a step inside a BPMN process: you choose the skill (classify, extract, decide, converse), define the prompt and the context data, and wire the output to the next task. No code is required; what is required is process design and human validation during the pilot.
They are GDPR compliant when deployed with data in the EU, encryption in transit and at rest, traceable processing and documented legal basis. For automated decisions with significant legal effects (GDPR article 22), meaningful human oversight must be preserved. The EU AI Act (Regulation 2024/1689) adds extra requirements for high-risk systems.
Three layers. RAG (Retrieval-Augmented Generation) so the agent answers only over verifiable data; business rules that validate the output before any action is taken; and human-in-the-loop with a confidence threshold that escalates uncertain cases to a reviewer. The agent's reasoning trace is stored in the BPM audit log.
Not in the real scenarios we observe. Agents take over the repetitive, low-value work (classifying, extracting, drafting), leaving humans to handle the final decision, exceptions and customer relationships. What changes is capacity: the same team can process several times more volume without losing quality.
Dokuflex is multi-model: it integrates GPT (OpenAI/Azure), Claude (Anthropic) and open-source models deployable on private infrastructure. Customers choose based on cost, latency, data sovereignty or the need for on-premise deployment. This avoids lock-in to a single vendor.
Metered consumption (tokens or agent executions) plus the BPM license. Dokuflex bills per process or per execution volume, with predictable cost per unit of work. Enterprise environments can choose an annual flat fee with a fixed token budget.
Technically you can invoke a standalone AI agent from a script. But without a BPMS you will have no orchestration, traceability, human oversight or version governance. The BPMS is what turns an isolated AI agent into a productive, auditable and maintainable process. It is the difference between a demo and production.
Indicative information. Adoption figures and ROI ranges are market estimates and typical observed outcomes; they do not constitute a guarantee of results. For specific cases, consult your IT and compliance teams.
Model the flow, add AI agents where they add value, keep human oversight in place and measure impact from the very first pilot.